llama.cpp/ggml-metal.metal

6883 lines
244 KiB
Metal
Raw Normal View History

#define GGML_COMMON_DECL_METAL
#define GGML_COMMON_IMPL_METAL
#include "ggml-common.h"
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
#include <metal_stdlib>
using namespace metal;
#define MAX(x, y) ((x) > (y) ? (x) : (y))
#define MIN(x, y) ((x) < (y) ? (x) : (y))
#define SWAP(x, y) { auto tmp = (x); (x) = (y); (y) = tmp; }
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
#define N_SIMDWIDTH 32 // assuming SIMD group size is 32
enum ggml_sort_order {
ggml : mul_mat_id use the same tensor for all the experts (#6387) * ggml : update mul_mat_id to use the same tensor for all the experts * update cuda * minor * update metal * update test-backend-ops * fix cuda * Update ggml-metal.m Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * update convert.py * update convert-hf-to-gguf.py * update convert.py for mixtral hf models * Update convert-hf-to-gguf.py Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * cuda : support non-pow-2 number of experts * allow quantize to work for split and merged experts models in the same way * cleanup + disable mmap automatically with split tensors models * update imatrix * test-backend-ops : test qwen argsort * update grok model loading * llama : add merged experts tensors to the grok tensor map * minor * gguf : bump version * fix quantizing of merged experts * convert-hf-to-gguf.py : update grok (untested) * make linter happy * cuda/argsort : use shared memory instead of pool memory * convert : fix grok tensor names * metal : add support for non-pow-2 argsort * llama : more loader cleanup, better error checking * cuda : fix warning * llama : still use mmap for loading old models, but copy the data to a host buffer * add review note * llama : remove ffn tensor counting + add sanity check ggml-ci * convert : fix handling of n_experts == None ggml-ci * imatrix : fix ncall counters * llama : produce error if imatrix size does not match * quantize : terminate on errors + trace logs ggml-ci * metal : pad shared memory to 16 bytes --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-04-03 15:07:05 +02:00
GGML_SORT_ORDER_ASC,
GGML_SORT_ORDER_DESC,
};
// general-purpose kernel for addition, multiplication and division of two tensors
// pros: works for non-contiguous tensors, supports broadcast across all dims
llama : custom attention mask + parallel decoding + no context swaps (#3228) * tests : verify that RoPE is "additive" * llama : replace ggml_diag_mask_inf with ggml_add (custom -inf mask) * ggml : ggml_rope now takes a vector with positions instead of n_past * metal : add rope_f16 kernel + optimize cpy kernels * llama : unified KV cache + batch inference API * llama : add new llama_decode() API that works with llama_batch * llama : add cell_max heuristic for more efficient kv_cache * llama : extend llama_kv_cache API * llama : more robust cell_max heuristic + wip shift * metal : disable concurrency optimization * llama : add llama_kv_cache_shift_seq + no more context swaps * llama : apply K-cache roping for Falcon and Baichuan * speculative : fix KV cache management * parallel : example for serving multiple users in parallel * parallel : disable hot-plug to avoid cache fragmentation * fixes : speculative KV cache + llama worst-case graph * llama : extend batch API to select which logits to output * llama : fix worst case graph build * ggml-cuda : update rope implementation for parallel decoding (#3254) * ggml-cuda : update rope implementation for parallel decoding * better solution for p0 computation * fix rope * simpler rope implementation --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * make : add parallel to build + fix static functions in llama.cpp * simple : fix token counting * parallel : various improvements * llama : fix cell_max logic + rename functions * parallel : try smaller batches when the KV cache is fragmented * parallel : fix sequence termination criteria * llama : silence errors KV cache errors * parallel : remove new line from prompt * parallel : process system prompt once + configurable paramters + llama API * parallel : remove question with short answers * parallel : count cache misses * parallel : print misses on each request * parallel : minor * llama : fix n_kv to never become 0 * parallel : rename hot-plug to continuous-batching * llama : improve llama_batch API + simplify parallel example * simple : add parallel decoding support * simple : improve comments + free batch * ggml-cuda : add rope f16, restore performance with parallel decoding (#3272) * ggml-cuda : add rope f16, restore performance * offload KQ_mask with all models * fix rope shift --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * llama : disable MPI for now ggml-ci * train : make KQ_pos memory buffer permanent via dummy scale op * ggml : revert change to ggml_cpy, add ggml_cont_Nd instead (#3275) ggml-ci * parallel : fix bug (extra BOS) + smaller token_prev array * parallel : fix cases where the input prompts can overflow the batch * parallel : add disabled experimental batch chunking in powers of two * llama : llama.h formatting + comments * simple : add README.md * llama : fix kv cache heuristic when context is less than 32 * parallel : fix crash when `-n -1` * llama : simplify returns if/else branches * metal : use mm kernels for batch size > 2 * examples : utilize new llama_get_logits_ith() * examples : add example for batched decoding * examples : do not eval prompt 2 times (close #3348) * server : clear the KV cache beyond n_past before llama_decode * server : avoid context swaps by shifting the KV cache --------- Co-authored-by: slaren <slarengh@gmail.com>
2023-09-28 18:04:36 +02:00
// cons: not very efficient
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
kernel void kernel_add(
llama : custom attention mask + parallel decoding + no context swaps (#3228) * tests : verify that RoPE is "additive" * llama : replace ggml_diag_mask_inf with ggml_add (custom -inf mask) * ggml : ggml_rope now takes a vector with positions instead of n_past * metal : add rope_f16 kernel + optimize cpy kernels * llama : unified KV cache + batch inference API * llama : add new llama_decode() API that works with llama_batch * llama : add cell_max heuristic for more efficient kv_cache * llama : extend llama_kv_cache API * llama : more robust cell_max heuristic + wip shift * metal : disable concurrency optimization * llama : add llama_kv_cache_shift_seq + no more context swaps * llama : apply K-cache roping for Falcon and Baichuan * speculative : fix KV cache management * parallel : example for serving multiple users in parallel * parallel : disable hot-plug to avoid cache fragmentation * fixes : speculative KV cache + llama worst-case graph * llama : extend batch API to select which logits to output * llama : fix worst case graph build * ggml-cuda : update rope implementation for parallel decoding (#3254) * ggml-cuda : update rope implementation for parallel decoding * better solution for p0 computation * fix rope * simpler rope implementation --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * make : add parallel to build + fix static functions in llama.cpp * simple : fix token counting * parallel : various improvements * llama : fix cell_max logic + rename functions * parallel : try smaller batches when the KV cache is fragmented * parallel : fix sequence termination criteria * llama : silence errors KV cache errors * parallel : remove new line from prompt * parallel : process system prompt once + configurable paramters + llama API * parallel : remove question with short answers * parallel : count cache misses * parallel : print misses on each request * parallel : minor * llama : fix n_kv to never become 0 * parallel : rename hot-plug to continuous-batching * llama : improve llama_batch API + simplify parallel example * simple : add parallel decoding support * simple : improve comments + free batch * ggml-cuda : add rope f16, restore performance with parallel decoding (#3272) * ggml-cuda : add rope f16, restore performance * offload KQ_mask with all models * fix rope shift --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * llama : disable MPI for now ggml-ci * train : make KQ_pos memory buffer permanent via dummy scale op * ggml : revert change to ggml_cpy, add ggml_cont_Nd instead (#3275) ggml-ci * parallel : fix bug (extra BOS) + smaller token_prev array * parallel : fix cases where the input prompts can overflow the batch * parallel : add disabled experimental batch chunking in powers of two * llama : llama.h formatting + comments * simple : add README.md * llama : fix kv cache heuristic when context is less than 32 * parallel : fix crash when `-n -1` * llama : simplify returns if/else branches * metal : use mm kernels for batch size > 2 * examples : utilize new llama_get_logits_ith() * examples : add example for batched decoding * examples : do not eval prompt 2 times (close #3348) * server : clear the KV cache beyond n_past before llama_decode * server : avoid context swaps by shifting the KV cache --------- Co-authored-by: slaren <slarengh@gmail.com>
2023-09-28 18:04:36 +02:00
device const char * src0,
device const char * src1,
device char * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
constant int64_t & ne03,
constant uint64_t & nb00,
constant uint64_t & nb01,
constant uint64_t & nb02,
constant uint64_t & nb03,
llama : custom attention mask + parallel decoding + no context swaps (#3228) * tests : verify that RoPE is "additive" * llama : replace ggml_diag_mask_inf with ggml_add (custom -inf mask) * ggml : ggml_rope now takes a vector with positions instead of n_past * metal : add rope_f16 kernel + optimize cpy kernels * llama : unified KV cache + batch inference API * llama : add new llama_decode() API that works with llama_batch * llama : add cell_max heuristic for more efficient kv_cache * llama : extend llama_kv_cache API * llama : more robust cell_max heuristic + wip shift * metal : disable concurrency optimization * llama : add llama_kv_cache_shift_seq + no more context swaps * llama : apply K-cache roping for Falcon and Baichuan * speculative : fix KV cache management * parallel : example for serving multiple users in parallel * parallel : disable hot-plug to avoid cache fragmentation * fixes : speculative KV cache + llama worst-case graph * llama : extend batch API to select which logits to output * llama : fix worst case graph build * ggml-cuda : update rope implementation for parallel decoding (#3254) * ggml-cuda : update rope implementation for parallel decoding * better solution for p0 computation * fix rope * simpler rope implementation --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * make : add parallel to build + fix static functions in llama.cpp * simple : fix token counting * parallel : various improvements * llama : fix cell_max logic + rename functions * parallel : try smaller batches when the KV cache is fragmented * parallel : fix sequence termination criteria * llama : silence errors KV cache errors * parallel : remove new line from prompt * parallel : process system prompt once + configurable paramters + llama API * parallel : remove question with short answers * parallel : count cache misses * parallel : print misses on each request * parallel : minor * llama : fix n_kv to never become 0 * parallel : rename hot-plug to continuous-batching * llama : improve llama_batch API + simplify parallel example * simple : add parallel decoding support * simple : improve comments + free batch * ggml-cuda : add rope f16, restore performance with parallel decoding (#3272) * ggml-cuda : add rope f16, restore performance * offload KQ_mask with all models * fix rope shift --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * llama : disable MPI for now ggml-ci * train : make KQ_pos memory buffer permanent via dummy scale op * ggml : revert change to ggml_cpy, add ggml_cont_Nd instead (#3275) ggml-ci * parallel : fix bug (extra BOS) + smaller token_prev array * parallel : fix cases where the input prompts can overflow the batch * parallel : add disabled experimental batch chunking in powers of two * llama : llama.h formatting + comments * simple : add README.md * llama : fix kv cache heuristic when context is less than 32 * parallel : fix crash when `-n -1` * llama : simplify returns if/else branches * metal : use mm kernels for batch size > 2 * examples : utilize new llama_get_logits_ith() * examples : add example for batched decoding * examples : do not eval prompt 2 times (close #3348) * server : clear the KV cache beyond n_past before llama_decode * server : avoid context swaps by shifting the KV cache --------- Co-authored-by: slaren <slarengh@gmail.com>
2023-09-28 18:04:36 +02:00
constant int64_t & ne10,
constant int64_t & ne11,
constant int64_t & ne12,
constant int64_t & ne13,
constant uint64_t & nb10,
constant uint64_t & nb11,
constant uint64_t & nb12,
constant uint64_t & nb13,
llama : custom attention mask + parallel decoding + no context swaps (#3228) * tests : verify that RoPE is "additive" * llama : replace ggml_diag_mask_inf with ggml_add (custom -inf mask) * ggml : ggml_rope now takes a vector with positions instead of n_past * metal : add rope_f16 kernel + optimize cpy kernels * llama : unified KV cache + batch inference API * llama : add new llama_decode() API that works with llama_batch * llama : add cell_max heuristic for more efficient kv_cache * llama : extend llama_kv_cache API * llama : more robust cell_max heuristic + wip shift * metal : disable concurrency optimization * llama : add llama_kv_cache_shift_seq + no more context swaps * llama : apply K-cache roping for Falcon and Baichuan * speculative : fix KV cache management * parallel : example for serving multiple users in parallel * parallel : disable hot-plug to avoid cache fragmentation * fixes : speculative KV cache + llama worst-case graph * llama : extend batch API to select which logits to output * llama : fix worst case graph build * ggml-cuda : update rope implementation for parallel decoding (#3254) * ggml-cuda : update rope implementation for parallel decoding * better solution for p0 computation * fix rope * simpler rope implementation --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * make : add parallel to build + fix static functions in llama.cpp * simple : fix token counting * parallel : various improvements * llama : fix cell_max logic + rename functions * parallel : try smaller batches when the KV cache is fragmented * parallel : fix sequence termination criteria * llama : silence errors KV cache errors * parallel : remove new line from prompt * parallel : process system prompt once + configurable paramters + llama API * parallel : remove question with short answers * parallel : count cache misses * parallel : print misses on each request * parallel : minor * llama : fix n_kv to never become 0 * parallel : rename hot-plug to continuous-batching * llama : improve llama_batch API + simplify parallel example * simple : add parallel decoding support * simple : improve comments + free batch * ggml-cuda : add rope f16, restore performance with parallel decoding (#3272) * ggml-cuda : add rope f16, restore performance * offload KQ_mask with all models * fix rope shift --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * llama : disable MPI for now ggml-ci * train : make KQ_pos memory buffer permanent via dummy scale op * ggml : revert change to ggml_cpy, add ggml_cont_Nd instead (#3275) ggml-ci * parallel : fix bug (extra BOS) + smaller token_prev array * parallel : fix cases where the input prompts can overflow the batch * parallel : add disabled experimental batch chunking in powers of two * llama : llama.h formatting + comments * simple : add README.md * llama : fix kv cache heuristic when context is less than 32 * parallel : fix crash when `-n -1` * llama : simplify returns if/else branches * metal : use mm kernels for batch size > 2 * examples : utilize new llama_get_logits_ith() * examples : add example for batched decoding * examples : do not eval prompt 2 times (close #3348) * server : clear the KV cache beyond n_past before llama_decode * server : avoid context swaps by shifting the KV cache --------- Co-authored-by: slaren <slarengh@gmail.com>
2023-09-28 18:04:36 +02:00
constant int64_t & ne0,
constant int64_t & ne1,
constant int64_t & ne2,
constant int64_t & ne3,
constant uint64_t & nb0,
constant uint64_t & nb1,
constant uint64_t & nb2,
constant uint64_t & nb3,
constant int64_t & offs,
llama : custom attention mask + parallel decoding + no context swaps (#3228) * tests : verify that RoPE is "additive" * llama : replace ggml_diag_mask_inf with ggml_add (custom -inf mask) * ggml : ggml_rope now takes a vector with positions instead of n_past * metal : add rope_f16 kernel + optimize cpy kernels * llama : unified KV cache + batch inference API * llama : add new llama_decode() API that works with llama_batch * llama : add cell_max heuristic for more efficient kv_cache * llama : extend llama_kv_cache API * llama : more robust cell_max heuristic + wip shift * metal : disable concurrency optimization * llama : add llama_kv_cache_shift_seq + no more context swaps * llama : apply K-cache roping for Falcon and Baichuan * speculative : fix KV cache management * parallel : example for serving multiple users in parallel * parallel : disable hot-plug to avoid cache fragmentation * fixes : speculative KV cache + llama worst-case graph * llama : extend batch API to select which logits to output * llama : fix worst case graph build * ggml-cuda : update rope implementation for parallel decoding (#3254) * ggml-cuda : update rope implementation for parallel decoding * better solution for p0 computation * fix rope * simpler rope implementation --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * make : add parallel to build + fix static functions in llama.cpp * simple : fix token counting * parallel : various improvements * llama : fix cell_max logic + rename functions * parallel : try smaller batches when the KV cache is fragmented * parallel : fix sequence termination criteria * llama : silence errors KV cache errors * parallel : remove new line from prompt * parallel : process system prompt once + configurable paramters + llama API * parallel : remove question with short answers * parallel : count cache misses * parallel : print misses on each request * parallel : minor * llama : fix n_kv to never become 0 * parallel : rename hot-plug to continuous-batching * llama : improve llama_batch API + simplify parallel example * simple : add parallel decoding support * simple : improve comments + free batch * ggml-cuda : add rope f16, restore performance with parallel decoding (#3272) * ggml-cuda : add rope f16, restore performance * offload KQ_mask with all models * fix rope shift --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * llama : disable MPI for now ggml-ci * train : make KQ_pos memory buffer permanent via dummy scale op * ggml : revert change to ggml_cpy, add ggml_cont_Nd instead (#3275) ggml-ci * parallel : fix bug (extra BOS) + smaller token_prev array * parallel : fix cases where the input prompts can overflow the batch * parallel : add disabled experimental batch chunking in powers of two * llama : llama.h formatting + comments * simple : add README.md * llama : fix kv cache heuristic when context is less than 32 * parallel : fix crash when `-n -1` * llama : simplify returns if/else branches * metal : use mm kernels for batch size > 2 * examples : utilize new llama_get_logits_ith() * examples : add example for batched decoding * examples : do not eval prompt 2 times (close #3348) * server : clear the KV cache beyond n_past before llama_decode * server : avoid context swaps by shifting the KV cache --------- Co-authored-by: slaren <slarengh@gmail.com>
2023-09-28 18:04:36 +02:00
uint3 tgpig[[threadgroup_position_in_grid]],
uint3 tpitg[[thread_position_in_threadgroup]],
uint3 ntg[[threads_per_threadgroup]]) {
const int64_t i03 = tgpig.z;
const int64_t i02 = tgpig.y;
const int64_t i01 = tgpig.x;
const int64_t i13 = i03 % ne13;
const int64_t i12 = i02 % ne12;
const int64_t i11 = i01 % ne11;
device const char * src0_ptr = src0 + i03*nb03 + i02*nb02 + i01*nb01 + offs;
device const char * src1_ptr = src1 + i13*nb13 + i12*nb12 + i11*nb11;
device char * dst_ptr = dst + i03*nb3 + i02*nb2 + i01*nb1 + offs;
llama : custom attention mask + parallel decoding + no context swaps (#3228) * tests : verify that RoPE is "additive" * llama : replace ggml_diag_mask_inf with ggml_add (custom -inf mask) * ggml : ggml_rope now takes a vector with positions instead of n_past * metal : add rope_f16 kernel + optimize cpy kernels * llama : unified KV cache + batch inference API * llama : add new llama_decode() API that works with llama_batch * llama : add cell_max heuristic for more efficient kv_cache * llama : extend llama_kv_cache API * llama : more robust cell_max heuristic + wip shift * metal : disable concurrency optimization * llama : add llama_kv_cache_shift_seq + no more context swaps * llama : apply K-cache roping for Falcon and Baichuan * speculative : fix KV cache management * parallel : example for serving multiple users in parallel * parallel : disable hot-plug to avoid cache fragmentation * fixes : speculative KV cache + llama worst-case graph * llama : extend batch API to select which logits to output * llama : fix worst case graph build * ggml-cuda : update rope implementation for parallel decoding (#3254) * ggml-cuda : update rope implementation for parallel decoding * better solution for p0 computation * fix rope * simpler rope implementation --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * make : add parallel to build + fix static functions in llama.cpp * simple : fix token counting * parallel : various improvements * llama : fix cell_max logic + rename functions * parallel : try smaller batches when the KV cache is fragmented * parallel : fix sequence termination criteria * llama : silence errors KV cache errors * parallel : remove new line from prompt * parallel : process system prompt once + configurable paramters + llama API * parallel : remove question with short answers * parallel : count cache misses * parallel : print misses on each request * parallel : minor * llama : fix n_kv to never become 0 * parallel : rename hot-plug to continuous-batching * llama : improve llama_batch API + simplify parallel example * simple : add parallel decoding support * simple : improve comments + free batch * ggml-cuda : add rope f16, restore performance with parallel decoding (#3272) * ggml-cuda : add rope f16, restore performance * offload KQ_mask with all models * fix rope shift --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * llama : disable MPI for now ggml-ci * train : make KQ_pos memory buffer permanent via dummy scale op * ggml : revert change to ggml_cpy, add ggml_cont_Nd instead (#3275) ggml-ci * parallel : fix bug (extra BOS) + smaller token_prev array * parallel : fix cases where the input prompts can overflow the batch * parallel : add disabled experimental batch chunking in powers of two * llama : llama.h formatting + comments * simple : add README.md * llama : fix kv cache heuristic when context is less than 32 * parallel : fix crash when `-n -1` * llama : simplify returns if/else branches * metal : use mm kernels for batch size > 2 * examples : utilize new llama_get_logits_ith() * examples : add example for batched decoding * examples : do not eval prompt 2 times (close #3348) * server : clear the KV cache beyond n_past before llama_decode * server : avoid context swaps by shifting the KV cache --------- Co-authored-by: slaren <slarengh@gmail.com>
2023-09-28 18:04:36 +02:00
for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) {
const int i10 = i0 % ne10;
*((device float *)(dst_ptr + i0*nb0)) = *((device float *)(src0_ptr + i0*nb00)) + *((device float *)(src1_ptr + i10*nb10));
}
}
kernel void kernel_mul(
device const char * src0,
device const char * src1,
device char * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
constant int64_t & ne03,
constant uint64_t & nb00,
constant uint64_t & nb01,
constant uint64_t & nb02,
constant uint64_t & nb03,
constant int64_t & ne10,
constant int64_t & ne11,
constant int64_t & ne12,
constant int64_t & ne13,
constant uint64_t & nb10,
constant uint64_t & nb11,
constant uint64_t & nb12,
constant uint64_t & nb13,
constant int64_t & ne0,
constant int64_t & ne1,
constant int64_t & ne2,
constant int64_t & ne3,
constant uint64_t & nb0,
constant uint64_t & nb1,
constant uint64_t & nb2,
constant uint64_t & nb3,
uint3 tgpig[[threadgroup_position_in_grid]],
uint3 tpitg[[thread_position_in_threadgroup]],
uint3 ntg[[threads_per_threadgroup]]) {
const int64_t i03 = tgpig.z;
const int64_t i02 = tgpig.y;
const int64_t i01 = tgpig.x;
llama : custom attention mask + parallel decoding + no context swaps (#3228) * tests : verify that RoPE is "additive" * llama : replace ggml_diag_mask_inf with ggml_add (custom -inf mask) * ggml : ggml_rope now takes a vector with positions instead of n_past * metal : add rope_f16 kernel + optimize cpy kernels * llama : unified KV cache + batch inference API * llama : add new llama_decode() API that works with llama_batch * llama : add cell_max heuristic for more efficient kv_cache * llama : extend llama_kv_cache API * llama : more robust cell_max heuristic + wip shift * metal : disable concurrency optimization * llama : add llama_kv_cache_shift_seq + no more context swaps * llama : apply K-cache roping for Falcon and Baichuan * speculative : fix KV cache management * parallel : example for serving multiple users in parallel * parallel : disable hot-plug to avoid cache fragmentation * fixes : speculative KV cache + llama worst-case graph * llama : extend batch API to select which logits to output * llama : fix worst case graph build * ggml-cuda : update rope implementation for parallel decoding (#3254) * ggml-cuda : update rope implementation for parallel decoding * better solution for p0 computation * fix rope * simpler rope implementation --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * make : add parallel to build + fix static functions in llama.cpp * simple : fix token counting * parallel : various improvements * llama : fix cell_max logic + rename functions * parallel : try smaller batches when the KV cache is fragmented * parallel : fix sequence termination criteria * llama : silence errors KV cache errors * parallel : remove new line from prompt * parallel : process system prompt once + configurable paramters + llama API * parallel : remove question with short answers * parallel : count cache misses * parallel : print misses on each request * parallel : minor * llama : fix n_kv to never become 0 * parallel : rename hot-plug to continuous-batching * llama : improve llama_batch API + simplify parallel example * simple : add parallel decoding support * simple : improve comments + free batch * ggml-cuda : add rope f16, restore performance with parallel decoding (#3272) * ggml-cuda : add rope f16, restore performance * offload KQ_mask with all models * fix rope shift --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * llama : disable MPI for now ggml-ci * train : make KQ_pos memory buffer permanent via dummy scale op * ggml : revert change to ggml_cpy, add ggml_cont_Nd instead (#3275) ggml-ci * parallel : fix bug (extra BOS) + smaller token_prev array * parallel : fix cases where the input prompts can overflow the batch * parallel : add disabled experimental batch chunking in powers of two * llama : llama.h formatting + comments * simple : add README.md * llama : fix kv cache heuristic when context is less than 32 * parallel : fix crash when `-n -1` * llama : simplify returns if/else branches * metal : use mm kernels for batch size > 2 * examples : utilize new llama_get_logits_ith() * examples : add example for batched decoding * examples : do not eval prompt 2 times (close #3348) * server : clear the KV cache beyond n_past before llama_decode * server : avoid context swaps by shifting the KV cache --------- Co-authored-by: slaren <slarengh@gmail.com>
2023-09-28 18:04:36 +02:00
const int64_t i13 = i03 % ne13;
const int64_t i12 = i02 % ne12;
const int64_t i11 = i01 % ne11;
device const char * src0_ptr = src0 + i03*nb03 + i02*nb02 + i01*nb01;
device const char * src1_ptr = src1 + i13*nb13 + i12*nb12 + i11*nb11;
device char * dst_ptr = dst + i03*nb3 + i02*nb2 + i01*nb1;
for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) {
const int i10 = i0 % ne10;
*((device float *)(dst_ptr + i0*nb0)) = *((device float *)(src0_ptr + i0*nb00)) * *((device float *)(src1_ptr + i10*nb10));
}
}
kernel void kernel_div(
device const char * src0,
device const char * src1,
device char * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
constant int64_t & ne03,
constant uint64_t & nb00,
constant uint64_t & nb01,
constant uint64_t & nb02,
constant uint64_t & nb03,
constant int64_t & ne10,
constant int64_t & ne11,
constant int64_t & ne12,
constant int64_t & ne13,
constant uint64_t & nb10,
constant uint64_t & nb11,
constant uint64_t & nb12,
constant uint64_t & nb13,
constant int64_t & ne0,
constant int64_t & ne1,
constant int64_t & ne2,
constant int64_t & ne3,
constant uint64_t & nb0,
constant uint64_t & nb1,
constant uint64_t & nb2,
constant uint64_t & nb3,
uint3 tgpig[[threadgroup_position_in_grid]],
uint3 tpitg[[thread_position_in_threadgroup]],
uint3 ntg[[threads_per_threadgroup]]) {
const int64_t i03 = tgpig.z;
const int64_t i02 = tgpig.y;
const int64_t i01 = tgpig.x;
const int64_t i13 = i03 % ne13;
const int64_t i12 = i02 % ne12;
const int64_t i11 = i01 % ne11;
device const char * src0_ptr = src0 + i03*nb03 + i02*nb02 + i01*nb01;
device const char * src1_ptr = src1 + i13*nb13 + i12*nb12 + i11*nb11;
device char * dst_ptr = dst + i03*nb3 + i02*nb2 + i01*nb1;
for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) {
const int i10 = i0 % ne10;
*((device float *)(dst_ptr + i0*nb0)) = *((device float *)(src0_ptr + i0*nb00)) / *((device float *)(src1_ptr + i10*nb10));
llama : custom attention mask + parallel decoding + no context swaps (#3228) * tests : verify that RoPE is "additive" * llama : replace ggml_diag_mask_inf with ggml_add (custom -inf mask) * ggml : ggml_rope now takes a vector with positions instead of n_past * metal : add rope_f16 kernel + optimize cpy kernels * llama : unified KV cache + batch inference API * llama : add new llama_decode() API that works with llama_batch * llama : add cell_max heuristic for more efficient kv_cache * llama : extend llama_kv_cache API * llama : more robust cell_max heuristic + wip shift * metal : disable concurrency optimization * llama : add llama_kv_cache_shift_seq + no more context swaps * llama : apply K-cache roping for Falcon and Baichuan * speculative : fix KV cache management * parallel : example for serving multiple users in parallel * parallel : disable hot-plug to avoid cache fragmentation * fixes : speculative KV cache + llama worst-case graph * llama : extend batch API to select which logits to output * llama : fix worst case graph build * ggml-cuda : update rope implementation for parallel decoding (#3254) * ggml-cuda : update rope implementation for parallel decoding * better solution for p0 computation * fix rope * simpler rope implementation --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * make : add parallel to build + fix static functions in llama.cpp * simple : fix token counting * parallel : various improvements * llama : fix cell_max logic + rename functions * parallel : try smaller batches when the KV cache is fragmented * parallel : fix sequence termination criteria * llama : silence errors KV cache errors * parallel : remove new line from prompt * parallel : process system prompt once + configurable paramters + llama API * parallel : remove question with short answers * parallel : count cache misses * parallel : print misses on each request * parallel : minor * llama : fix n_kv to never become 0 * parallel : rename hot-plug to continuous-batching * llama : improve llama_batch API + simplify parallel example * simple : add parallel decoding support * simple : improve comments + free batch * ggml-cuda : add rope f16, restore performance with parallel decoding (#3272) * ggml-cuda : add rope f16, restore performance * offload KQ_mask with all models * fix rope shift --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * llama : disable MPI for now ggml-ci * train : make KQ_pos memory buffer permanent via dummy scale op * ggml : revert change to ggml_cpy, add ggml_cont_Nd instead (#3275) ggml-ci * parallel : fix bug (extra BOS) + smaller token_prev array * parallel : fix cases where the input prompts can overflow the batch * parallel : add disabled experimental batch chunking in powers of two * llama : llama.h formatting + comments * simple : add README.md * llama : fix kv cache heuristic when context is less than 32 * parallel : fix crash when `-n -1` * llama : simplify returns if/else branches * metal : use mm kernels for batch size > 2 * examples : utilize new llama_get_logits_ith() * examples : add example for batched decoding * examples : do not eval prompt 2 times (close #3348) * server : clear the KV cache beyond n_past before llama_decode * server : avoid context swaps by shifting the KV cache --------- Co-authored-by: slaren <slarengh@gmail.com>
2023-09-28 18:04:36 +02:00
}
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
}
// assumption: src1 is a row
// broadcast src1 into src0
kernel void kernel_add_row(
device const float4 * src0,
device const float4 * src1,
device float4 * dst,
constant uint64_t & nb [[buffer(28)]],
uint tpig[[thread_position_in_grid]]) {
dst[tpig] = src0[tpig] + src1[tpig % nb];
}
kernel void kernel_mul_row(
device const float4 * src0,
device const float4 * src1,
device float4 * dst,
constant uint64_t & nb [[buffer(28)]],
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
uint tpig[[thread_position_in_grid]]) {
dst[tpig] = src0[tpig] * src1[tpig % nb];
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
}
kernel void kernel_div_row(
device const float4 * src0,
device const float4 * src1,
device float4 * dst,
constant uint64_t & nb [[buffer(28)]],
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
uint tpig[[thread_position_in_grid]]) {
dst[tpig] = src0[tpig] / src1[tpig % nb];
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
}
kernel void kernel_scale(
device const float * src0,
device float * dst,
constant float & scale,
uint tpig[[thread_position_in_grid]]) {
dst[tpig] = src0[tpig] * scale;
}
kernel void kernel_scale_4(
device const float4 * src0,
device float4 * dst,
constant float & scale,
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
uint tpig[[thread_position_in_grid]]) {
dst[tpig] = src0[tpig] * scale;
}
kernel void kernel_clamp(
device const float * src0,
device float * dst,
constant float & min,
constant float & max,
uint tpig[[thread_position_in_grid]]) {
dst[tpig] = src0[tpig] < min ? min : (src0[tpig] > max ? max : src0[tpig]);
}
kernel void kernel_relu(
device const float * src0,
device float * dst,
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
uint tpig[[thread_position_in_grid]]) {
dst[tpig] = max(0.0f, src0[tpig]);
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
}
kernel void kernel_tanh(
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
device const float * src0,
device float * dst,
uint tpig[[thread_position_in_grid]]) {
device const float & x = src0[tpig];
dst[tpig] = precise::tanh(x);
}
constant float GELU_COEF_A = 0.044715f;
constant float GELU_QUICK_COEF = -1.702f;
constant float SQRT_2_OVER_PI = 0.79788456080286535587989211986876f;
kernel void kernel_gelu(
device const float * src0,
device float * dst,
uint tpig[[thread_position_in_grid]]) {
device const float & x = src0[tpig];
dst[tpig] = 0.5f*x*(1.0f + precise::tanh(SQRT_2_OVER_PI*x*(1.0f + GELU_COEF_A*x*x)));
}
kernel void kernel_gelu_4(
device const float4 * src0,
device float4 * dst,
uint tpig[[thread_position_in_grid]]) {
device const float4 & x = src0[tpig];
// BEWARE !!!
// Simply using "tanh" instead of "precise::tanh" will sometimes results in NaNs!
// This was observed with Falcon 7B and 40B models
//
dst[tpig] = 0.5f*x*(1.0f + precise::tanh(SQRT_2_OVER_PI*x*(1.0f + GELU_COEF_A*x*x)));
}
kernel void kernel_gelu_quick(
device const float * src0,
device float * dst,
uint tpig[[thread_position_in_grid]]) {
device const float & x = src0[tpig];
dst[tpig] = x*(1.0f/(1.0f+exp(GELU_QUICK_COEF*x)));
}
kernel void kernel_gelu_quick_4(
device const float4 * src0,
device float4 * dst,
uint tpig[[thread_position_in_grid]]) {
device const float4 & x = src0[tpig];
dst[tpig] = x*(1.0f/(1.0f+exp(GELU_QUICK_COEF*x)));
}
kernel void kernel_silu(
device const float * src0,
device float * dst,
uint tpig[[thread_position_in_grid]]) {
device const float & x = src0[tpig];
dst[tpig] = x / (1.0f + exp(-x));
}
kernel void kernel_silu_4(
device const float4 * src0,
device float4 * dst,
uint tpig[[thread_position_in_grid]]) {
device const float4 & x = src0[tpig];
dst[tpig] = x / (1.0f + exp(-x));
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
}
kernel void kernel_sqr(
device const float * src0,
device float * dst,
uint tpig[[thread_position_in_grid]]) {
dst[tpig] = src0[tpig] * src0[tpig];
}
kernel void kernel_sum_rows(
device const float * src0,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
constant int64_t & ne03,
constant uint64_t & nb00,
constant uint64_t & nb01,
constant uint64_t & nb02,
constant uint64_t & nb03,
constant int64_t & ne10,
constant int64_t & ne11,
constant int64_t & ne12,
constant int64_t & ne13,
constant uint64_t & nb10,
constant uint64_t & nb11,
constant uint64_t & nb12,
constant uint64_t & nb13,
constant int64_t & ne0,
constant int64_t & ne1,
constant int64_t & ne2,
constant int64_t & ne3,
constant uint64_t & nb0,
constant uint64_t & nb1,
constant uint64_t & nb2,
constant uint64_t & nb3,
uint3 tpig[[thread_position_in_grid]]) {
int64_t i3 = tpig.z;
int64_t i2 = tpig.y;
int64_t i1 = tpig.x;
if (i3 >= ne03 || i2 >= ne02 || i1 >= ne01) {
return;
}
device const float * src_row = (device const float *) ((device const char *) src0 + i1*nb01 + i2*nb02 + i3*nb03);
device float * dst_row = (device float *) ((device char *) dst + i1*nb1 + i2*nb2 + i3*nb3);
float row_sum = 0;
for (int64_t i0 = 0; i0 < ne00; i0++) {
row_sum += src_row[i0];
}
dst_row[0] = row_sum;
}
ggml : add Flash Attention (#5021) * ggml : add ggml_flash_attn_ext API * ggml : fix GQA support in ggml_flash_attn_ext * ggml : online attention (CPU) * metal : initial implementation * metal : f16 precision * metal : reduce branches * metal : specialize for head size * wip : 8 rows per simd group * wip : 4 rows per simd group * wip : template for rows per warp * metal : parallelize across KV size * metal : parallel reduce across heads * metal : efficient flash_attn_f16 implementation * metal : avoid redundant loads of the attention * metal : scale and mask in matrix form * metal : fix comment * llama : avoid ggml_cast, use F32 query * metal : add parallel reduce version (disabled) * metal : move output into local memory + optimize - the result from each simdgroup now stays in the registers - significantly reduced SRAM usage - more efficient skipping of -INF blocks - avoid simdgroup barrier in hot loop - add comments * metal : add tests, fix scaling, support C > 32 * metal : improve precision * ggml : fix f16 mad * metal : minor * metal : support Q > 8 * tests : add ATTN tests * metal : disable buffer allocation logs * tests : more * metal : faster inner loop for C == 32 * metal : fix array initialization * tests : ifdef * ggml : switch to padded F16 mask for ggml_soft_max, ggml_flash_attn_ext * ggml : fix ggml_soft_max mask requirement * cuda : fix soft_max to use correct mask size * cuda : add flash_attn kernel (wip) * metal : optimize softmax for C > 32 * metal : optimize softmax * tests : minor fix * cuda : avoid zeroing fragments * tests : update dims * cuda : fix __hisinf() result check * cuda : avoid warp_reduce for smax * cuda : use int instead of int64_t Noticeably improves performance (thanks to Johannes) * cuda : make loops use the same loop values Thanks Johannes again for the tip * cuda : unroll some of the loops * cuda : avoid __hisinf branches * cuda : use half2 in softmax * cuda : switch to 1 warp for bs > 16 * cuda : speed-up reduce part of the kernel * cuda : unroll Q*K^T loop * cuda : fix -INF block check * cuda : simplify softmax * cuda : fix matrix names * cuda : minor * llama : adapt to F16 KQ_pos * llama : adapt new models to F16 KQ_mask * ggml : fix F16 store (ARM NEON) * llama : fix type of KQ_mask and KQ_pos * ggml : fix CPU soft_max * tests : add hs=256 * cuda : fix build * metal : improve perf via smaller int registers * cuda : adapt soft_max to F16 mask and pos * CUDA: faster FlashAttention, kernel for bs == 1 * 16 cols for Phi-2 * no vec for hs, no hs==256 ncols==32 for Volta * adjust kernel selection logic * 4 warps, 256 stride for all D * no ncols == 64 * Multiple parallel blocks for batch size 1 * fix compile warnings * fix excessive KQ_b loads * fix cmake build * fix KV cache padding, NaN from INFINITY (#6438) * llama : flash_attn cparam + fix defrag * server: support flash_attn param * server: bench: enable flash_attn param * CUDA: refactor host code, dyn. par. blocks * fix flash_attn_vec_f16 race condition * flush softmax exp below threshold to 0 * store temp KQ in registers * Calculate KQ as FP32 if KQV has GGML_PREC_F32 * Add __hgt2_mask implementation for CUDA 11 * fix KQ FP32 precision fpr parallel_blocks > 1 * llama-bench : add -fa,--flash-attn arg * metal : add BS=1 kernel for flash attention (#6508) * metal : add BS=1 kernel for flash attention (wip) * metal : support more than 1 warps * metal : opts * metal : opt * metal : switch to parallel reduce * metal : reduce registers * metal : simplify * metal : initial FA vec kernel * metal : use F32 attention accumulators * batched-bench : add fattn arg * llama : simplify llama_build_kv_store ggml-ci * llama : adapt build_olmo to changes * ggml : fix arm fp16 store on windows * metal : clean-up * metal : clean-up kernel code * metal : minor * tests : remove benchmarks ggml-ci * ggml : fix avx512 const correctness ggml-ci * ggml : fix soft_max with bias on CPU ggml-ci * common : print --flash-attn in help * ggml : fix num dimensions in ggml_flash_attn_ext * llama : force disable flash attention for incompatible models * ggml : ggml_soft_max support F16/F32 mask/pos ggml-ci * cuda : uint -> uint32_t * cuda : "constexpr dim3" -> "const dim3" ggml-ci * cuda : try to fix __hgt2_mask ggml-ci * ggml : add TODO's for F16/F32 mask/pos support in other backends * llama : replace bool need_kq_pos with use_alibi * llama : prep ALiBi support for BERT models ggml-ci * llama : fix n_batch requirements ggml-ci * cont * server : add help for --flash-attn arg * llama : disable FA for AMD * tests : remove TMP_ATTN_BENCH ggml-ci * llama : support save/load state with FA enabled ggml-ci * ci : add CUDA save-load-state tests ggml-ci * llama : llama_kv_cache_clear zeroes data + fix save-load seq ggml-ci * llama : fix copy-paste errors, add TODO * llama : disallow incompatible states * llama : update llama_state_get_size after v_trans field * metal : remove tmp log * llama : add static reminder for llama_state_get_size * metal : fix max nsg ggml-ci * ci : fix arg order ggml-ci --------- Co-authored-by: Johannes Gäßler <johannesg@5d6.de> Co-authored-by: Pierrick HYMBERT <pierrick.hymbert@gmail.com>
2024-04-30 11:16:08 +02:00
template<typename T>
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
kernel void kernel_soft_max(
ggml : add Flash Attention (#5021) * ggml : add ggml_flash_attn_ext API * ggml : fix GQA support in ggml_flash_attn_ext * ggml : online attention (CPU) * metal : initial implementation * metal : f16 precision * metal : reduce branches * metal : specialize for head size * wip : 8 rows per simd group * wip : 4 rows per simd group * wip : template for rows per warp * metal : parallelize across KV size * metal : parallel reduce across heads * metal : efficient flash_attn_f16 implementation * metal : avoid redundant loads of the attention * metal : scale and mask in matrix form * metal : fix comment * llama : avoid ggml_cast, use F32 query * metal : add parallel reduce version (disabled) * metal : move output into local memory + optimize - the result from each simdgroup now stays in the registers - significantly reduced SRAM usage - more efficient skipping of -INF blocks - avoid simdgroup barrier in hot loop - add comments * metal : add tests, fix scaling, support C > 32 * metal : improve precision * ggml : fix f16 mad * metal : minor * metal : support Q > 8 * tests : add ATTN tests * metal : disable buffer allocation logs * tests : more * metal : faster inner loop for C == 32 * metal : fix array initialization * tests : ifdef * ggml : switch to padded F16 mask for ggml_soft_max, ggml_flash_attn_ext * ggml : fix ggml_soft_max mask requirement * cuda : fix soft_max to use correct mask size * cuda : add flash_attn kernel (wip) * metal : optimize softmax for C > 32 * metal : optimize softmax * tests : minor fix * cuda : avoid zeroing fragments * tests : update dims * cuda : fix __hisinf() result check * cuda : avoid warp_reduce for smax * cuda : use int instead of int64_t Noticeably improves performance (thanks to Johannes) * cuda : make loops use the same loop values Thanks Johannes again for the tip * cuda : unroll some of the loops * cuda : avoid __hisinf branches * cuda : use half2 in softmax * cuda : switch to 1 warp for bs > 16 * cuda : speed-up reduce part of the kernel * cuda : unroll Q*K^T loop * cuda : fix -INF block check * cuda : simplify softmax * cuda : fix matrix names * cuda : minor * llama : adapt to F16 KQ_pos * llama : adapt new models to F16 KQ_mask * ggml : fix F16 store (ARM NEON) * llama : fix type of KQ_mask and KQ_pos * ggml : fix CPU soft_max * tests : add hs=256 * cuda : fix build * metal : improve perf via smaller int registers * cuda : adapt soft_max to F16 mask and pos * CUDA: faster FlashAttention, kernel for bs == 1 * 16 cols for Phi-2 * no vec for hs, no hs==256 ncols==32 for Volta * adjust kernel selection logic * 4 warps, 256 stride for all D * no ncols == 64 * Multiple parallel blocks for batch size 1 * fix compile warnings * fix excessive KQ_b loads * fix cmake build * fix KV cache padding, NaN from INFINITY (#6438) * llama : flash_attn cparam + fix defrag * server: support flash_attn param * server: bench: enable flash_attn param * CUDA: refactor host code, dyn. par. blocks * fix flash_attn_vec_f16 race condition * flush softmax exp below threshold to 0 * store temp KQ in registers * Calculate KQ as FP32 if KQV has GGML_PREC_F32 * Add __hgt2_mask implementation for CUDA 11 * fix KQ FP32 precision fpr parallel_blocks > 1 * llama-bench : add -fa,--flash-attn arg * metal : add BS=1 kernel for flash attention (#6508) * metal : add BS=1 kernel for flash attention (wip) * metal : support more than 1 warps * metal : opts * metal : opt * metal : switch to parallel reduce * metal : reduce registers * metal : simplify * metal : initial FA vec kernel * metal : use F32 attention accumulators * batched-bench : add fattn arg * llama : simplify llama_build_kv_store ggml-ci * llama : adapt build_olmo to changes * ggml : fix arm fp16 store on windows * metal : clean-up * metal : clean-up kernel code * metal : minor * tests : remove benchmarks ggml-ci * ggml : fix avx512 const correctness ggml-ci * ggml : fix soft_max with bias on CPU ggml-ci * common : print --flash-attn in help * ggml : fix num dimensions in ggml_flash_attn_ext * llama : force disable flash attention for incompatible models * ggml : ggml_soft_max support F16/F32 mask/pos ggml-ci * cuda : uint -> uint32_t * cuda : "constexpr dim3" -> "const dim3" ggml-ci * cuda : try to fix __hgt2_mask ggml-ci * ggml : add TODO's for F16/F32 mask/pos support in other backends * llama : replace bool need_kq_pos with use_alibi * llama : prep ALiBi support for BERT models ggml-ci * llama : fix n_batch requirements ggml-ci * cont * server : add help for --flash-attn arg * llama : disable FA for AMD * tests : remove TMP_ATTN_BENCH ggml-ci * llama : support save/load state with FA enabled ggml-ci * ci : add CUDA save-load-state tests ggml-ci * llama : llama_kv_cache_clear zeroes data + fix save-load seq ggml-ci * llama : fix copy-paste errors, add TODO * llama : disallow incompatible states * llama : update llama_state_get_size after v_trans field * metal : remove tmp log * llama : add static reminder for llama_state_get_size * metal : fix max nsg ggml-ci * ci : fix arg order ggml-ci --------- Co-authored-by: Johannes Gäßler <johannesg@5d6.de> Co-authored-by: Pierrick HYMBERT <pierrick.hymbert@gmail.com>
2024-04-30 11:16:08 +02:00
device const char * src0,
device const char * src1,
device const char * src2,
device char * dst,
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
constant float & scale,
constant float & max_bias,
constant float & m0,
constant float & m1,
constant uint32_t & n_head_log2,
threadgroup float * buf [[threadgroup(0)]],
uint tgpig[[threadgroup_position_in_grid]],
uint tpitg[[thread_position_in_threadgroup]],
uint sgitg[[simdgroup_index_in_threadgroup]],
uint tiisg[[thread_index_in_simdgroup]],
uint ntg[[threads_per_threadgroup]]) {
const int64_t i03 = (tgpig) / (ne02*ne01);
const int64_t i02 = (tgpig - i03*ne02*ne01) / ne01;
const int64_t i01 = (tgpig - i03*ne02*ne01 - i02*ne01);
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
ggml : add Flash Attention (#5021) * ggml : add ggml_flash_attn_ext API * ggml : fix GQA support in ggml_flash_attn_ext * ggml : online attention (CPU) * metal : initial implementation * metal : f16 precision * metal : reduce branches * metal : specialize for head size * wip : 8 rows per simd group * wip : 4 rows per simd group * wip : template for rows per warp * metal : parallelize across KV size * metal : parallel reduce across heads * metal : efficient flash_attn_f16 implementation * metal : avoid redundant loads of the attention * metal : scale and mask in matrix form * metal : fix comment * llama : avoid ggml_cast, use F32 query * metal : add parallel reduce version (disabled) * metal : move output into local memory + optimize - the result from each simdgroup now stays in the registers - significantly reduced SRAM usage - more efficient skipping of -INF blocks - avoid simdgroup barrier in hot loop - add comments * metal : add tests, fix scaling, support C > 32 * metal : improve precision * ggml : fix f16 mad * metal : minor * metal : support Q > 8 * tests : add ATTN tests * metal : disable buffer allocation logs * tests : more * metal : faster inner loop for C == 32 * metal : fix array initialization * tests : ifdef * ggml : switch to padded F16 mask for ggml_soft_max, ggml_flash_attn_ext * ggml : fix ggml_soft_max mask requirement * cuda : fix soft_max to use correct mask size * cuda : add flash_attn kernel (wip) * metal : optimize softmax for C > 32 * metal : optimize softmax * tests : minor fix * cuda : avoid zeroing fragments * tests : update dims * cuda : fix __hisinf() result check * cuda : avoid warp_reduce for smax * cuda : use int instead of int64_t Noticeably improves performance (thanks to Johannes) * cuda : make loops use the same loop values Thanks Johannes again for the tip * cuda : unroll some of the loops * cuda : avoid __hisinf branches * cuda : use half2 in softmax * cuda : switch to 1 warp for bs > 16 * cuda : speed-up reduce part of the kernel * cuda : unroll Q*K^T loop * cuda : fix -INF block check * cuda : simplify softmax * cuda : fix matrix names * cuda : minor * llama : adapt to F16 KQ_pos * llama : adapt new models to F16 KQ_mask * ggml : fix F16 store (ARM NEON) * llama : fix type of KQ_mask and KQ_pos * ggml : fix CPU soft_max * tests : add hs=256 * cuda : fix build * metal : improve perf via smaller int registers * cuda : adapt soft_max to F16 mask and pos * CUDA: faster FlashAttention, kernel for bs == 1 * 16 cols for Phi-2 * no vec for hs, no hs==256 ncols==32 for Volta * adjust kernel selection logic * 4 warps, 256 stride for all D * no ncols == 64 * Multiple parallel blocks for batch size 1 * fix compile warnings * fix excessive KQ_b loads * fix cmake build * fix KV cache padding, NaN from INFINITY (#6438) * llama : flash_attn cparam + fix defrag * server: support flash_attn param * server: bench: enable flash_attn param * CUDA: refactor host code, dyn. par. blocks * fix flash_attn_vec_f16 race condition * flush softmax exp below threshold to 0 * store temp KQ in registers * Calculate KQ as FP32 if KQV has GGML_PREC_F32 * Add __hgt2_mask implementation for CUDA 11 * fix KQ FP32 precision fpr parallel_blocks > 1 * llama-bench : add -fa,--flash-attn arg * metal : add BS=1 kernel for flash attention (#6508) * metal : add BS=1 kernel for flash attention (wip) * metal : support more than 1 warps * metal : opts * metal : opt * metal : switch to parallel reduce * metal : reduce registers * metal : simplify * metal : initial FA vec kernel * metal : use F32 attention accumulators * batched-bench : add fattn arg * llama : simplify llama_build_kv_store ggml-ci * llama : adapt build_olmo to changes * ggml : fix arm fp16 store on windows * metal : clean-up * metal : clean-up kernel code * metal : minor * tests : remove benchmarks ggml-ci * ggml : fix avx512 const correctness ggml-ci * ggml : fix soft_max with bias on CPU ggml-ci * common : print --flash-attn in help * ggml : fix num dimensions in ggml_flash_attn_ext * llama : force disable flash attention for incompatible models * ggml : ggml_soft_max support F16/F32 mask/pos ggml-ci * cuda : uint -> uint32_t * cuda : "constexpr dim3" -> "const dim3" ggml-ci * cuda : try to fix __hgt2_mask ggml-ci * ggml : add TODO's for F16/F32 mask/pos support in other backends * llama : replace bool need_kq_pos with use_alibi * llama : prep ALiBi support for BERT models ggml-ci * llama : fix n_batch requirements ggml-ci * cont * server : add help for --flash-attn arg * llama : disable FA for AMD * tests : remove TMP_ATTN_BENCH ggml-ci * llama : support save/load state with FA enabled ggml-ci * ci : add CUDA save-load-state tests ggml-ci * llama : llama_kv_cache_clear zeroes data + fix save-load seq ggml-ci * llama : fix copy-paste errors, add TODO * llama : disallow incompatible states * llama : update llama_state_get_size after v_trans field * metal : remove tmp log * llama : add static reminder for llama_state_get_size * metal : fix max nsg ggml-ci * ci : fix arg order ggml-ci --------- Co-authored-by: Johannes Gäßler <johannesg@5d6.de> Co-authored-by: Pierrick HYMBERT <pierrick.hymbert@gmail.com>
2024-04-30 11:16:08 +02:00
device const float * psrc0 = (device const float *) src0 + (i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00);
device const T * pmask = src1 != src0 ? (device const T *) src1 + i01*ne00 : nullptr;
device const T * ppos = src2 != src0 ? (device const T *) src2 : nullptr;
device float * pdst = (device float *) dst + (i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00);
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
float slope = 0.0f;
// ALiBi
if (max_bias > 0.0f) {
const int64_t h = i02;
const float base = h < n_head_log2 ? m0 : m1;
const int exp = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
slope = pow(base, exp);
}
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
// parallel max
float lmax = -INFINITY;
for (int i00 = tpitg; i00 < ne00; i00 += ntg) {
lmax = MAX(lmax, psrc0[i00]*scale + (pmask ? pmask[i00] : 0.0f) + (ppos ? slope*ppos[i00] : 0.0f));
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
}
// find the max value in the block
float max_val = simd_max(lmax);
if (ntg > N_SIMDWIDTH) {
if (sgitg == 0) {
buf[tiisg] = -INFINITY;
}
threadgroup_barrier(mem_flags::mem_threadgroup);
if (tiisg == 0) {
buf[sgitg] = max_val;
}
threadgroup_barrier(mem_flags::mem_threadgroup);
max_val = buf[tiisg];
max_val = simd_max(max_val);
}
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
// parallel sum
float lsum = 0.0f;
for (int i00 = tpitg; i00 < ne00; i00 += ntg) {
const float exp_psrc0 = exp((psrc0[i00]*scale + (pmask ? pmask[i00] : 0.0f) + (ppos ? slope*ppos[i00] : 0.0f)) - max_val);
lsum += exp_psrc0;
pdst[i00] = exp_psrc0;
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
}
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 13:04:25 +01:00
// This barrier fixes a failing test
// ref: https://github.com/ggerganov/ggml/pull/621#discussion_r1425156335
threadgroup_barrier(mem_flags::mem_none);
float sum = simd_sum(lsum);
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 13:04:25 +01:00
if (ntg > N_SIMDWIDTH) {
if (sgitg == 0) {
buf[tiisg] = 0.0f;
}
threadgroup_barrier(mem_flags::mem_threadgroup);
if (tiisg == 0) {
buf[sgitg] = sum;
}
threadgroup_barrier(mem_flags::mem_threadgroup);
sum = buf[tiisg];
sum = simd_sum(sum);
}
const float inv_sum = 1.0f/sum;
for (int i00 = tpitg; i00 < ne00; i00 += ntg) {
pdst[i00] *= inv_sum;
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
}
}
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
ggml : add Flash Attention (#5021) * ggml : add ggml_flash_attn_ext API * ggml : fix GQA support in ggml_flash_attn_ext * ggml : online attention (CPU) * metal : initial implementation * metal : f16 precision * metal : reduce branches * metal : specialize for head size * wip : 8 rows per simd group * wip : 4 rows per simd group * wip : template for rows per warp * metal : parallelize across KV size * metal : parallel reduce across heads * metal : efficient flash_attn_f16 implementation * metal : avoid redundant loads of the attention * metal : scale and mask in matrix form * metal : fix comment * llama : avoid ggml_cast, use F32 query * metal : add parallel reduce version (disabled) * metal : move output into local memory + optimize - the result from each simdgroup now stays in the registers - significantly reduced SRAM usage - more efficient skipping of -INF blocks - avoid simdgroup barrier in hot loop - add comments * metal : add tests, fix scaling, support C > 32 * metal : improve precision * ggml : fix f16 mad * metal : minor * metal : support Q > 8 * tests : add ATTN tests * metal : disable buffer allocation logs * tests : more * metal : faster inner loop for C == 32 * metal : fix array initialization * tests : ifdef * ggml : switch to padded F16 mask for ggml_soft_max, ggml_flash_attn_ext * ggml : fix ggml_soft_max mask requirement * cuda : fix soft_max to use correct mask size * cuda : add flash_attn kernel (wip) * metal : optimize softmax for C > 32 * metal : optimize softmax * tests : minor fix * cuda : avoid zeroing fragments * tests : update dims * cuda : fix __hisinf() result check * cuda : avoid warp_reduce for smax * cuda : use int instead of int64_t Noticeably improves performance (thanks to Johannes) * cuda : make loops use the same loop values Thanks Johannes again for the tip * cuda : unroll some of the loops * cuda : avoid __hisinf branches * cuda : use half2 in softmax * cuda : switch to 1 warp for bs > 16 * cuda : speed-up reduce part of the kernel * cuda : unroll Q*K^T loop * cuda : fix -INF block check * cuda : simplify softmax * cuda : fix matrix names * cuda : minor * llama : adapt to F16 KQ_pos * llama : adapt new models to F16 KQ_mask * ggml : fix F16 store (ARM NEON) * llama : fix type of KQ_mask and KQ_pos * ggml : fix CPU soft_max * tests : add hs=256 * cuda : fix build * metal : improve perf via smaller int registers * cuda : adapt soft_max to F16 mask and pos * CUDA: faster FlashAttention, kernel for bs == 1 * 16 cols for Phi-2 * no vec for hs, no hs==256 ncols==32 for Volta * adjust kernel selection logic * 4 warps, 256 stride for all D * no ncols == 64 * Multiple parallel blocks for batch size 1 * fix compile warnings * fix excessive KQ_b loads * fix cmake build * fix KV cache padding, NaN from INFINITY (#6438) * llama : flash_attn cparam + fix defrag * server: support flash_attn param * server: bench: enable flash_attn param * CUDA: refactor host code, dyn. par. blocks * fix flash_attn_vec_f16 race condition * flush softmax exp below threshold to 0 * store temp KQ in registers * Calculate KQ as FP32 if KQV has GGML_PREC_F32 * Add __hgt2_mask implementation for CUDA 11 * fix KQ FP32 precision fpr parallel_blocks > 1 * llama-bench : add -fa,--flash-attn arg * metal : add BS=1 kernel for flash attention (#6508) * metal : add BS=1 kernel for flash attention (wip) * metal : support more than 1 warps * metal : opts * metal : opt * metal : switch to parallel reduce * metal : reduce registers * metal : simplify * metal : initial FA vec kernel * metal : use F32 attention accumulators * batched-bench : add fattn arg * llama : simplify llama_build_kv_store ggml-ci * llama : adapt build_olmo to changes * ggml : fix arm fp16 store on windows * metal : clean-up * metal : clean-up kernel code * metal : minor * tests : remove benchmarks ggml-ci * ggml : fix avx512 const correctness ggml-ci * ggml : fix soft_max with bias on CPU ggml-ci * common : print --flash-attn in help * ggml : fix num dimensions in ggml_flash_attn_ext * llama : force disable flash attention for incompatible models * ggml : ggml_soft_max support F16/F32 mask/pos ggml-ci * cuda : uint -> uint32_t * cuda : "constexpr dim3" -> "const dim3" ggml-ci * cuda : try to fix __hgt2_mask ggml-ci * ggml : add TODO's for F16/F32 mask/pos support in other backends * llama : replace bool need_kq_pos with use_alibi * llama : prep ALiBi support for BERT models ggml-ci * llama : fix n_batch requirements ggml-ci * cont * server : add help for --flash-attn arg * llama : disable FA for AMD * tests : remove TMP_ATTN_BENCH ggml-ci * llama : support save/load state with FA enabled ggml-ci * ci : add CUDA save-load-state tests ggml-ci * llama : llama_kv_cache_clear zeroes data + fix save-load seq ggml-ci * llama : fix copy-paste errors, add TODO * llama : disallow incompatible states * llama : update llama_state_get_size after v_trans field * metal : remove tmp log * llama : add static reminder for llama_state_get_size * metal : fix max nsg ggml-ci * ci : fix arg order ggml-ci --------- Co-authored-by: Johannes Gäßler <johannesg@5d6.de> Co-authored-by: Pierrick HYMBERT <pierrick.hymbert@gmail.com>
2024-04-30 11:16:08 +02:00
template<typename T>
kernel void kernel_soft_max_4(
ggml : add Flash Attention (#5021) * ggml : add ggml_flash_attn_ext API * ggml : fix GQA support in ggml_flash_attn_ext * ggml : online attention (CPU) * metal : initial implementation * metal : f16 precision * metal : reduce branches * metal : specialize for head size * wip : 8 rows per simd group * wip : 4 rows per simd group * wip : template for rows per warp * metal : parallelize across KV size * metal : parallel reduce across heads * metal : efficient flash_attn_f16 implementation * metal : avoid redundant loads of the attention * metal : scale and mask in matrix form * metal : fix comment * llama : avoid ggml_cast, use F32 query * metal : add parallel reduce version (disabled) * metal : move output into local memory + optimize - the result from each simdgroup now stays in the registers - significantly reduced SRAM usage - more efficient skipping of -INF blocks - avoid simdgroup barrier in hot loop - add comments * metal : add tests, fix scaling, support C > 32 * metal : improve precision * ggml : fix f16 mad * metal : minor * metal : support Q > 8 * tests : add ATTN tests * metal : disable buffer allocation logs * tests : more * metal : faster inner loop for C == 32 * metal : fix array initialization * tests : ifdef * ggml : switch to padded F16 mask for ggml_soft_max, ggml_flash_attn_ext * ggml : fix ggml_soft_max mask requirement * cuda : fix soft_max to use correct mask size * cuda : add flash_attn kernel (wip) * metal : optimize softmax for C > 32 * metal : optimize softmax * tests : minor fix * cuda : avoid zeroing fragments * tests : update dims * cuda : fix __hisinf() result check * cuda : avoid warp_reduce for smax * cuda : use int instead of int64_t Noticeably improves performance (thanks to Johannes) * cuda : make loops use the same loop values Thanks Johannes again for the tip * cuda : unroll some of the loops * cuda : avoid __hisinf branches * cuda : use half2 in softmax * cuda : switch to 1 warp for bs > 16 * cuda : speed-up reduce part of the kernel * cuda : unroll Q*K^T loop * cuda : fix -INF block check * cuda : simplify softmax * cuda : fix matrix names * cuda : minor * llama : adapt to F16 KQ_pos * llama : adapt new models to F16 KQ_mask * ggml : fix F16 store (ARM NEON) * llama : fix type of KQ_mask and KQ_pos * ggml : fix CPU soft_max * tests : add hs=256 * cuda : fix build * metal : improve perf via smaller int registers * cuda : adapt soft_max to F16 mask and pos * CUDA: faster FlashAttention, kernel for bs == 1 * 16 cols for Phi-2 * no vec for hs, no hs==256 ncols==32 for Volta * adjust kernel selection logic * 4 warps, 256 stride for all D * no ncols == 64 * Multiple parallel blocks for batch size 1 * fix compile warnings * fix excessive KQ_b loads * fix cmake build * fix KV cache padding, NaN from INFINITY (#6438) * llama : flash_attn cparam + fix defrag * server: support flash_attn param * server: bench: enable flash_attn param * CUDA: refactor host code, dyn. par. blocks * fix flash_attn_vec_f16 race condition * flush softmax exp below threshold to 0 * store temp KQ in registers * Calculate KQ as FP32 if KQV has GGML_PREC_F32 * Add __hgt2_mask implementation for CUDA 11 * fix KQ FP32 precision fpr parallel_blocks > 1 * llama-bench : add -fa,--flash-attn arg * metal : add BS=1 kernel for flash attention (#6508) * metal : add BS=1 kernel for flash attention (wip) * metal : support more than 1 warps * metal : opts * metal : opt * metal : switch to parallel reduce * metal : reduce registers * metal : simplify * metal : initial FA vec kernel * metal : use F32 attention accumulators * batched-bench : add fattn arg * llama : simplify llama_build_kv_store ggml-ci * llama : adapt build_olmo to changes * ggml : fix arm fp16 store on windows * metal : clean-up * metal : clean-up kernel code * metal : minor * tests : remove benchmarks ggml-ci * ggml : fix avx512 const correctness ggml-ci * ggml : fix soft_max with bias on CPU ggml-ci * common : print --flash-attn in help * ggml : fix num dimensions in ggml_flash_attn_ext * llama : force disable flash attention for incompatible models * ggml : ggml_soft_max support F16/F32 mask/pos ggml-ci * cuda : uint -> uint32_t * cuda : "constexpr dim3" -> "const dim3" ggml-ci * cuda : try to fix __hgt2_mask ggml-ci * ggml : add TODO's for F16/F32 mask/pos support in other backends * llama : replace bool need_kq_pos with use_alibi * llama : prep ALiBi support for BERT models ggml-ci * llama : fix n_batch requirements ggml-ci * cont * server : add help for --flash-attn arg * llama : disable FA for AMD * tests : remove TMP_ATTN_BENCH ggml-ci * llama : support save/load state with FA enabled ggml-ci * ci : add CUDA save-load-state tests ggml-ci * llama : llama_kv_cache_clear zeroes data + fix save-load seq ggml-ci * llama : fix copy-paste errors, add TODO * llama : disallow incompatible states * llama : update llama_state_get_size after v_trans field * metal : remove tmp log * llama : add static reminder for llama_state_get_size * metal : fix max nsg ggml-ci * ci : fix arg order ggml-ci --------- Co-authored-by: Johannes Gäßler <johannesg@5d6.de> Co-authored-by: Pierrick HYMBERT <pierrick.hymbert@gmail.com>
2024-04-30 11:16:08 +02:00
device const char * src0,
device const char * src1,
device const char * src2,
device char * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
constant float & scale,
constant float & max_bias,
constant float & m0,
constant float & m1,
constant uint32_t & n_head_log2,
threadgroup float * buf [[threadgroup(0)]],
uint tgpig[[threadgroup_position_in_grid]],
uint tpitg[[thread_position_in_threadgroup]],
uint sgitg[[simdgroup_index_in_threadgroup]],
uint tiisg[[thread_index_in_simdgroup]],
uint ntg[[threads_per_threadgroup]]) {
const int64_t i03 = (tgpig) / (ne02*ne01);
const int64_t i02 = (tgpig - i03*ne02*ne01) / ne01;
const int64_t i01 = (tgpig - i03*ne02*ne01 - i02*ne01);
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
ggml : add Flash Attention (#5021) * ggml : add ggml_flash_attn_ext API * ggml : fix GQA support in ggml_flash_attn_ext * ggml : online attention (CPU) * metal : initial implementation * metal : f16 precision * metal : reduce branches * metal : specialize for head size * wip : 8 rows per simd group * wip : 4 rows per simd group * wip : template for rows per warp * metal : parallelize across KV size * metal : parallel reduce across heads * metal : efficient flash_attn_f16 implementation * metal : avoid redundant loads of the attention * metal : scale and mask in matrix form * metal : fix comment * llama : avoid ggml_cast, use F32 query * metal : add parallel reduce version (disabled) * metal : move output into local memory + optimize - the result from each simdgroup now stays in the registers - significantly reduced SRAM usage - more efficient skipping of -INF blocks - avoid simdgroup barrier in hot loop - add comments * metal : add tests, fix scaling, support C > 32 * metal : improve precision * ggml : fix f16 mad * metal : minor * metal : support Q > 8 * tests : add ATTN tests * metal : disable buffer allocation logs * tests : more * metal : faster inner loop for C == 32 * metal : fix array initialization * tests : ifdef * ggml : switch to padded F16 mask for ggml_soft_max, ggml_flash_attn_ext * ggml : fix ggml_soft_max mask requirement * cuda : fix soft_max to use correct mask size * cuda : add flash_attn kernel (wip) * metal : optimize softmax for C > 32 * metal : optimize softmax * tests : minor fix * cuda : avoid zeroing fragments * tests : update dims * cuda : fix __hisinf() result check * cuda : avoid warp_reduce for smax * cuda : use int instead of int64_t Noticeably improves performance (thanks to Johannes) * cuda : make loops use the same loop values Thanks Johannes again for the tip * cuda : unroll some of the loops * cuda : avoid __hisinf branches * cuda : use half2 in softmax * cuda : switch to 1 warp for bs > 16 * cuda : speed-up reduce part of the kernel * cuda : unroll Q*K^T loop * cuda : fix -INF block check * cuda : simplify softmax * cuda : fix matrix names * cuda : minor * llama : adapt to F16 KQ_pos * llama : adapt new models to F16 KQ_mask * ggml : fix F16 store (ARM NEON) * llama : fix type of KQ_mask and KQ_pos * ggml : fix CPU soft_max * tests : add hs=256 * cuda : fix build * metal : improve perf via smaller int registers * cuda : adapt soft_max to F16 mask and pos * CUDA: faster FlashAttention, kernel for bs == 1 * 16 cols for Phi-2 * no vec for hs, no hs==256 ncols==32 for Volta * adjust kernel selection logic * 4 warps, 256 stride for all D * no ncols == 64 * Multiple parallel blocks for batch size 1 * fix compile warnings * fix excessive KQ_b loads * fix cmake build * fix KV cache padding, NaN from INFINITY (#6438) * llama : flash_attn cparam + fix defrag * server: support flash_attn param * server: bench: enable flash_attn param * CUDA: refactor host code, dyn. par. blocks * fix flash_attn_vec_f16 race condition * flush softmax exp below threshold to 0 * store temp KQ in registers * Calculate KQ as FP32 if KQV has GGML_PREC_F32 * Add __hgt2_mask implementation for CUDA 11 * fix KQ FP32 precision fpr parallel_blocks > 1 * llama-bench : add -fa,--flash-attn arg * metal : add BS=1 kernel for flash attention (#6508) * metal : add BS=1 kernel for flash attention (wip) * metal : support more than 1 warps * metal : opts * metal : opt * metal : switch to parallel reduce * metal : reduce registers * metal : simplify * metal : initial FA vec kernel * metal : use F32 attention accumulators * batched-bench : add fattn arg * llama : simplify llama_build_kv_store ggml-ci * llama : adapt build_olmo to changes * ggml : fix arm fp16 store on windows * metal : clean-up * metal : clean-up kernel code * metal : minor * tests : remove benchmarks ggml-ci * ggml : fix avx512 const correctness ggml-ci * ggml : fix soft_max with bias on CPU ggml-ci * common : print --flash-attn in help * ggml : fix num dimensions in ggml_flash_attn_ext * llama : force disable flash attention for incompatible models * ggml : ggml_soft_max support F16/F32 mask/pos ggml-ci * cuda : uint -> uint32_t * cuda : "constexpr dim3" -> "const dim3" ggml-ci * cuda : try to fix __hgt2_mask ggml-ci * ggml : add TODO's for F16/F32 mask/pos support in other backends * llama : replace bool need_kq_pos with use_alibi * llama : prep ALiBi support for BERT models ggml-ci * llama : fix n_batch requirements ggml-ci * cont * server : add help for --flash-attn arg * llama : disable FA for AMD * tests : remove TMP_ATTN_BENCH ggml-ci * llama : support save/load state with FA enabled ggml-ci * ci : add CUDA save-load-state tests ggml-ci * llama : llama_kv_cache_clear zeroes data + fix save-load seq ggml-ci * llama : fix copy-paste errors, add TODO * llama : disallow incompatible states * llama : update llama_state_get_size after v_trans field * metal : remove tmp log * llama : add static reminder for llama_state_get_size * metal : fix max nsg ggml-ci * ci : fix arg order ggml-ci --------- Co-authored-by: Johannes Gäßler <johannesg@5d6.de> Co-authored-by: Pierrick HYMBERT <pierrick.hymbert@gmail.com>
2024-04-30 11:16:08 +02:00
device const float4 * psrc4 = (device const float4 *) src0 + (i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00)/4;
device const T * pmask = src1 != src0 ? (device const T *) src1 + i01*ne00/4 : nullptr;
device const T * ppos = src2 != src0 ? (device const T *) src2 : nullptr;
device float4 * pdst4 = (device float4 *) dst + (i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00)/4;
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
float slope = 0.0f;
if (max_bias > 0.0f) {
const int64_t h = i02;
const float base = h < n_head_log2 ? m0 : m1;
const int exp = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
slope = pow(base, exp);
}
// parallel max
float4 lmax4 = -INFINITY;
for (int i00 = tpitg; i00 < ne00/4; i00 += ntg) {
ggml : add Flash Attention (#5021) * ggml : add ggml_flash_attn_ext API * ggml : fix GQA support in ggml_flash_attn_ext * ggml : online attention (CPU) * metal : initial implementation * metal : f16 precision * metal : reduce branches * metal : specialize for head size * wip : 8 rows per simd group * wip : 4 rows per simd group * wip : template for rows per warp * metal : parallelize across KV size * metal : parallel reduce across heads * metal : efficient flash_attn_f16 implementation * metal : avoid redundant loads of the attention * metal : scale and mask in matrix form * metal : fix comment * llama : avoid ggml_cast, use F32 query * metal : add parallel reduce version (disabled) * metal : move output into local memory + optimize - the result from each simdgroup now stays in the registers - significantly reduced SRAM usage - more efficient skipping of -INF blocks - avoid simdgroup barrier in hot loop - add comments * metal : add tests, fix scaling, support C > 32 * metal : improve precision * ggml : fix f16 mad * metal : minor * metal : support Q > 8 * tests : add ATTN tests * metal : disable buffer allocation logs * tests : more * metal : faster inner loop for C == 32 * metal : fix array initialization * tests : ifdef * ggml : switch to padded F16 mask for ggml_soft_max, ggml_flash_attn_ext * ggml : fix ggml_soft_max mask requirement * cuda : fix soft_max to use correct mask size * cuda : add flash_attn kernel (wip) * metal : optimize softmax for C > 32 * metal : optimize softmax * tests : minor fix * cuda : avoid zeroing fragments * tests : update dims * cuda : fix __hisinf() result check * cuda : avoid warp_reduce for smax * cuda : use int instead of int64_t Noticeably improves performance (thanks to Johannes) * cuda : make loops use the same loop values Thanks Johannes again for the tip * cuda : unroll some of the loops * cuda : avoid __hisinf branches * cuda : use half2 in softmax * cuda : switch to 1 warp for bs > 16 * cuda : speed-up reduce part of the kernel * cuda : unroll Q*K^T loop * cuda : fix -INF block check * cuda : simplify softmax * cuda : fix matrix names * cuda : minor * llama : adapt to F16 KQ_pos * llama : adapt new models to F16 KQ_mask * ggml : fix F16 store (ARM NEON) * llama : fix type of KQ_mask and KQ_pos * ggml : fix CPU soft_max * tests : add hs=256 * cuda : fix build * metal : improve perf via smaller int registers * cuda : adapt soft_max to F16 mask and pos * CUDA: faster FlashAttention, kernel for bs == 1 * 16 cols for Phi-2 * no vec for hs, no hs==256 ncols==32 for Volta * adjust kernel selection logic * 4 warps, 256 stride for all D * no ncols == 64 * Multiple parallel blocks for batch size 1 * fix compile warnings * fix excessive KQ_b loads * fix cmake build * fix KV cache padding, NaN from INFINITY (#6438) * llama : flash_attn cparam + fix defrag * server: support flash_attn param * server: bench: enable flash_attn param * CUDA: refactor host code, dyn. par. blocks * fix flash_attn_vec_f16 race condition * flush softmax exp below threshold to 0 * store temp KQ in registers * Calculate KQ as FP32 if KQV has GGML_PREC_F32 * Add __hgt2_mask implementation for CUDA 11 * fix KQ FP32 precision fpr parallel_blocks > 1 * llama-bench : add -fa,--flash-attn arg * metal : add BS=1 kernel for flash attention (#6508) * metal : add BS=1 kernel for flash attention (wip) * metal : support more than 1 warps * metal : opts * metal : opt * metal : switch to parallel reduce * metal : reduce registers * metal : simplify * metal : initial FA vec kernel * metal : use F32 attention accumulators * batched-bench : add fattn arg * llama : simplify llama_build_kv_store ggml-ci * llama : adapt build_olmo to changes * ggml : fix arm fp16 store on windows * metal : clean-up * metal : clean-up kernel code * metal : minor * tests : remove benchmarks ggml-ci * ggml : fix avx512 const correctness ggml-ci * ggml : fix soft_max with bias on CPU ggml-ci * common : print --flash-attn in help * ggml : fix num dimensions in ggml_flash_attn_ext * llama : force disable flash attention for incompatible models * ggml : ggml_soft_max support F16/F32 mask/pos ggml-ci * cuda : uint -> uint32_t * cuda : "constexpr dim3" -> "const dim3" ggml-ci * cuda : try to fix __hgt2_mask ggml-ci * ggml : add TODO's for F16/F32 mask/pos support in other backends * llama : replace bool need_kq_pos with use_alibi * llama : prep ALiBi support for BERT models ggml-ci * llama : fix n_batch requirements ggml-ci * cont * server : add help for --flash-attn arg * llama : disable FA for AMD * tests : remove TMP_ATTN_BENCH ggml-ci * llama : support save/load state with FA enabled ggml-ci * ci : add CUDA save-load-state tests ggml-ci * llama : llama_kv_cache_clear zeroes data + fix save-load seq ggml-ci * llama : fix copy-paste errors, add TODO * llama : disallow incompatible states * llama : update llama_state_get_size after v_trans field * metal : remove tmp log * llama : add static reminder for llama_state_get_size * metal : fix max nsg ggml-ci * ci : fix arg order ggml-ci --------- Co-authored-by: Johannes Gäßler <johannesg@5d6.de> Co-authored-by: Pierrick HYMBERT <pierrick.hymbert@gmail.com>
2024-04-30 11:16:08 +02:00
lmax4 = fmax(lmax4, psrc4[i00]*scale + (float4)((pmask ? pmask[i00] : 0.0f) + (ppos ? slope*ppos[i00] : 0.0f)));
}
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
const float lmax = MAX(MAX(lmax4[0], lmax4[1]), MAX(lmax4[2], lmax4[3]));
float max_val = simd_max(lmax);
if (ntg > N_SIMDWIDTH) {
if (sgitg == 0) {
buf[tiisg] = -INFINITY;
}
threadgroup_barrier(mem_flags::mem_threadgroup);
if (tiisg == 0) {
buf[sgitg] = max_val;
}
threadgroup_barrier(mem_flags::mem_threadgroup);
max_val = buf[tiisg];
max_val = simd_max(max_val);
}
// parallel sum
float4 lsum4 = 0.0f;
for (int i00 = tpitg; i00 < ne00/4; i00 += ntg) {
ggml : add Flash Attention (#5021) * ggml : add ggml_flash_attn_ext API * ggml : fix GQA support in ggml_flash_attn_ext * ggml : online attention (CPU) * metal : initial implementation * metal : f16 precision * metal : reduce branches * metal : specialize for head size * wip : 8 rows per simd group * wip : 4 rows per simd group * wip : template for rows per warp * metal : parallelize across KV size * metal : parallel reduce across heads * metal : efficient flash_attn_f16 implementation * metal : avoid redundant loads of the attention * metal : scale and mask in matrix form * metal : fix comment * llama : avoid ggml_cast, use F32 query * metal : add parallel reduce version (disabled) * metal : move output into local memory + optimize - the result from each simdgroup now stays in the registers - significantly reduced SRAM usage - more efficient skipping of -INF blocks - avoid simdgroup barrier in hot loop - add comments * metal : add tests, fix scaling, support C > 32 * metal : improve precision * ggml : fix f16 mad * metal : minor * metal : support Q > 8 * tests : add ATTN tests * metal : disable buffer allocation logs * tests : more * metal : faster inner loop for C == 32 * metal : fix array initialization * tests : ifdef * ggml : switch to padded F16 mask for ggml_soft_max, ggml_flash_attn_ext * ggml : fix ggml_soft_max mask requirement * cuda : fix soft_max to use correct mask size * cuda : add flash_attn kernel (wip) * metal : optimize softmax for C > 32 * metal : optimize softmax * tests : minor fix * cuda : avoid zeroing fragments * tests : update dims * cuda : fix __hisinf() result check * cuda : avoid warp_reduce for smax * cuda : use int instead of int64_t Noticeably improves performance (thanks to Johannes) * cuda : make loops use the same loop values Thanks Johannes again for the tip * cuda : unroll some of the loops * cuda : avoid __hisinf branches * cuda : use half2 in softmax * cuda : switch to 1 warp for bs > 16 * cuda : speed-up reduce part of the kernel * cuda : unroll Q*K^T loop * cuda : fix -INF block check * cuda : simplify softmax * cuda : fix matrix names * cuda : minor * llama : adapt to F16 KQ_pos * llama : adapt new models to F16 KQ_mask * ggml : fix F16 store (ARM NEON) * llama : fix type of KQ_mask and KQ_pos * ggml : fix CPU soft_max * tests : add hs=256 * cuda : fix build * metal : improve perf via smaller int registers * cuda : adapt soft_max to F16 mask and pos * CUDA: faster FlashAttention, kernel for bs == 1 * 16 cols for Phi-2 * no vec for hs, no hs==256 ncols==32 for Volta * adjust kernel selection logic * 4 warps, 256 stride for all D * no ncols == 64 * Multiple parallel blocks for batch size 1 * fix compile warnings * fix excessive KQ_b loads * fix cmake build * fix KV cache padding, NaN from INFINITY (#6438) * llama : flash_attn cparam + fix defrag * server: support flash_attn param * server: bench: enable flash_attn param * CUDA: refactor host code, dyn. par. blocks * fix flash_attn_vec_f16 race condition * flush softmax exp below threshold to 0 * store temp KQ in registers * Calculate KQ as FP32 if KQV has GGML_PREC_F32 * Add __hgt2_mask implementation for CUDA 11 * fix KQ FP32 precision fpr parallel_blocks > 1 * llama-bench : add -fa,--flash-attn arg * metal : add BS=1 kernel for flash attention (#6508) * metal : add BS=1 kernel for flash attention (wip) * metal : support more than 1 warps * metal : opts * metal : opt * metal : switch to parallel reduce * metal : reduce registers * metal : simplify * metal : initial FA vec kernel * metal : use F32 attention accumulators * batched-bench : add fattn arg * llama : simplify llama_build_kv_store ggml-ci * llama : adapt build_olmo to changes * ggml : fix arm fp16 store on windows * metal : clean-up * metal : clean-up kernel code * metal : minor * tests : remove benchmarks ggml-ci * ggml : fix avx512 const correctness ggml-ci * ggml : fix soft_max with bias on CPU ggml-ci * common : print --flash-attn in help * ggml : fix num dimensions in ggml_flash_attn_ext * llama : force disable flash attention for incompatible models * ggml : ggml_soft_max support F16/F32 mask/pos ggml-ci * cuda : uint -> uint32_t * cuda : "constexpr dim3" -> "const dim3" ggml-ci * cuda : try to fix __hgt2_mask ggml-ci * ggml : add TODO's for F16/F32 mask/pos support in other backends * llama : replace bool need_kq_pos with use_alibi * llama : prep ALiBi support for BERT models ggml-ci * llama : fix n_batch requirements ggml-ci * cont * server : add help for --flash-attn arg * llama : disable FA for AMD * tests : remove TMP_ATTN_BENCH ggml-ci * llama : support save/load state with FA enabled ggml-ci * ci : add CUDA save-load-state tests ggml-ci * llama : llama_kv_cache_clear zeroes data + fix save-load seq ggml-ci * llama : fix copy-paste errors, add TODO * llama : disallow incompatible states * llama : update llama_state_get_size after v_trans field * metal : remove tmp log * llama : add static reminder for llama_state_get_size * metal : fix max nsg ggml-ci * ci : fix arg order ggml-ci --------- Co-authored-by: Johannes Gäßler <johannesg@5d6.de> Co-authored-by: Pierrick HYMBERT <pierrick.hymbert@gmail.com>
2024-04-30 11:16:08 +02:00
const float4 exp_psrc4 = exp((psrc4[i00]*scale + (float4)((pmask ? pmask[i00] : 0.0f) + (ppos ? slope*ppos[i00] : 0.0f))) - max_val);
lsum4 += exp_psrc4;
pdst4[i00] = exp_psrc4;
}
const float lsum = lsum4[0] + lsum4[1] + lsum4[2] + lsum4[3];
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 13:04:25 +01:00
// This barrier fixes a failing test
// ref: https://github.com/ggerganov/ggml/pull/621#discussion_r1425156335
threadgroup_barrier(mem_flags::mem_none);
float sum = simd_sum(lsum);
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 13:04:25 +01:00
if (ntg > N_SIMDWIDTH) {
if (sgitg == 0) {
buf[tiisg] = 0.0f;
}
threadgroup_barrier(mem_flags::mem_threadgroup);
if (tiisg == 0) {
buf[sgitg] = sum;
}
threadgroup_barrier(mem_flags::mem_threadgroup);
sum = buf[tiisg];
sum = simd_sum(sum);
}
const float inv_sum = 1.0f/sum;
for (int i00 = tpitg; i00 < ne00/4; i00 += ntg) {
pdst4[i00] *= inv_sum;
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
}
}
ggml : add Flash Attention (#5021) * ggml : add ggml_flash_attn_ext API * ggml : fix GQA support in ggml_flash_attn_ext * ggml : online attention (CPU) * metal : initial implementation * metal : f16 precision * metal : reduce branches * metal : specialize for head size * wip : 8 rows per simd group * wip : 4 rows per simd group * wip : template for rows per warp * metal : parallelize across KV size * metal : parallel reduce across heads * metal : efficient flash_attn_f16 implementation * metal : avoid redundant loads of the attention * metal : scale and mask in matrix form * metal : fix comment * llama : avoid ggml_cast, use F32 query * metal : add parallel reduce version (disabled) * metal : move output into local memory + optimize - the result from each simdgroup now stays in the registers - significantly reduced SRAM usage - more efficient skipping of -INF blocks - avoid simdgroup barrier in hot loop - add comments * metal : add tests, fix scaling, support C > 32 * metal : improve precision * ggml : fix f16 mad * metal : minor * metal : support Q > 8 * tests : add ATTN tests * metal : disable buffer allocation logs * tests : more * metal : faster inner loop for C == 32 * metal : fix array initialization * tests : ifdef * ggml : switch to padded F16 mask for ggml_soft_max, ggml_flash_attn_ext * ggml : fix ggml_soft_max mask requirement * cuda : fix soft_max to use correct mask size * cuda : add flash_attn kernel (wip) * metal : optimize softmax for C > 32 * metal : optimize softmax * tests : minor fix * cuda : avoid zeroing fragments * tests : update dims * cuda : fix __hisinf() result check * cuda : avoid warp_reduce for smax * cuda : use int instead of int64_t Noticeably improves performance (thanks to Johannes) * cuda : make loops use the same loop values Thanks Johannes again for the tip * cuda : unroll some of the loops * cuda : avoid __hisinf branches * cuda : use half2 in softmax * cuda : switch to 1 warp for bs > 16 * cuda : speed-up reduce part of the kernel * cuda : unroll Q*K^T loop * cuda : fix -INF block check * cuda : simplify softmax * cuda : fix matrix names * cuda : minor * llama : adapt to F16 KQ_pos * llama : adapt new models to F16 KQ_mask * ggml : fix F16 store (ARM NEON) * llama : fix type of KQ_mask and KQ_pos * ggml : fix CPU soft_max * tests : add hs=256 * cuda : fix build * metal : improve perf via smaller int registers * cuda : adapt soft_max to F16 mask and pos * CUDA: faster FlashAttention, kernel for bs == 1 * 16 cols for Phi-2 * no vec for hs, no hs==256 ncols==32 for Volta * adjust kernel selection logic * 4 warps, 256 stride for all D * no ncols == 64 * Multiple parallel blocks for batch size 1 * fix compile warnings * fix excessive KQ_b loads * fix cmake build * fix KV cache padding, NaN from INFINITY (#6438) * llama : flash_attn cparam + fix defrag * server: support flash_attn param * server: bench: enable flash_attn param * CUDA: refactor host code, dyn. par. blocks * fix flash_attn_vec_f16 race condition * flush softmax exp below threshold to 0 * store temp KQ in registers * Calculate KQ as FP32 if KQV has GGML_PREC_F32 * Add __hgt2_mask implementation for CUDA 11 * fix KQ FP32 precision fpr parallel_blocks > 1 * llama-bench : add -fa,--flash-attn arg * metal : add BS=1 kernel for flash attention (#6508) * metal : add BS=1 kernel for flash attention (wip) * metal : support more than 1 warps * metal : opts * metal : opt * metal : switch to parallel reduce * metal : reduce registers * metal : simplify * metal : initial FA vec kernel * metal : use F32 attention accumulators * batched-bench : add fattn arg * llama : simplify llama_build_kv_store ggml-ci * llama : adapt build_olmo to changes * ggml : fix arm fp16 store on windows * metal : clean-up * metal : clean-up kernel code * metal : minor * tests : remove benchmarks ggml-ci * ggml : fix avx512 const correctness ggml-ci * ggml : fix soft_max with bias on CPU ggml-ci * common : print --flash-attn in help * ggml : fix num dimensions in ggml_flash_attn_ext * llama : force disable flash attention for incompatible models * ggml : ggml_soft_max support F16/F32 mask/pos ggml-ci * cuda : uint -> uint32_t * cuda : "constexpr dim3" -> "const dim3" ggml-ci * cuda : try to fix __hgt2_mask ggml-ci * ggml : add TODO's for F16/F32 mask/pos support in other backends * llama : replace bool need_kq_pos with use_alibi * llama : prep ALiBi support for BERT models ggml-ci * llama : fix n_batch requirements ggml-ci * cont * server : add help for --flash-attn arg * llama : disable FA for AMD * tests : remove TMP_ATTN_BENCH ggml-ci * llama : support save/load state with FA enabled ggml-ci * ci : add CUDA save-load-state tests ggml-ci * llama : llama_kv_cache_clear zeroes data + fix save-load seq ggml-ci * llama : fix copy-paste errors, add TODO * llama : disallow incompatible states * llama : update llama_state_get_size after v_trans field * metal : remove tmp log * llama : add static reminder for llama_state_get_size * metal : fix max nsg ggml-ci * ci : fix arg order ggml-ci --------- Co-authored-by: Johannes Gäßler <johannesg@5d6.de> Co-authored-by: Pierrick HYMBERT <pierrick.hymbert@gmail.com>
2024-04-30 11:16:08 +02:00
typedef decltype(kernel_soft_max<float>) kernel_soft_max_t;
typedef decltype(kernel_soft_max_4<float4>) kernel_soft_max_4_t;
template [[host_name("kernel_soft_max_f16")]] kernel kernel_soft_max_t kernel_soft_max<half>;
template [[host_name("kernel_soft_max_f32")]] kernel kernel_soft_max_t kernel_soft_max<float>;
template [[host_name("kernel_soft_max_f16_4")]] kernel kernel_soft_max_4_t kernel_soft_max_4<half4>;
template [[host_name("kernel_soft_max_f32_4")]] kernel kernel_soft_max_4_t kernel_soft_max_4<float4>;
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
kernel void kernel_diag_mask_inf(
device const float * src0,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int & n_past,
uint3 tpig[[thread_position_in_grid]]) {
const int64_t i02 = tpig[2];
const int64_t i01 = tpig[1];
const int64_t i00 = tpig[0];
if (i00 > n_past + i01) {
dst[i02*ne01*ne00 + i01*ne00 + i00] = -INFINITY;
} else {
dst[i02*ne01*ne00 + i01*ne00 + i00] = src0[i02*ne01*ne00 + i01*ne00 + i00];
}
}
kernel void kernel_diag_mask_inf_8(
device const float4 * src0,
device float4 * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int & n_past,
uint3 tpig[[thread_position_in_grid]]) {
const int64_t i = 2*tpig[0];
dst[i+0] = src0[i+0];
dst[i+1] = src0[i+1];
int64_t i4 = 4*i;
const int64_t i02 = i4/(ne00*ne01); i4 -= i02*ne00*ne01;
const int64_t i01 = i4/(ne00); i4 -= i01*ne00;
const int64_t i00 = i4;
for (int k = 3; k >= 0; --k) {
if (i00 + 4 + k <= n_past + i01) {
break;
}
dst[i+1][k] = -INFINITY;
if (i00 + k > n_past + i01) {
dst[i][k] = -INFINITY;
}
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
}
}
kernel void kernel_norm(
device const void * src0,
device float * dst,
constant int64_t & ne00,
constant uint64_t & nb01,
constant float & eps,
threadgroup float * sum [[threadgroup(0)]],
uint tgpig[[threadgroup_position_in_grid]],
uint tpitg[[thread_position_in_threadgroup]],
uint ntg[[threads_per_threadgroup]]) {
device const float * x = (device const float *) ((device const char *) src0 + tgpig*nb01);
// MEAN
// parallel sum
sum[tpitg] = 0.0f;
for (int i00 = tpitg; i00 < ne00; i00 += ntg) {
sum[tpitg] += x[i00];
}
// reduce
threadgroup_barrier(mem_flags::mem_threadgroup);
for (uint i = ntg/2; i > 0; i /= 2) {
if (tpitg < i) {
sum[tpitg] += sum[tpitg + i];
}
threadgroup_barrier(mem_flags::mem_threadgroup);
}
const float mean = sum[0] / ne00;
// recenter and VARIANCE
threadgroup_barrier(mem_flags::mem_threadgroup);
device float * y = dst + tgpig*ne00;
sum[tpitg] = 0.0f;
for (int i00 = tpitg; i00 < ne00; i00 += ntg) {
y[i00] = x[i00] - mean;
sum[tpitg] += y[i00] * y[i00];
}
// reduce
threadgroup_barrier(mem_flags::mem_threadgroup);
for (uint i = ntg/2; i > 0; i /= 2) {
if (tpitg < i) {
sum[tpitg] += sum[tpitg + i];
}
threadgroup_barrier(mem_flags::mem_threadgroup);
}
const float variance = sum[0] / ne00;
const float scale = 1.0f/sqrt(variance + eps);
for (int i00 = tpitg; i00 < ne00; i00 += ntg) {
y[i00] = y[i00] * scale;
}
}
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
kernel void kernel_rms_norm(
device const void * src0,
device float * dst,
constant int64_t & ne00,
constant uint64_t & nb01,
constant float & eps,
threadgroup float * buf [[threadgroup(0)]],
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
uint tgpig[[threadgroup_position_in_grid]],
uint tpitg[[thread_position_in_threadgroup]],
uint sgitg[[simdgroup_index_in_threadgroup]],
uint tiisg[[thread_index_in_simdgroup]],
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
uint ntg[[threads_per_threadgroup]]) {
device const float4 * x = (device const float4 *) ((device const char *) src0 + tgpig*nb01);
float4 sumf = 0;
float all_sum = 0;
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
// parallel sum
for (int i00 = tpitg; i00 < ne00/4; i00 += ntg) {
sumf += x[i00] * x[i00];
}
all_sum = sumf[0] + sumf[1] + sumf[2] + sumf[3];
all_sum = simd_sum(all_sum);
if (ntg > N_SIMDWIDTH) {
if (sgitg == 0) {
buf[tiisg] = 0.0f;
}
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
threadgroup_barrier(mem_flags::mem_threadgroup);
if (tiisg == 0) {
buf[sgitg] = all_sum;
}
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
threadgroup_barrier(mem_flags::mem_threadgroup);
all_sum = buf[tiisg];
all_sum = simd_sum(all_sum);
}
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
const float mean = all_sum/ne00;
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
const float scale = 1.0f/sqrt(mean + eps);
device float4 * y = (device float4 *) (dst + tgpig*ne00);
for (int i00 = tpitg; i00 < ne00/4; i00 += ntg) {
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
y[i00] = x[i00] * scale;
}
}
kernel void kernel_group_norm(
device const float * src0,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
constant uint64_t & nb00,
constant uint64_t & nb01,
constant uint64_t & nb02,
constant int32_t & n_groups,
constant float & eps,
threadgroup float * buf [[threadgroup(0)]],
uint tgpig[[threadgroup_position_in_grid]],
uint tpitg[[thread_position_in_threadgroup]],
uint sgitg[[simdgroup_index_in_threadgroup]],
uint tiisg[[thread_index_in_simdgroup]],
uint ntg[[threads_per_threadgroup]]) {
const int64_t ne = ne00*ne01*ne02;
const int64_t gs = ne00*ne01*((ne02 + n_groups - 1) / n_groups);
int start = tgpig * gs;
int end = start + gs;
start += tpitg;
if (end >= ne) {
end = ne;
}
float tmp = 0.0f; // partial sum for thread in warp
for (int j = start; j < end; j += ntg) {
tmp += src0[j];
}
threadgroup_barrier(mem_flags::mem_threadgroup);
tmp = simd_sum(tmp);
if (ntg > N_SIMDWIDTH) {
if (sgitg == 0) {
buf[tiisg] = 0.0f;
}
threadgroup_barrier(mem_flags::mem_threadgroup);
if (tiisg == 0) {
buf[sgitg] = tmp;
}
threadgroup_barrier(mem_flags::mem_threadgroup);
tmp = buf[tiisg];
tmp = simd_sum(tmp);
}
const float mean = tmp / gs;
tmp = 0.0f;
for (int j = start; j < end; j += ntg) {
float xi = src0[j] - mean;
dst[j] = xi;
tmp += xi * xi;
}
tmp = simd_sum(tmp);
if (ntg > N_SIMDWIDTH) {
if (sgitg == 0) {
buf[tiisg] = 0.0f;
}
threadgroup_barrier(mem_flags::mem_threadgroup);
if (tiisg == 0) {
buf[sgitg] = tmp;
}
threadgroup_barrier(mem_flags::mem_threadgroup);
tmp = buf[tiisg];
tmp = simd_sum(tmp);
}
const float variance = tmp / gs;
const float scale = 1.0f/sqrt(variance + eps);
for (int j = start; j < end; j += ntg) {
dst[j] *= scale;
}
}
// function for calculate inner product between half a q4_0 block and 16 floats (yl), sumy is SUM(yl[i])
// il indicates where the q4 quants begin (0 or QK4_0/4)
// we assume that the yl's have been multiplied with the appropriate scale factor
// that corresponds to the missing bit shifts (1, 1/16, 1/256, 1/4096)
inline float block_q_n_dot_y(device const block_q4_0 * qb_curr, float sumy, thread float * yl, int il) {
float d = qb_curr->d;
float2 acc = 0.f;
device const uint16_t * qs = ((device const uint16_t *)qb_curr + 1 + il/2);
for (int i = 0; i < 8; i+=2) {
acc[0] += yl[i + 0] * (qs[i / 2] & 0x000F)
+ yl[i + 1] * (qs[i / 2] & 0x0F00);
acc[1] += yl[i + 8] * (qs[i / 2] & 0x00F0)
+ yl[i + 9] * (qs[i / 2] & 0xF000);
}
return d * (sumy * -8.f + acc[0] + acc[1]);
}
// function for calculate inner product between half a q4_1 block and 16 floats (yl), sumy is SUM(yl[i])
// il indicates where the q4 quants begin (0 or QK4_0/4)
// we assume that the yl's have been multiplied with the appropriate scale factor
// that corresponds to the missing bit shifts (1, 1/16, 1/256, 1/4096)
inline float block_q_n_dot_y(device const block_q4_1 * qb_curr, float sumy, thread float * yl, int il) {
float d = qb_curr->d;
float m = qb_curr->m;
float2 acc = 0.f;
device const uint16_t * qs = ((device const uint16_t *)qb_curr + 2 + il/2);
for (int i = 0; i < 8; i+=2) {
acc[0] += yl[i + 0] * (qs[i / 2] & 0x000F)
+ yl[i + 1] * (qs[i / 2] & 0x0F00);
acc[1] += yl[i + 8] * (qs[i / 2] & 0x00F0)
+ yl[i + 9] * (qs[i / 2] & 0xF000);
}
return d * (acc[0] + acc[1]) + sumy * m;
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
}
// function for calculate inner product between half a q5_0 block and 16 floats (yl), sumy is SUM(yl[i])
// il indicates where the q5 quants begin (0 or QK5_0/4)
// we assume that the yl's have been multiplied with the appropriate scale factor
// that corresponds to the missing bit shifts (1, 1/16, 1/256, 1/4096)
inline float block_q_n_dot_y(device const block_q5_0 * qb_curr, float sumy, thread float * yl, int il) {
float d = qb_curr->d;
float2 acc = 0.f;
device const uint16_t * qs = ((device const uint16_t *)qb_curr + 3 + il/2);
const uint32_t qh = *((device const uint32_t *)qb_curr->qh);
for (int i = 0; i < 8; i+=2) {
acc[0] += yl[i + 0] * ((qs[i / 2] & 0x000F) | ((qh >> (i+0+il ) << 4 ) & 0x00010))
+ yl[i + 1] * ((qs[i / 2] & 0x0F00) | ((qh >> (i+1+il ) << 12) & 0x01000));
acc[1] += yl[i + 8] * ((qs[i / 2] & 0x00F0) | ((qh >> (i+0+il+QK5_0/2) << 8 ) & 0x00100))
+ yl[i + 9] * ((qs[i / 2] & 0xF000) | ((qh >> (i+1+il+QK5_0/2) << 16) & 0x10000));
}
return d * (sumy * -16.f + acc[0] + acc[1]);
}
// function for calculate inner product between half a q5_1 block and 16 floats (yl), sumy is SUM(yl[i])
// il indicates where the q5 quants begin (0 or QK5_1/4)
// we assume that the yl's have been multiplied with the appropriate scale factor
// that corresponds to the missing bit shifts (1, 1/16, 1/256, 1/4096)
inline float block_q_n_dot_y(device const block_q5_1 * qb_curr, float sumy, thread float * yl, int il) {
float d = qb_curr->d;
float m = qb_curr->m;
float2 acc = 0.f;
device const uint16_t * qs = ((device const uint16_t *)qb_curr + 4 + il/2);
const uint32_t qh = *((device const uint32_t *)qb_curr->qh);
for (int i = 0; i < 8; i+=2) {
acc[0] += yl[i + 0] * ((qs[i / 2] & 0x000F) | ((qh >> (i+0+il ) << 4 ) & 0x00010))
+ yl[i + 1] * ((qs[i / 2] & 0x0F00) | ((qh >> (i+1+il ) << 12) & 0x01000));
acc[1] += yl[i + 8] * ((qs[i / 2] & 0x00F0) | ((qh >> (i+0+il+QK5_0/2) << 8 ) & 0x00100))
+ yl[i + 9] * ((qs[i / 2] & 0xF000) | ((qh >> (i+1+il+QK5_0/2) << 16) & 0x10000));
}
return d * (acc[0] + acc[1]) + sumy * m;
}
// putting them in the kernel cause a significant performance penalty
#define N_DST 4 // each SIMD group works on 4 rows
#define N_SIMDGROUP 2 // number of SIMD groups in a thread group
//Note: This is a template, but strictly speaking it only applies to
// quantizations where the block size is 32. It also does not
// guard against the number of rows not being divisible by
// N_DST, so this is another explicit assumption of the implementation.
template<typename block_q_type, int nr, int nsg, int nw>
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 13:04:25 +01:00
void mul_vec_q_n_f32_impl(
device const void * src0,
device const float * src1,
device float * dst,
int64_t ne00,
int64_t ne01,
int64_t ne02,
int64_t ne10,
int64_t ne12,
int64_t ne0,
int64_t ne1,
uint r2,
uint r3,
threadgroup int8_t * shared_values,
uint3 tgpig, uint tiisg, uint sgitg) {
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
const int nb = ne00/QK4_0;
const int r0 = tgpig.x;
const int r1 = tgpig.y;
const int im = tgpig.z;
const int first_row = (r0 * nsg + sgitg) * nr;
const uint i12 = im%ne12;
const uint i13 = im/ne12;
const uint offset0 = first_row * nb + (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
device const block_q_type * x = (device const block_q_type *) src0 + offset0;
device const float * y = (device const float *) src1 + r1*ne10 + im*ne00*ne1;
float yl[16]; // src1 vector cache
float sumf[nr] = {0.f};
const int ix = (tiisg/2);
const int il = (tiisg%2)*8;
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
device const float * yb = y + ix * QK4_0 + il;
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
// each thread in a SIMD group deals with half a block.
for (int ib = ix; ib < nb; ib += nw/2) {
float sumy = 0;
for (int i = 0; i < 8; i += 2) {
sumy += yb[i] + yb[i+1];
yl[i+0] = yb[i+ 0];
yl[i+1] = yb[i+ 1]/256.f;
sumy += yb[i+16] + yb[i+17];
yl[i+8] = yb[i+16]/16.f;
yl[i+9] = yb[i+17]/4096.f;
}
for (int row = 0; row < nr; row++) {
sumf[row] += block_q_n_dot_y(x+ib+row*nb, sumy, yl, il);
}
yb += QK4_0 * 16;
}
for (int row = 0; row < nr; ++row) {
const float tot = simd_sum(sumf[row]);
if (tiisg == 0 && first_row + row < ne01) {
dst[im*ne0*ne1 + r1*ne0 + first_row + row] = tot;
}
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
}
}
kernel void kernel_mul_mv_q4_0_f32(
device const void * src0,
device const float * src1,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
constant uint64_t & nb00,
constant uint64_t & nb01,
constant uint64_t & nb02,
constant int64_t & ne10,
constant int64_t & ne11,
constant int64_t & ne12,
constant uint64_t & nb10,
constant uint64_t & nb11,
constant uint64_t & nb12,
constant int64_t & ne0,
constant int64_t & ne1,
constant uint & r2,
constant uint & r3,
uint3 tgpig[[threadgroup_position_in_grid]],
uint tiisg[[thread_index_in_simdgroup]],
uint sgitg[[simdgroup_index_in_threadgroup]]) {
mul_vec_q_n_f32_impl<block_q4_0, N_DST, N_SIMDGROUP, N_SIMDWIDTH>(src0,src1,dst,ne00,ne01,ne02,ne10,ne12,ne0,ne1,r2,r3,nullptr,tgpig,tiisg,sgitg);
}
kernel void kernel_mul_mv_q4_1_f32(
device const void * src0,
device const float * src1,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
constant uint64_t & nb00,
constant uint64_t & nb01,
constant uint64_t & nb02,
constant int64_t & ne10,
constant int64_t & ne11,
constant int64_t & ne12,
constant uint64_t & nb10,
constant uint64_t & nb11,
constant uint64_t & nb12,
constant int64_t & ne0,
constant int64_t & ne1,
constant uint & r2,
constant uint & r3,
uint3 tgpig[[threadgroup_position_in_grid]],
uint tiisg[[thread_index_in_simdgroup]],
uint sgitg[[simdgroup_index_in_threadgroup]]) {
mul_vec_q_n_f32_impl<block_q4_1, N_DST, N_SIMDGROUP, N_SIMDWIDTH>(src0,src1,dst,ne00,ne01,ne02,ne10,ne12,ne0,ne1,r2,r3,nullptr,tgpig,tiisg,sgitg);
}
kernel void kernel_mul_mv_q5_0_f32(
device const void * src0,
device const float * src1,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
constant uint64_t & nb00,
constant uint64_t & nb01,
constant uint64_t & nb02,
constant int64_t & ne10,
constant int64_t & ne11,
constant int64_t & ne12,
constant uint64_t & nb10,
constant uint64_t & nb11,
constant uint64_t & nb12,
constant int64_t & ne0,
constant int64_t & ne1,
constant uint & r2,
constant uint & r3,
uint3 tgpig[[threadgroup_position_in_grid]],
uint tiisg[[thread_index_in_simdgroup]],
uint sgitg[[simdgroup_index_in_threadgroup]]) {
mul_vec_q_n_f32_impl<block_q5_0, N_DST, N_SIMDGROUP, N_SIMDWIDTH>(src0,src1,dst,ne00,ne01,ne02,ne10,ne12,ne0,ne1,r2,r3,nullptr,tgpig,tiisg,sgitg);
}
kernel void kernel_mul_mv_q5_1_f32(
device const void * src0,
device const float * src1,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
constant uint64_t & nb00,
constant uint64_t & nb01,
constant uint64_t & nb02,
constant int64_t & ne10,
constant int64_t & ne11,
constant int64_t & ne12,
constant uint64_t & nb10,
constant uint64_t & nb11,
constant uint64_t & nb12,
constant int64_t & ne0,
constant int64_t & ne1,
constant uint & r2,
constant uint & r3,
uint3 tgpig[[threadgroup_position_in_grid]],
uint tiisg[[thread_index_in_simdgroup]],
uint sgitg[[simdgroup_index_in_threadgroup]]) {
mul_vec_q_n_f32_impl<block_q5_1, N_DST, N_SIMDGROUP, N_SIMDWIDTH>(src0,src1,dst,ne00,ne01,ne02,ne10,ne12,ne0,ne1,r2,r3,nullptr,tgpig,tiisg,sgitg);
}
#define NB_Q8_0 8
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 13:04:25 +01:00
void kernel_mul_mv_q8_0_f32_impl(
device const void * src0,
device const float * src1,
device float * dst,
int64_t ne00,
int64_t ne01,
int64_t ne02,
int64_t ne10,
int64_t ne12,
int64_t ne0,
int64_t ne1,
uint r2,
uint r3,
threadgroup int8_t * shared_values,
uint3 tgpig,
uint tiisg,
uint sgitg) {
const int nr = N_DST;
const int nsg = N_SIMDGROUP;
const int nw = N_SIMDWIDTH;
const int nb = ne00/QK8_0;
const int r0 = tgpig.x;
const int r1 = tgpig.y;
const int im = tgpig.z;
const int first_row = (r0 * nsg + sgitg) * nr;
const uint i12 = im%ne12;
const uint i13 = im/ne12;
const uint offset0 = first_row * nb + (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
device const block_q8_0 * x = (device const block_q8_0 *) src0 + offset0;
device const float * y = (device const float *) src1 + r1*ne10 + im*ne00*ne1;
float yl[NB_Q8_0];
float sumf[nr]={0.f};
const int ix = tiisg/4;
const int il = tiisg%4;
device const float * yb = y + ix * QK8_0 + NB_Q8_0*il;
// each thread in a SIMD group deals with NB_Q8_0 quants at a time
for (int ib = ix; ib < nb; ib += nw/4) {
for (int i = 0; i < NB_Q8_0; ++i) {
yl[i] = yb[i];
}
for (int row = 0; row < nr; row++) {
device const int8_t * qs = x[ib+row*nb].qs + NB_Q8_0*il;
float sumq = 0.f;
for (int iq = 0; iq < NB_Q8_0; ++iq) {
sumq += qs[iq] * yl[iq];
}
sumf[row] += sumq*x[ib+row*nb].d;
}
yb += NB_Q8_0 * nw;
}
for (int row = 0; row < nr; ++row) {
const float tot = simd_sum(sumf[row]);
if (tiisg == 0 && first_row + row < ne01) {
dst[r1*ne0 + im*ne0*ne1 + first_row + row] = tot;
}
}
}
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 13:04:25 +01:00
[[host_name("kernel_mul_mv_q8_0_f32")]]
kernel void kernel_mul_mv_q8_0_f32(
device const void * src0,
device const float * src1,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
constant uint64_t & nb00,
constant uint64_t & nb01,
constant uint64_t & nb02,
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 13:04:25 +01:00
constant int64_t & ne10,
constant int64_t & ne11,
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 13:04:25 +01:00
constant int64_t & ne12,
constant uint64_t & nb10,
constant uint64_t & nb11,
constant uint64_t & nb12,
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 13:04:25 +01:00
constant int64_t & ne0,
constant int64_t & ne1,
constant uint & r2,
constant uint & r3,
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 13:04:25 +01:00
uint3 tgpig[[threadgroup_position_in_grid]],
uint tiisg[[thread_index_in_simdgroup]],
uint sgitg[[simdgroup_index_in_threadgroup]]) {
kernel_mul_mv_q8_0_f32_impl(src0,src1,dst,ne00,ne01,ne02,ne10,ne12,ne0,ne1,r2,r3,nullptr,tgpig,tiisg,sgitg);
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 13:04:25 +01:00
}
#define N_F32_F32 4
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 13:04:25 +01:00
void kernel_mul_mv_f32_f32_impl(
device const char * src0,
device const char * src1,
device float * dst,
int64_t ne00,
int64_t ne01,
int64_t ne02,
uint64_t nb00,
uint64_t nb01,
uint64_t nb02,
int64_t ne10,
int64_t ne11,
int64_t ne12,
uint64_t nb10,
uint64_t nb11,
uint64_t nb12,
int64_t ne0,
int64_t ne1,
uint r2,
uint r3,
uint3 tgpig,
uint tiisg) {
const int64_t r0 = tgpig.x;
const int64_t rb = tgpig.y*N_F32_F32;
const int64_t im = tgpig.z;
const uint i12 = im%ne12;
const uint i13 = im/ne12;
const uint offset0 = r0*nb01 + (i12/r2)*nb02 + (i13/r3)*nb02*ne02;
device const float * x = (device const float *) (src0 + offset0);
if (ne00 < 128) {
for (int row = 0; row < N_F32_F32; ++row) {
int r1 = rb + row;
if (r1 >= ne11) {
break;
}
device const float * y = (device const float *) (src1 + r1*nb11 + im*nb12);
float sumf = 0;
for (int i = tiisg; i < ne00; i += 32) {
sumf += (float) x[i] * (float) y[i];
}
float all_sum = simd_sum(sumf);
if (tiisg == 0) {
dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
}
}
} else {
device const float4 * x4 = (device const float4 *)x;
for (int row = 0; row < N_F32_F32; ++row) {
int r1 = rb + row;
if (r1 >= ne11) {
break;
}
device const float * y = (device const float *) (src1 + r1*nb11 + im*nb12);
device const float4 * y4 = (device const float4 *) y;
float sumf = 0;
for (int i = tiisg; i < ne00/4; i += 32) {
for (int k = 0; k < 4; ++k) sumf += (float) x4[i][k] * y4[i][k];
}
float all_sum = simd_sum(sumf);
if (tiisg == 0) {
for (int i = 4*(ne00/4); i < ne00; ++i) all_sum += (float) x[i] * y[i];
dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
}
}
}
}
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 13:04:25 +01:00
[[host_name("kernel_mul_mv_f32_f32")]]
kernel void kernel_mul_mv_f32_f32(
device const char * src0,
device const char * src1,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
constant uint64_t & nb00,
constant uint64_t & nb01,
constant uint64_t & nb02,
constant int64_t & ne10,
constant int64_t & ne11,
constant int64_t & ne12,
constant uint64_t & nb10,
constant uint64_t & nb11,
constant uint64_t & nb12,
constant int64_t & ne0,
constant int64_t & ne1,
constant uint & r2,
constant uint & r3,
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 13:04:25 +01:00
uint3 tgpig[[threadgroup_position_in_grid]],
uint tiisg[[thread_index_in_simdgroup]]) {
kernel_mul_mv_f32_f32_impl(src0, src1, dst, ne00, ne01, ne02, nb00, nb01, nb02, ne10, ne11, ne12, nb10, nb11, nb12, ne0, ne1, r2, r3, tgpig, tiisg);
}
#define N_F16_F16 4
kernel void kernel_mul_mv_f16_f16(
device const char * src0,
device const char * src1,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
constant uint64_t & nb00,
constant uint64_t & nb01,
constant uint64_t & nb02,
constant int64_t & ne10,
constant int64_t & ne11,
constant int64_t & ne12,
constant uint64_t & nb10,
constant uint64_t & nb11,
constant uint64_t & nb12,
constant int64_t & ne0,
constant int64_t & ne1,
constant uint & r2,
constant uint & r3,
uint3 tgpig[[threadgroup_position_in_grid]],
uint tiisg[[thread_index_in_simdgroup]]) {
const int64_t r0 = tgpig.x;
const int64_t rb = tgpig.y*N_F16_F16;
const int64_t im = tgpig.z;
const uint i12 = im%ne12;
const uint i13 = im/ne12;
const uint offset0 = r0*nb01 + (i12/r2)*nb02 + (i13/r3)*nb02*ne02;
device const half * x = (device const half *) (src0 + offset0);
if (ne00 < 128) {
for (int row = 0; row < N_F16_F16; ++row) {
int r1 = rb + row;
if (r1 >= ne11) {
break;
}
device const half * y = (device const half *) (src1 + r1*nb11 + im*nb12);
float sumf = 0;
for (int i = tiisg; i < ne00; i += 32) {
sumf += (half) x[i] * (half) y[i];
}
float all_sum = simd_sum(sumf);
if (tiisg == 0) {
dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
}
}
} else {
device const half4 * x4 = (device const half4 *)x;
for (int row = 0; row < N_F16_F16; ++row) {
int r1 = rb + row;
if (r1 >= ne11) {
break;
}
device const half * y = (device const half *) (src1 + r1*nb11 + im*nb12);
device const half4 * y4 = (device const half4 *) y;
float sumf = 0;
for (int i = tiisg; i < ne00/4; i += 32) {
for (int k = 0; k < 4; ++k) sumf += (half) x4[i][k] * y4[i][k];
}
float all_sum = simd_sum(sumf);
if (tiisg == 0) {
for (int i = 4*(ne00/4); i < ne00; ++i) all_sum += (half) x[i] * y[i];
dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
}
}
}
}
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 13:04:25 +01:00
void kernel_mul_mv_f16_f32_1row_impl(
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
device const char * src0,
device const char * src1,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
constant uint64_t & nb00,
constant uint64_t & nb01,
constant uint64_t & nb02,
constant int64_t & ne10,
constant int64_t & ne11,
constant int64_t & ne12,
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
constant uint64_t & nb10,
constant uint64_t & nb11,
constant uint64_t & nb12,
constant int64_t & ne0,
constant int64_t & ne1,
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 13:04:25 +01:00
constant uint & r2,
constant uint & r3,
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
uint3 tgpig[[threadgroup_position_in_grid]],
uint tiisg[[thread_index_in_simdgroup]]) {
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
const int64_t r0 = tgpig.x;
const int64_t r1 = tgpig.y;
const int64_t im = tgpig.z;
const uint i12 = im%ne12;
const uint i13 = im/ne12;
const uint offset0 = r0*nb01 + (i12/r2)*nb02 + (i13/r3)*nb02*ne02;
device const half * x = (device const half *) (src0 + offset0);
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
device const float * y = (device const float *) (src1 + r1*nb11 + im*nb12);
float sumf = 0;
if (ne00 < 128) {
for (int i = tiisg; i < ne00; i += 32) {
sumf += (float) x[i] * (float) y[i];
}
float all_sum = simd_sum(sumf);
if (tiisg == 0) {
dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
}
} else {
device const half4 * x4 = (device const half4 *) x;
device const float4 * y4 = (device const float4 *) y;
for (int i = tiisg; i < ne00/4; i += 32) {
for (int k = 0; k < 4; ++k) sumf += (float)x4[i][k] * y4[i][k];
}
float all_sum = simd_sum(sumf);
if (tiisg == 0) {
for (int i = 4*(ne00/4); i < ne00; ++i) all_sum += (float) x[i] * y[i];
dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
}
}
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 13:04:25 +01:00
}
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 13:04:25 +01:00
[[host_name("kernel_mul_mv_f16_f32_1row")]]
kernel void kernel_mul_mv_f16_f32_1row(
device const char * src0,
device const char * src1,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
constant uint64_t & nb00,
constant uint64_t & nb01,
constant uint64_t & nb02,
constant int64_t & ne10,
constant int64_t & ne11,
constant int64_t & ne12,
constant uint64_t & nb10,
constant uint64_t & nb11,
constant uint64_t & nb12,
constant int64_t & ne0,
constant int64_t & ne1,
constant uint & r2,
constant uint & r3,
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 13:04:25 +01:00
uint3 tgpig[[threadgroup_position_in_grid]],
uint tiisg[[thread_index_in_simdgroup]]) {
kernel_mul_mv_f16_f32_1row_impl(src0, src1, dst, ne00, ne01, ne02, nb00, nb01, nb02, ne10, ne11, ne12, nb10, nb11, nb12, ne0, ne1, r2, r3, tgpig, tiisg);
}
#define N_F16_F32 4
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 13:04:25 +01:00
void kernel_mul_mv_f16_f32_impl(
device const char * src0,
device const char * src1,
device float * dst,
int64_t ne00,
int64_t ne01,
int64_t ne02,
uint64_t nb00,
uint64_t nb01,
uint64_t nb02,
int64_t ne10,
int64_t ne11,
int64_t ne12,
uint64_t nb10,
uint64_t nb11,
uint64_t nb12,
int64_t ne0,
int64_t ne1,
uint r2,
uint r3,
uint3 tgpig,
uint tiisg) {
const int64_t r0 = tgpig.x;
const int64_t rb = tgpig.y*N_F16_F32;
const int64_t im = tgpig.z;
const uint i12 = im%ne12;
const uint i13 = im/ne12;
const uint offset0 = r0*nb01 + (i12/r2)*nb02 + (i13/r3)*nb02*ne02;
device const half * x = (device const half *) (src0 + offset0);
if (ne00 < 128) {
for (int row = 0; row < N_F16_F32; ++row) {
int r1 = rb + row;
if (r1 >= ne11) {
break;
}
device const float * y = (device const float *) (src1 + r1*nb11 + im*nb12);
float sumf = 0;
for (int i = tiisg; i < ne00; i += 32) {
sumf += (float) x[i] * (float) y[i];
}
float all_sum = simd_sum(sumf);
if (tiisg == 0) {
dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
}
}
} else {
device const half4 * x4 = (device const half4 *)x;
for (int row = 0; row < N_F16_F32; ++row) {
int r1 = rb + row;
if (r1 >= ne11) {
break;
}
device const float * y = (device const float *) (src1 + r1*nb11 + im*nb12);
device const float4 * y4 = (device const float4 *) y;
float sumf = 0;
for (int i = tiisg; i < ne00/4; i += 32) {
for (int k = 0; k < 4; ++k) sumf += (float) x4[i][k] * y4[i][k];
}
float all_sum = simd_sum(sumf);
if (tiisg == 0) {
for (int i = 4*(ne00/4); i < ne00; ++i) all_sum += (float) x[i] * y[i];
dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
}
}
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
}
}
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 13:04:25 +01:00
[[host_name("kernel_mul_mv_f16_f32")]]
kernel void kernel_mul_mv_f16_f32(
device const char * src0,
device const char * src1,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
constant uint64_t & nb00,
constant uint64_t & nb01,
constant uint64_t & nb02,
constant int64_t & ne10,
constant int64_t & ne11,
constant int64_t & ne12,
constant uint64_t & nb10,
constant uint64_t & nb11,
constant uint64_t & nb12,
constant int64_t & ne0,
constant int64_t & ne1,
constant uint & r2,
constant uint & r3,
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 13:04:25 +01:00
uint3 tgpig[[threadgroup_position_in_grid]],
uint tiisg[[thread_index_in_simdgroup]]) {
kernel_mul_mv_f16_f32_impl(src0, src1, dst, ne00, ne01, ne02, nb00, nb01, nb02, ne10, ne11, ne12, nb10, nb11, nb12, ne0, ne1, r2, r3, tgpig, tiisg);
}
// Assumes row size (ne00) is a multiple of 4
kernel void kernel_mul_mv_f16_f32_l4(
device const char * src0,
device const char * src1,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
constant uint64_t & nb00,
constant uint64_t & nb01,
constant uint64_t & nb02,
constant int64_t & ne10,
constant int64_t & ne11,
constant int64_t & ne12,
constant uint64_t & nb10,
constant uint64_t & nb11,
constant uint64_t & nb12,
constant int64_t & ne0,
constant int64_t & ne1,
constant uint & r2,
constant uint & r3,
uint3 tgpig[[threadgroup_position_in_grid]],
uint tiisg[[thread_index_in_simdgroup]]) {
const int nrows = ne11;
const int64_t r0 = tgpig.x;
const int64_t im = tgpig.z;
const uint i12 = im%ne12;
const uint i13 = im/ne12;
const uint offset0 = r0*nb01 + (i12/r2)*nb02 + (i13/r3)*nb02*ne02;
device const half4 * x4 = (device const half4 *) (src0 + offset0);
for (int r1 = 0; r1 < nrows; ++r1) {
device const float4 * y4 = (device const float4 *) (src1 + r1*nb11 + im*nb12);
float sumf = 0;
for (int i = tiisg; i < ne00/4; i += 32) {
for (int k = 0; k < 4; ++k) sumf += (float) x4[i][k] * y4[i][k];
}
float all_sum = simd_sum(sumf);
if (tiisg == 0) {
dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
}
}
}
kernel void kernel_alibi_f32(
device const float * src0,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
constant int64_t & ne03,
constant uint64_t & nb00,
constant uint64_t & nb01,
constant uint64_t & nb02,
constant uint64_t & nb03,
constant int64_t & ne0,
constant int64_t & ne1,
constant int64_t & ne2,
constant int64_t & ne3,
constant uint64_t & nb0,
constant uint64_t & nb1,
constant uint64_t & nb2,
constant uint64_t & nb3,
constant float & m0,
constant float & m1,
constant int & n_heads_log2_floor,
uint3 tgpig[[threadgroup_position_in_grid]],
uint3 tpitg[[thread_position_in_threadgroup]],
uint3 ntg[[threads_per_threadgroup]]) {
const int64_t i03 = tgpig[2];
const int64_t i02 = tgpig[1];
const int64_t i01 = tgpig[0];
const int64_t n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
const int64_t i3 = n / (ne2*ne1*ne0);
const int64_t i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0);
const int64_t i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0;
//const int64_t i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0);
const int64_t k = i3*ne3 + i2;
float m_k;
if (k < n_heads_log2_floor) {
m_k = pow(m0, k + 1);
} else {
m_k = pow(m1, 2 * (k - n_heads_log2_floor) + 1);
}
device char * dst_row = (device char *) dst + i3*nb3 + i2*nb2 + i1*nb1;
device const char * src_row = (device char *) src0 + i03*nb03 + i02*nb02 + i01*nb01;
for (int64_t i00 = tpitg.x; i00 < ne00; i00 += ntg.x) {
const float src_v = *(device float *)(src_row + i00*nb00);
device float * dst_v = (device float *)(dst_row + i00*nb0);
*dst_v = i00 * m_k + src_v;
}
}
static float rope_yarn_ramp(const float low, const float high, const int i0) {
const float y = (i0 / 2 - low) / max(0.001f, high - low);
return 1.0f - min(1.0f, max(0.0f, y));
}
// YaRN algorithm based on LlamaYaRNScaledRotaryEmbedding.py from https://github.com/jquesnelle/yarn
// MIT licensed. Copyright (c) 2023 Jeffrey Quesnelle and Bowen Peng.
static void rope_yarn(
float theta_extrap, float freq_scale, float corr_dims[2], int64_t i0, float ext_factor, float mscale,
thread float * cos_theta, thread float * sin_theta
) {
// Get n-d rotational scaling corrected for extrapolation
float theta_interp = freq_scale * theta_extrap;
float theta = theta_interp;
if (ext_factor != 0.0f) {
float ramp_mix = rope_yarn_ramp(corr_dims[0], corr_dims[1], i0) * ext_factor;
theta = theta_interp * (1 - ramp_mix) + theta_extrap * ramp_mix;
// Get n-d magnitude scaling corrected for interpolation
mscale *= 1.0f + 0.1f * log(1.0f / freq_scale);
}
*cos_theta = cos(theta) * mscale;
*sin_theta = sin(theta) * mscale;
}
// Apparently solving `n_rot = 2pi * x * base^((2 * max_pos_emb) / n_dims)` for x, we get
// `corr_fac(n_rot) = n_dims * log(max_pos_emb / (n_rot * 2pi)) / (2 * log(base))`
static float rope_yarn_corr_factor(int n_dims, int n_orig_ctx, float n_rot, float base) {
return n_dims * log(n_orig_ctx / (n_rot * 2 * M_PI_F)) / (2 * log(base));
}
static void rope_yarn_corr_dims(
int n_dims, int n_orig_ctx, float freq_base, float beta_fast, float beta_slow, float dims[2]
) {
// start and end correction dims
dims[0] = max(0.0f, floor(rope_yarn_corr_factor(n_dims, n_orig_ctx, beta_fast, freq_base)));
dims[1] = min(n_dims - 1.0f, ceil(rope_yarn_corr_factor(n_dims, n_orig_ctx, beta_slow, freq_base)));
}
llama : custom attention mask + parallel decoding + no context swaps (#3228) * tests : verify that RoPE is "additive" * llama : replace ggml_diag_mask_inf with ggml_add (custom -inf mask) * ggml : ggml_rope now takes a vector with positions instead of n_past * metal : add rope_f16 kernel + optimize cpy kernels * llama : unified KV cache + batch inference API * llama : add new llama_decode() API that works with llama_batch * llama : add cell_max heuristic for more efficient kv_cache * llama : extend llama_kv_cache API * llama : more robust cell_max heuristic + wip shift * metal : disable concurrency optimization * llama : add llama_kv_cache_shift_seq + no more context swaps * llama : apply K-cache roping for Falcon and Baichuan * speculative : fix KV cache management * parallel : example for serving multiple users in parallel * parallel : disable hot-plug to avoid cache fragmentation * fixes : speculative KV cache + llama worst-case graph * llama : extend batch API to select which logits to output * llama : fix worst case graph build * ggml-cuda : update rope implementation for parallel decoding (#3254) * ggml-cuda : update rope implementation for parallel decoding * better solution for p0 computation * fix rope * simpler rope implementation --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * make : add parallel to build + fix static functions in llama.cpp * simple : fix token counting * parallel : various improvements * llama : fix cell_max logic + rename functions * parallel : try smaller batches when the KV cache is fragmented * parallel : fix sequence termination criteria * llama : silence errors KV cache errors * parallel : remove new line from prompt * parallel : process system prompt once + configurable paramters + llama API * parallel : remove question with short answers * parallel : count cache misses * parallel : print misses on each request * parallel : minor * llama : fix n_kv to never become 0 * parallel : rename hot-plug to continuous-batching * llama : improve llama_batch API + simplify parallel example * simple : add parallel decoding support * simple : improve comments + free batch * ggml-cuda : add rope f16, restore performance with parallel decoding (#3272) * ggml-cuda : add rope f16, restore performance * offload KQ_mask with all models * fix rope shift --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * llama : disable MPI for now ggml-ci * train : make KQ_pos memory buffer permanent via dummy scale op * ggml : revert change to ggml_cpy, add ggml_cont_Nd instead (#3275) ggml-ci * parallel : fix bug (extra BOS) + smaller token_prev array * parallel : fix cases where the input prompts can overflow the batch * parallel : add disabled experimental batch chunking in powers of two * llama : llama.h formatting + comments * simple : add README.md * llama : fix kv cache heuristic when context is less than 32 * parallel : fix crash when `-n -1` * llama : simplify returns if/else branches * metal : use mm kernels for batch size > 2 * examples : utilize new llama_get_logits_ith() * examples : add example for batched decoding * examples : do not eval prompt 2 times (close #3348) * server : clear the KV cache beyond n_past before llama_decode * server : avoid context swaps by shifting the KV cache --------- Co-authored-by: slaren <slarengh@gmail.com>
2023-09-28 18:04:36 +02:00
typedef void (rope_t)(
device const void * src0,
device const int32_t * src1,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
constant int64_t & ne03,
constant uint64_t & nb00,
constant uint64_t & nb01,
constant uint64_t & nb02,
constant uint64_t & nb03,
constant int64_t & ne0,
constant int64_t & ne1,
constant int64_t & ne2,
constant int64_t & ne3,
constant uint64_t & nb0,
constant uint64_t & nb1,
constant uint64_t & nb2,
constant uint64_t & nb3,
constant int & n_past,
constant int & n_dims,
constant int & mode,
constant int & n_orig_ctx,
llama : custom attention mask + parallel decoding + no context swaps (#3228) * tests : verify that RoPE is "additive" * llama : replace ggml_diag_mask_inf with ggml_add (custom -inf mask) * ggml : ggml_rope now takes a vector with positions instead of n_past * metal : add rope_f16 kernel + optimize cpy kernels * llama : unified KV cache + batch inference API * llama : add new llama_decode() API that works with llama_batch * llama : add cell_max heuristic for more efficient kv_cache * llama : extend llama_kv_cache API * llama : more robust cell_max heuristic + wip shift * metal : disable concurrency optimization * llama : add llama_kv_cache_shift_seq + no more context swaps * llama : apply K-cache roping for Falcon and Baichuan * speculative : fix KV cache management * parallel : example for serving multiple users in parallel * parallel : disable hot-plug to avoid cache fragmentation * fixes : speculative KV cache + llama worst-case graph * llama : extend batch API to select which logits to output * llama : fix worst case graph build * ggml-cuda : update rope implementation for parallel decoding (#3254) * ggml-cuda : update rope implementation for parallel decoding * better solution for p0 computation * fix rope * simpler rope implementation --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * make : add parallel to build + fix static functions in llama.cpp * simple : fix token counting * parallel : various improvements * llama : fix cell_max logic + rename functions * parallel : try smaller batches when the KV cache is fragmented * parallel : fix sequence termination criteria * llama : silence errors KV cache errors * parallel : remove new line from prompt * parallel : process system prompt once + configurable paramters + llama API * parallel : remove question with short answers * parallel : count cache misses * parallel : print misses on each request * parallel : minor * llama : fix n_kv to never become 0 * parallel : rename hot-plug to continuous-batching * llama : improve llama_batch API + simplify parallel example * simple : add parallel decoding support * simple : improve comments + free batch * ggml-cuda : add rope f16, restore performance with parallel decoding (#3272) * ggml-cuda : add rope f16, restore performance * offload KQ_mask with all models * fix rope shift --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * llama : disable MPI for now ggml-ci * train : make KQ_pos memory buffer permanent via dummy scale op * ggml : revert change to ggml_cpy, add ggml_cont_Nd instead (#3275) ggml-ci * parallel : fix bug (extra BOS) + smaller token_prev array * parallel : fix cases where the input prompts can overflow the batch * parallel : add disabled experimental batch chunking in powers of two * llama : llama.h formatting + comments * simple : add README.md * llama : fix kv cache heuristic when context is less than 32 * parallel : fix crash when `-n -1` * llama : simplify returns if/else branches * metal : use mm kernels for batch size > 2 * examples : utilize new llama_get_logits_ith() * examples : add example for batched decoding * examples : do not eval prompt 2 times (close #3348) * server : clear the KV cache beyond n_past before llama_decode * server : avoid context swaps by shifting the KV cache --------- Co-authored-by: slaren <slarengh@gmail.com>
2023-09-28 18:04:36 +02:00
constant float & freq_base,
constant float & freq_scale,
constant float & ext_factor,
constant float & attn_factor,
constant float & beta_fast,
constant float & beta_slow,
llama : custom attention mask + parallel decoding + no context swaps (#3228) * tests : verify that RoPE is "additive" * llama : replace ggml_diag_mask_inf with ggml_add (custom -inf mask) * ggml : ggml_rope now takes a vector with positions instead of n_past * metal : add rope_f16 kernel + optimize cpy kernels * llama : unified KV cache + batch inference API * llama : add new llama_decode() API that works with llama_batch * llama : add cell_max heuristic for more efficient kv_cache * llama : extend llama_kv_cache API * llama : more robust cell_max heuristic + wip shift * metal : disable concurrency optimization * llama : add llama_kv_cache_shift_seq + no more context swaps * llama : apply K-cache roping for Falcon and Baichuan * speculative : fix KV cache management * parallel : example for serving multiple users in parallel * parallel : disable hot-plug to avoid cache fragmentation * fixes : speculative KV cache + llama worst-case graph * llama : extend batch API to select which logits to output * llama : fix worst case graph build * ggml-cuda : update rope implementation for parallel decoding (#3254) * ggml-cuda : update rope implementation for parallel decoding * better solution for p0 computation * fix rope * simpler rope implementation --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * make : add parallel to build + fix static functions in llama.cpp * simple : fix token counting * parallel : various improvements * llama : fix cell_max logic + rename functions * parallel : try smaller batches when the KV cache is fragmented * parallel : fix sequence termination criteria * llama : silence errors KV cache errors * parallel : remove new line from prompt * parallel : process system prompt once + configurable paramters + llama API * parallel : remove question with short answers * parallel : count cache misses * parallel : print misses on each request * parallel : minor * llama : fix n_kv to never become 0 * parallel : rename hot-plug to continuous-batching * llama : improve llama_batch API + simplify parallel example * simple : add parallel decoding support * simple : improve comments + free batch * ggml-cuda : add rope f16, restore performance with parallel decoding (#3272) * ggml-cuda : add rope f16, restore performance * offload KQ_mask with all models * fix rope shift --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * llama : disable MPI for now ggml-ci * train : make KQ_pos memory buffer permanent via dummy scale op * ggml : revert change to ggml_cpy, add ggml_cont_Nd instead (#3275) ggml-ci * parallel : fix bug (extra BOS) + smaller token_prev array * parallel : fix cases where the input prompts can overflow the batch * parallel : add disabled experimental batch chunking in powers of two * llama : llama.h formatting + comments * simple : add README.md * llama : fix kv cache heuristic when context is less than 32 * parallel : fix crash when `-n -1` * llama : simplify returns if/else branches * metal : use mm kernels for batch size > 2 * examples : utilize new llama_get_logits_ith() * examples : add example for batched decoding * examples : do not eval prompt 2 times (close #3348) * server : clear the KV cache beyond n_past before llama_decode * server : avoid context swaps by shifting the KV cache --------- Co-authored-by: slaren <slarengh@gmail.com>
2023-09-28 18:04:36 +02:00
uint tiitg[[thread_index_in_threadgroup]],
uint3 tptg[[threads_per_threadgroup]],
uint3 tgpig[[threadgroup_position_in_grid]]);
template<typename T>
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
kernel void kernel_rope(
llama : custom attention mask + parallel decoding + no context swaps (#3228) * tests : verify that RoPE is "additive" * llama : replace ggml_diag_mask_inf with ggml_add (custom -inf mask) * ggml : ggml_rope now takes a vector with positions instead of n_past * metal : add rope_f16 kernel + optimize cpy kernels * llama : unified KV cache + batch inference API * llama : add new llama_decode() API that works with llama_batch * llama : add cell_max heuristic for more efficient kv_cache * llama : extend llama_kv_cache API * llama : more robust cell_max heuristic + wip shift * metal : disable concurrency optimization * llama : add llama_kv_cache_shift_seq + no more context swaps * llama : apply K-cache roping for Falcon and Baichuan * speculative : fix KV cache management * parallel : example for serving multiple users in parallel * parallel : disable hot-plug to avoid cache fragmentation * fixes : speculative KV cache + llama worst-case graph * llama : extend batch API to select which logits to output * llama : fix worst case graph build * ggml-cuda : update rope implementation for parallel decoding (#3254) * ggml-cuda : update rope implementation for parallel decoding * better solution for p0 computation * fix rope * simpler rope implementation --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * make : add parallel to build + fix static functions in llama.cpp * simple : fix token counting * parallel : various improvements * llama : fix cell_max logic + rename functions * parallel : try smaller batches when the KV cache is fragmented * parallel : fix sequence termination criteria * llama : silence errors KV cache errors * parallel : remove new line from prompt * parallel : process system prompt once + configurable paramters + llama API * parallel : remove question with short answers * parallel : count cache misses * parallel : print misses on each request * parallel : minor * llama : fix n_kv to never become 0 * parallel : rename hot-plug to continuous-batching * llama : improve llama_batch API + simplify parallel example * simple : add parallel decoding support * simple : improve comments + free batch * ggml-cuda : add rope f16, restore performance with parallel decoding (#3272) * ggml-cuda : add rope f16, restore performance * offload KQ_mask with all models * fix rope shift --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * llama : disable MPI for now ggml-ci * train : make KQ_pos memory buffer permanent via dummy scale op * ggml : revert change to ggml_cpy, add ggml_cont_Nd instead (#3275) ggml-ci * parallel : fix bug (extra BOS) + smaller token_prev array * parallel : fix cases where the input prompts can overflow the batch * parallel : add disabled experimental batch chunking in powers of two * llama : llama.h formatting + comments * simple : add README.md * llama : fix kv cache heuristic when context is less than 32 * parallel : fix crash when `-n -1` * llama : simplify returns if/else branches * metal : use mm kernels for batch size > 2 * examples : utilize new llama_get_logits_ith() * examples : add example for batched decoding * examples : do not eval prompt 2 times (close #3348) * server : clear the KV cache beyond n_past before llama_decode * server : avoid context swaps by shifting the KV cache --------- Co-authored-by: slaren <slarengh@gmail.com>
2023-09-28 18:04:36 +02:00
device const void * src0,
device const int32_t * src1,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
constant int64_t & ne03,
constant uint64_t & nb00,
constant uint64_t & nb01,
constant uint64_t & nb02,
constant uint64_t & nb03,
constant int64_t & ne0,
constant int64_t & ne1,
constant int64_t & ne2,
constant int64_t & ne3,
constant uint64_t & nb0,
constant uint64_t & nb1,
constant uint64_t & nb2,
constant uint64_t & nb3,
constant int & n_past,
constant int & n_dims,
constant int & mode,
constant int & n_orig_ctx,
llama : custom attention mask + parallel decoding + no context swaps (#3228) * tests : verify that RoPE is "additive" * llama : replace ggml_diag_mask_inf with ggml_add (custom -inf mask) * ggml : ggml_rope now takes a vector with positions instead of n_past * metal : add rope_f16 kernel + optimize cpy kernels * llama : unified KV cache + batch inference API * llama : add new llama_decode() API that works with llama_batch * llama : add cell_max heuristic for more efficient kv_cache * llama : extend llama_kv_cache API * llama : more robust cell_max heuristic + wip shift * metal : disable concurrency optimization * llama : add llama_kv_cache_shift_seq + no more context swaps * llama : apply K-cache roping for Falcon and Baichuan * speculative : fix KV cache management * parallel : example for serving multiple users in parallel * parallel : disable hot-plug to avoid cache fragmentation * fixes : speculative KV cache + llama worst-case graph * llama : extend batch API to select which logits to output * llama : fix worst case graph build * ggml-cuda : update rope implementation for parallel decoding (#3254) * ggml-cuda : update rope implementation for parallel decoding * better solution for p0 computation * fix rope * simpler rope implementation --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * make : add parallel to build + fix static functions in llama.cpp * simple : fix token counting * parallel : various improvements * llama : fix cell_max logic + rename functions * parallel : try smaller batches when the KV cache is fragmented * parallel : fix sequence termination criteria * llama : silence errors KV cache errors * parallel : remove new line from prompt * parallel : process system prompt once + configurable paramters + llama API * parallel : remove question with short answers * parallel : count cache misses * parallel : print misses on each request * parallel : minor * llama : fix n_kv to never become 0 * parallel : rename hot-plug to continuous-batching * llama : improve llama_batch API + simplify parallel example * simple : add parallel decoding support * simple : improve comments + free batch * ggml-cuda : add rope f16, restore performance with parallel decoding (#3272) * ggml-cuda : add rope f16, restore performance * offload KQ_mask with all models * fix rope shift --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * llama : disable MPI for now ggml-ci * train : make KQ_pos memory buffer permanent via dummy scale op * ggml : revert change to ggml_cpy, add ggml_cont_Nd instead (#3275) ggml-ci * parallel : fix bug (extra BOS) + smaller token_prev array * parallel : fix cases where the input prompts can overflow the batch * parallel : add disabled experimental batch chunking in powers of two * llama : llama.h formatting + comments * simple : add README.md * llama : fix kv cache heuristic when context is less than 32 * parallel : fix crash when `-n -1` * llama : simplify returns if/else branches * metal : use mm kernels for batch size > 2 * examples : utilize new llama_get_logits_ith() * examples : add example for batched decoding * examples : do not eval prompt 2 times (close #3348) * server : clear the KV cache beyond n_past before llama_decode * server : avoid context swaps by shifting the KV cache --------- Co-authored-by: slaren <slarengh@gmail.com>
2023-09-28 18:04:36 +02:00
constant float & freq_base,
constant float & freq_scale,
constant float & ext_factor,
constant float & attn_factor,
constant float & beta_fast,
constant float & beta_slow,
uint tiitg[[thread_index_in_threadgroup]],
uint3 tptg[[threads_per_threadgroup]],
uint3 tgpig[[threadgroup_position_in_grid]]) {
const int64_t i3 = tgpig[2];
const int64_t i2 = tgpig[1];
const int64_t i1 = tgpig[0];
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
const bool is_neox = mode & 2;
float corr_dims[2];
rope_yarn_corr_dims(n_dims, n_orig_ctx, freq_base, beta_fast, beta_slow, corr_dims);
llama : custom attention mask + parallel decoding + no context swaps (#3228) * tests : verify that RoPE is "additive" * llama : replace ggml_diag_mask_inf with ggml_add (custom -inf mask) * ggml : ggml_rope now takes a vector with positions instead of n_past * metal : add rope_f16 kernel + optimize cpy kernels * llama : unified KV cache + batch inference API * llama : add new llama_decode() API that works with llama_batch * llama : add cell_max heuristic for more efficient kv_cache * llama : extend llama_kv_cache API * llama : more robust cell_max heuristic + wip shift * metal : disable concurrency optimization * llama : add llama_kv_cache_shift_seq + no more context swaps * llama : apply K-cache roping for Falcon and Baichuan * speculative : fix KV cache management * parallel : example for serving multiple users in parallel * parallel : disable hot-plug to avoid cache fragmentation * fixes : speculative KV cache + llama worst-case graph * llama : extend batch API to select which logits to output * llama : fix worst case graph build * ggml-cuda : update rope implementation for parallel decoding (#3254) * ggml-cuda : update rope implementation for parallel decoding * better solution for p0 computation * fix rope * simpler rope implementation --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * make : add parallel to build + fix static functions in llama.cpp * simple : fix token counting * parallel : various improvements * llama : fix cell_max logic + rename functions * parallel : try smaller batches when the KV cache is fragmented * parallel : fix sequence termination criteria * llama : silence errors KV cache errors * parallel : remove new line from prompt * parallel : process system prompt once + configurable paramters + llama API * parallel : remove question with short answers * parallel : count cache misses * parallel : print misses on each request * parallel : minor * llama : fix n_kv to never become 0 * parallel : rename hot-plug to continuous-batching * llama : improve llama_batch API + simplify parallel example * simple : add parallel decoding support * simple : improve comments + free batch * ggml-cuda : add rope f16, restore performance with parallel decoding (#3272) * ggml-cuda : add rope f16, restore performance * offload KQ_mask with all models * fix rope shift --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * llama : disable MPI for now ggml-ci * train : make KQ_pos memory buffer permanent via dummy scale op * ggml : revert change to ggml_cpy, add ggml_cont_Nd instead (#3275) ggml-ci * parallel : fix bug (extra BOS) + smaller token_prev array * parallel : fix cases where the input prompts can overflow the batch * parallel : add disabled experimental batch chunking in powers of two * llama : llama.h formatting + comments * simple : add README.md * llama : fix kv cache heuristic when context is less than 32 * parallel : fix crash when `-n -1` * llama : simplify returns if/else branches * metal : use mm kernels for batch size > 2 * examples : utilize new llama_get_logits_ith() * examples : add example for batched decoding * examples : do not eval prompt 2 times (close #3348) * server : clear the KV cache beyond n_past before llama_decode * server : avoid context swaps by shifting the KV cache --------- Co-authored-by: slaren <slarengh@gmail.com>
2023-09-28 18:04:36 +02:00
device const int32_t * pos = src1;
const int64_t p = pos[i2];
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
const float theta_0 = (float)p;
const float inv_ndims = -1.f/n_dims;
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
if (!is_neox) {
for (int64_t i0 = 2*tiitg; i0 < ne0; i0 += 2*tptg.x) {
const float theta = theta_0 * pow(freq_base, inv_ndims*i0);
float cos_theta, sin_theta;
rope_yarn(theta, freq_scale, corr_dims, i0, ext_factor, attn_factor, &cos_theta, &sin_theta);
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
llama : custom attention mask + parallel decoding + no context swaps (#3228) * tests : verify that RoPE is "additive" * llama : replace ggml_diag_mask_inf with ggml_add (custom -inf mask) * ggml : ggml_rope now takes a vector with positions instead of n_past * metal : add rope_f16 kernel + optimize cpy kernels * llama : unified KV cache + batch inference API * llama : add new llama_decode() API that works with llama_batch * llama : add cell_max heuristic for more efficient kv_cache * llama : extend llama_kv_cache API * llama : more robust cell_max heuristic + wip shift * metal : disable concurrency optimization * llama : add llama_kv_cache_shift_seq + no more context swaps * llama : apply K-cache roping for Falcon and Baichuan * speculative : fix KV cache management * parallel : example for serving multiple users in parallel * parallel : disable hot-plug to avoid cache fragmentation * fixes : speculative KV cache + llama worst-case graph * llama : extend batch API to select which logits to output * llama : fix worst case graph build * ggml-cuda : update rope implementation for parallel decoding (#3254) * ggml-cuda : update rope implementation for parallel decoding * better solution for p0 computation * fix rope * simpler rope implementation --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * make : add parallel to build + fix static functions in llama.cpp * simple : fix token counting * parallel : various improvements * llama : fix cell_max logic + rename functions * parallel : try smaller batches when the KV cache is fragmented * parallel : fix sequence termination criteria * llama : silence errors KV cache errors * parallel : remove new line from prompt * parallel : process system prompt once + configurable paramters + llama API * parallel : remove question with short answers * parallel : count cache misses * parallel : print misses on each request * parallel : minor * llama : fix n_kv to never become 0 * parallel : rename hot-plug to continuous-batching * llama : improve llama_batch API + simplify parallel example * simple : add parallel decoding support * simple : improve comments + free batch * ggml-cuda : add rope f16, restore performance with parallel decoding (#3272) * ggml-cuda : add rope f16, restore performance * offload KQ_mask with all models * fix rope shift --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * llama : disable MPI for now ggml-ci * train : make KQ_pos memory buffer permanent via dummy scale op * ggml : revert change to ggml_cpy, add ggml_cont_Nd instead (#3275) ggml-ci * parallel : fix bug (extra BOS) + smaller token_prev array * parallel : fix cases where the input prompts can overflow the batch * parallel : add disabled experimental batch chunking in powers of two * llama : llama.h formatting + comments * simple : add README.md * llama : fix kv cache heuristic when context is less than 32 * parallel : fix crash when `-n -1` * llama : simplify returns if/else branches * metal : use mm kernels for batch size > 2 * examples : utilize new llama_get_logits_ith() * examples : add example for batched decoding * examples : do not eval prompt 2 times (close #3348) * server : clear the KV cache beyond n_past before llama_decode * server : avoid context swaps by shifting the KV cache --------- Co-authored-by: slaren <slarengh@gmail.com>
2023-09-28 18:04:36 +02:00
device const T * const src = (device T *)((device char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
device T * dst_data = (device T *)((device char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
llama : custom attention mask + parallel decoding + no context swaps (#3228) * tests : verify that RoPE is "additive" * llama : replace ggml_diag_mask_inf with ggml_add (custom -inf mask) * ggml : ggml_rope now takes a vector with positions instead of n_past * metal : add rope_f16 kernel + optimize cpy kernels * llama : unified KV cache + batch inference API * llama : add new llama_decode() API that works with llama_batch * llama : add cell_max heuristic for more efficient kv_cache * llama : extend llama_kv_cache API * llama : more robust cell_max heuristic + wip shift * metal : disable concurrency optimization * llama : add llama_kv_cache_shift_seq + no more context swaps * llama : apply K-cache roping for Falcon and Baichuan * speculative : fix KV cache management * parallel : example for serving multiple users in parallel * parallel : disable hot-plug to avoid cache fragmentation * fixes : speculative KV cache + llama worst-case graph * llama : extend batch API to select which logits to output * llama : fix worst case graph build * ggml-cuda : update rope implementation for parallel decoding (#3254) * ggml-cuda : update rope implementation for parallel decoding * better solution for p0 computation * fix rope * simpler rope implementation --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * make : add parallel to build + fix static functions in llama.cpp * simple : fix token counting * parallel : various improvements * llama : fix cell_max logic + rename functions * parallel : try smaller batches when the KV cache is fragmented * parallel : fix sequence termination criteria * llama : silence errors KV cache errors * parallel : remove new line from prompt * parallel : process system prompt once + configurable paramters + llama API * parallel : remove question with short answers * parallel : count cache misses * parallel : print misses on each request * parallel : minor * llama : fix n_kv to never become 0 * parallel : rename hot-plug to continuous-batching * llama : improve llama_batch API + simplify parallel example * simple : add parallel decoding support * simple : improve comments + free batch * ggml-cuda : add rope f16, restore performance with parallel decoding (#3272) * ggml-cuda : add rope f16, restore performance * offload KQ_mask with all models * fix rope shift --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * llama : disable MPI for now ggml-ci * train : make KQ_pos memory buffer permanent via dummy scale op * ggml : revert change to ggml_cpy, add ggml_cont_Nd instead (#3275) ggml-ci * parallel : fix bug (extra BOS) + smaller token_prev array * parallel : fix cases where the input prompts can overflow the batch * parallel : add disabled experimental batch chunking in powers of two * llama : llama.h formatting + comments * simple : add README.md * llama : fix kv cache heuristic when context is less than 32 * parallel : fix crash when `-n -1` * llama : simplify returns if/else branches * metal : use mm kernels for batch size > 2 * examples : utilize new llama_get_logits_ith() * examples : add example for batched decoding * examples : do not eval prompt 2 times (close #3348) * server : clear the KV cache beyond n_past before llama_decode * server : avoid context swaps by shifting the KV cache --------- Co-authored-by: slaren <slarengh@gmail.com>
2023-09-28 18:04:36 +02:00
const T x0 = src[0];
const T x1 = src[1];
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
dst_data[0] = x0*cos_theta - x1*sin_theta;
dst_data[1] = x0*sin_theta + x1*cos_theta;
}
} else {
for (int64_t ic = 2*tiitg; ic < ne0; ic += 2*tptg.x) {
if (ic < n_dims) {
const int64_t ib = 0;
// simplified from `(ib * n_dims + ic) * inv_ndims`
const float cur_rot = inv_ndims*ic - ib;
const float theta = theta_0 * pow(freq_base, cur_rot);
float cos_theta, sin_theta;
rope_yarn(theta, freq_scale, corr_dims, cur_rot, ext_factor, attn_factor, &cos_theta, &sin_theta);
llm : add Falcon support (#2717) * llama : refactor GGUF constants into static maps * llama : check if model architecture is known * llama : refactor llama_model_load_internal() * gguf : add KV constant maps * llm : read arch-specific KVs * convert : add dummy scores + types * falcon : load tensor data (CPU only) * llama : fix loading progress bar * llama : add arch member to llama_model * falcon : CPU inference working * falcon : support non-40B models * falcon : minor * llama : minor updates ggml-ci * convert-falcon-hf-to-gguf.py : fix special token mapping * llama.cpp : llama default UNK token = id 0 * llama.cpp : fix bpe tokenizer * llama.cpp : fix the fix of bpe tokenizer * ggml : pass eps to ggml_norm * metal : implement RoPE (mode = 2) + avoid ggml_repeat * ggml : ggml_repeat always creates new tensor * falcon : copy-paste self-attention from LLaMA * metal : print extra compute pipeline info * falcon : minor changes (still chasing the Metal problem) * llama.cpp : fix linefeed token * metal : fix GELU kernel numerical stability by using precise::tanh * metal : temporary workaround for the concurrency optimization bug * falcon : add CUDA offloading (#2739) * llama : better model naming and size reporting * llama : prep new tokenizer support * llama : advanced BPE tokenizer based on ggllm.cpp imlpementation * llama : remove oboslete comment ggml-ci * common : remove obsolete BPE API + disable test-tokenizer-1 * llama : revert BPE special-case in llama_byte_to_token() * cuda : add TODOs for RoPE NeoX implementation * llama : default special tokens based on vocab type * perplexity : add log for start of tokenization --------- Co-authored-by: klosax <131523366+klosax@users.noreply.github.com> Co-authored-by: slaren <slarengh@gmail.com>
2023-08-23 22:08:04 +02:00
const int64_t i0 = ib*n_dims + ic/2;
llama : custom attention mask + parallel decoding + no context swaps (#3228) * tests : verify that RoPE is "additive" * llama : replace ggml_diag_mask_inf with ggml_add (custom -inf mask) * ggml : ggml_rope now takes a vector with positions instead of n_past * metal : add rope_f16 kernel + optimize cpy kernels * llama : unified KV cache + batch inference API * llama : add new llama_decode() API that works with llama_batch * llama : add cell_max heuristic for more efficient kv_cache * llama : extend llama_kv_cache API * llama : more robust cell_max heuristic + wip shift * metal : disable concurrency optimization * llama : add llama_kv_cache_shift_seq + no more context swaps * llama : apply K-cache roping for Falcon and Baichuan * speculative : fix KV cache management * parallel : example for serving multiple users in parallel * parallel : disable hot-plug to avoid cache fragmentation * fixes : speculative KV cache + llama worst-case graph * llama : extend batch API to select which logits to output * llama : fix worst case graph build * ggml-cuda : update rope implementation for parallel decoding (#3254) * ggml-cuda : update rope implementation for parallel decoding * better solution for p0 computation * fix rope * simpler rope implementation --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * make : add parallel to build + fix static functions in llama.cpp * simple : fix token counting * parallel : various improvements * llama : fix cell_max logic + rename functions * parallel : try smaller batches when the KV cache is fragmented * parallel : fix sequence termination criteria * llama : silence errors KV cache errors * parallel : remove new line from prompt * parallel : process system prompt once + configurable paramters + llama API * parallel : remove question with short answers * parallel : count cache misses * parallel : print misses on each request * parallel : minor * llama : fix n_kv to never become 0 * parallel : rename hot-plug to continuous-batching * llama : improve llama_batch API + simplify parallel example * simple : add parallel decoding support * simple : improve comments + free batch * ggml-cuda : add rope f16, restore performance with parallel decoding (#3272) * ggml-cuda : add rope f16, restore performance * offload KQ_mask with all models * fix rope shift --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * llama : disable MPI for now ggml-ci * train : make KQ_pos memory buffer permanent via dummy scale op * ggml : revert change to ggml_cpy, add ggml_cont_Nd instead (#3275) ggml-ci * parallel : fix bug (extra BOS) + smaller token_prev array * parallel : fix cases where the input prompts can overflow the batch * parallel : add disabled experimental batch chunking in powers of two * llama : llama.h formatting + comments * simple : add README.md * llama : fix kv cache heuristic when context is less than 32 * parallel : fix crash when `-n -1` * llama : simplify returns if/else branches * metal : use mm kernels for batch size > 2 * examples : utilize new llama_get_logits_ith() * examples : add example for batched decoding * examples : do not eval prompt 2 times (close #3348) * server : clear the KV cache beyond n_past before llama_decode * server : avoid context swaps by shifting the KV cache --------- Co-authored-by: slaren <slarengh@gmail.com>
2023-09-28 18:04:36 +02:00
device const T * const src = (device T *)((device char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
device T * dst_data = (device T *)((device char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
llm : add Falcon support (#2717) * llama : refactor GGUF constants into static maps * llama : check if model architecture is known * llama : refactor llama_model_load_internal() * gguf : add KV constant maps * llm : read arch-specific KVs * convert : add dummy scores + types * falcon : load tensor data (CPU only) * llama : fix loading progress bar * llama : add arch member to llama_model * falcon : CPU inference working * falcon : support non-40B models * falcon : minor * llama : minor updates ggml-ci * convert-falcon-hf-to-gguf.py : fix special token mapping * llama.cpp : llama default UNK token = id 0 * llama.cpp : fix bpe tokenizer * llama.cpp : fix the fix of bpe tokenizer * ggml : pass eps to ggml_norm * metal : implement RoPE (mode = 2) + avoid ggml_repeat * ggml : ggml_repeat always creates new tensor * falcon : copy-paste self-attention from LLaMA * metal : print extra compute pipeline info * falcon : minor changes (still chasing the Metal problem) * llama.cpp : fix linefeed token * metal : fix GELU kernel numerical stability by using precise::tanh * metal : temporary workaround for the concurrency optimization bug * falcon : add CUDA offloading (#2739) * llama : better model naming and size reporting * llama : prep new tokenizer support * llama : advanced BPE tokenizer based on ggllm.cpp imlpementation * llama : remove oboslete comment ggml-ci * common : remove obsolete BPE API + disable test-tokenizer-1 * llama : revert BPE special-case in llama_byte_to_token() * cuda : add TODOs for RoPE NeoX implementation * llama : default special tokens based on vocab type * perplexity : add log for start of tokenization --------- Co-authored-by: klosax <131523366+klosax@users.noreply.github.com> Co-authored-by: slaren <slarengh@gmail.com>
2023-08-23 22:08:04 +02:00
const float x0 = src[0];
const float x1 = src[n_dims/2];
dst_data[0] = x0*cos_theta - x1*sin_theta;
dst_data[n_dims/2] = x0*sin_theta + x1*cos_theta;
} else {
const int64_t i0 = ic;
device const T * const src = (device T *)((device char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
device T * dst_data = (device T *)((device char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
dst_data[0] = src[0];
dst_data[1] = src[1];
llm : add Falcon support (#2717) * llama : refactor GGUF constants into static maps * llama : check if model architecture is known * llama : refactor llama_model_load_internal() * gguf : add KV constant maps * llm : read arch-specific KVs * convert : add dummy scores + types * falcon : load tensor data (CPU only) * llama : fix loading progress bar * llama : add arch member to llama_model * falcon : CPU inference working * falcon : support non-40B models * falcon : minor * llama : minor updates ggml-ci * convert-falcon-hf-to-gguf.py : fix special token mapping * llama.cpp : llama default UNK token = id 0 * llama.cpp : fix bpe tokenizer * llama.cpp : fix the fix of bpe tokenizer * ggml : pass eps to ggml_norm * metal : implement RoPE (mode = 2) + avoid ggml_repeat * ggml : ggml_repeat always creates new tensor * falcon : copy-paste self-attention from LLaMA * metal : print extra compute pipeline info * falcon : minor changes (still chasing the Metal problem) * llama.cpp : fix linefeed token * metal : fix GELU kernel numerical stability by using precise::tanh * metal : temporary workaround for the concurrency optimization bug * falcon : add CUDA offloading (#2739) * llama : better model naming and size reporting * llama : prep new tokenizer support * llama : advanced BPE tokenizer based on ggllm.cpp imlpementation * llama : remove oboslete comment ggml-ci * common : remove obsolete BPE API + disable test-tokenizer-1 * llama : revert BPE special-case in llama_byte_to_token() * cuda : add TODOs for RoPE NeoX implementation * llama : default special tokens based on vocab type * perplexity : add log for start of tokenization --------- Co-authored-by: klosax <131523366+klosax@users.noreply.github.com> Co-authored-by: slaren <slarengh@gmail.com>
2023-08-23 22:08:04 +02:00
}
}
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
}
}
llama : custom attention mask + parallel decoding + no context swaps (#3228) * tests : verify that RoPE is "additive" * llama : replace ggml_diag_mask_inf with ggml_add (custom -inf mask) * ggml : ggml_rope now takes a vector with positions instead of n_past * metal : add rope_f16 kernel + optimize cpy kernels * llama : unified KV cache + batch inference API * llama : add new llama_decode() API that works with llama_batch * llama : add cell_max heuristic for more efficient kv_cache * llama : extend llama_kv_cache API * llama : more robust cell_max heuristic + wip shift * metal : disable concurrency optimization * llama : add llama_kv_cache_shift_seq + no more context swaps * llama : apply K-cache roping for Falcon and Baichuan * speculative : fix KV cache management * parallel : example for serving multiple users in parallel * parallel : disable hot-plug to avoid cache fragmentation * fixes : speculative KV cache + llama worst-case graph * llama : extend batch API to select which logits to output * llama : fix worst case graph build * ggml-cuda : update rope implementation for parallel decoding (#3254) * ggml-cuda : update rope implementation for parallel decoding * better solution for p0 computation * fix rope * simpler rope implementation --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * make : add parallel to build + fix static functions in llama.cpp * simple : fix token counting * parallel : various improvements * llama : fix cell_max logic + rename functions * parallel : try smaller batches when the KV cache is fragmented * parallel : fix sequence termination criteria * llama : silence errors KV cache errors * parallel : remove new line from prompt * parallel : process system prompt once + configurable paramters + llama API * parallel : remove question with short answers * parallel : count cache misses * parallel : print misses on each request * parallel : minor * llama : fix n_kv to never become 0 * parallel : rename hot-plug to continuous-batching * llama : improve llama_batch API + simplify parallel example * simple : add parallel decoding support * simple : improve comments + free batch * ggml-cuda : add rope f16, restore performance with parallel decoding (#3272) * ggml-cuda : add rope f16, restore performance * offload KQ_mask with all models * fix rope shift --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * llama : disable MPI for now ggml-ci * train : make KQ_pos memory buffer permanent via dummy scale op * ggml : revert change to ggml_cpy, add ggml_cont_Nd instead (#3275) ggml-ci * parallel : fix bug (extra BOS) + smaller token_prev array * parallel : fix cases where the input prompts can overflow the batch * parallel : add disabled experimental batch chunking in powers of two * llama : llama.h formatting + comments * simple : add README.md * llama : fix kv cache heuristic when context is less than 32 * parallel : fix crash when `-n -1` * llama : simplify returns if/else branches * metal : use mm kernels for batch size > 2 * examples : utilize new llama_get_logits_ith() * examples : add example for batched decoding * examples : do not eval prompt 2 times (close #3348) * server : clear the KV cache beyond n_past before llama_decode * server : avoid context swaps by shifting the KV cache --------- Co-authored-by: slaren <slarengh@gmail.com>
2023-09-28 18:04:36 +02:00
template [[host_name("kernel_rope_f32")]] kernel rope_t kernel_rope<float>;
template [[host_name("kernel_rope_f16")]] kernel rope_t kernel_rope<half>;
typedef void (im2col_t)(
device const float * x,
device char * dst,
constant int32_t & ofs0,
constant int32_t & ofs1,
constant int32_t & IW,
constant int32_t & IH,
constant int32_t & CHW,
constant int32_t & s0,
constant int32_t & s1,
constant int32_t & p0,
constant int32_t & p1,
constant int32_t & d0,
constant int32_t & d1,
uint3 tgpig[[threadgroup_position_in_grid]],
uint3 tgpg[[threadgroups_per_grid]],
uint3 tpitg[[thread_position_in_threadgroup]],
uint3 ntg[[threads_per_threadgroup]]);
template <typename T>
kernel void kernel_im2col(
device const float * x,
device char * dst,
constant int32_t & ofs0,
constant int32_t & ofs1,
constant int32_t & IW,
constant int32_t & IH,
constant int32_t & CHW,
constant int32_t & s0,
constant int32_t & s1,
constant int32_t & p0,
constant int32_t & p1,
constant int32_t & d0,
constant int32_t & d1,
uint3 tgpig[[threadgroup_position_in_grid]],
uint3 tgpg[[threadgroups_per_grid]],
uint3 tpitg[[thread_position_in_threadgroup]],
uint3 ntg[[threads_per_threadgroup]]) {
const int32_t iiw = tgpig[2] * s0 + tpitg[2] * d0 - p0;
const int32_t iih = tgpig[1] * s1 + tpitg[1] * d1 - p1;
const int32_t offset_dst =
(tpitg[0] * tgpg[1] * tgpg[2] + tgpig[1] * tgpg[2] + tgpig[2]) * CHW +
(tgpig[0] * (ntg[1] * ntg[2]) + tpitg[1] * ntg[2] + tpitg[2]);
device T * pdst = (device T *) (dst);
if (iih < 0 || iih >= IH || iiw < 0 || iiw >= IW) {
pdst[offset_dst] = 0.0f;
} else {
const int32_t offset_src = tpitg[0] * ofs0 + tgpig[0] * ofs1;
pdst[offset_dst] = x[offset_src + iih * IW + iiw];
}
}
template [[host_name("kernel_im2col_f32")]] kernel im2col_t kernel_im2col<float>;
template [[host_name("kernel_im2col_f16")]] kernel im2col_t kernel_im2col<half>;
kernel void kernel_upscale_f32(
device const char * src0,
device char * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
constant int64_t & ne03,
constant uint64_t & nb00,
constant uint64_t & nb01,
constant uint64_t & nb02,
constant uint64_t & nb03,
constant int64_t & ne0,
constant int64_t & ne1,
constant int64_t & ne2,
constant int64_t & ne3,
constant uint64_t & nb0,
constant uint64_t & nb1,
constant uint64_t & nb2,
constant uint64_t & nb3,
constant int32_t & sf,
uint3 tgpig[[threadgroup_position_in_grid]],
uint3 tpitg[[thread_position_in_threadgroup]],
uint3 ntg[[threads_per_threadgroup]]) {
const int64_t i3 = tgpig.z;
const int64_t i2 = tgpig.y;
const int64_t i1 = tgpig.x;
const int64_t i03 = i3;
const int64_t i02 = i2;
const int64_t i01 = i1/sf;
device const float * src0_ptr = (device const float *) (src0 + i03*nb03 + i02*nb02 + i01*nb01);
device float * dst_ptr = (device float *) (dst + i3*nb3 + i2*nb2 + i1*nb1);
for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) {
dst_ptr[i0] = src0_ptr[i0/sf];
}
}
kernel void kernel_pad_f32(
device const char * src0,
device char * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
constant int64_t & ne03,
constant uint64_t & nb00,
constant uint64_t & nb01,
constant uint64_t & nb02,
constant uint64_t & nb03,
constant int64_t & ne0,
constant int64_t & ne1,
constant int64_t & ne2,
constant int64_t & ne3,
constant uint64_t & nb0,
constant uint64_t & nb1,
constant uint64_t & nb2,
constant uint64_t & nb3,
uint3 tgpig[[threadgroup_position_in_grid]],
uint3 tpitg[[thread_position_in_threadgroup]],
uint3 ntg[[threads_per_threadgroup]]) {
const int64_t i3 = tgpig.z;
const int64_t i2 = tgpig.y;
const int64_t i1 = tgpig.x;
const int64_t i03 = i3;
const int64_t i02 = i2;
const int64_t i01 = i1;
device const float * src0_ptr = (device const float *) (src0 + i03*nb03 + i02*nb02 + i01*nb01);
device float * dst_ptr = (device float *) (dst + i3*nb3 + i2*nb2 + i1*nb1);
if (i1 < ne01 && i2 < ne02 && i3 < ne03) {
for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) {
if (i0 < ne00) {
dst_ptr[i0] = src0_ptr[i0];
} else {
dst_ptr[i0] = 0.0f;
}
}
return;
}
for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) {
dst_ptr[i0] = 0.0f;
}
}
kernel void kernel_arange_f32(
device char * dst,
constant int64_t & ne0,
constant float & start,
constant float & step,
uint3 tgpig[[threadgroup_position_in_grid]],
uint3 tpitg[[thread_position_in_threadgroup]],
uint3 ntg[[threads_per_threadgroup]]) {
device float * dst_ptr = (device float *) dst;
for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) {
dst_ptr[i0] = start + step * i0;
}
}
kernel void kernel_timestep_embedding_f32(
device const char * src0,
device char * dst,
constant uint64_t & nb1,
constant int & dim,
constant int & max_period,
uint3 tgpig[[threadgroup_position_in_grid]],
uint3 tpitg[[thread_position_in_threadgroup]],
uint3 ntg[[threads_per_threadgroup]]) {
int i = tgpig.x;
device float * embed_data = (device float *)(dst + i*nb1);
int half_ = dim / 2;
for (int j = tpitg.x; j < half_; j += ntg.x) {
float timestep = ((device float *)src0)[i];
float freq = (float)exp(-log((float)max_period) * j / half_);
float arg = timestep * freq;
embed_data[j ] = cos(arg);
embed_data[j + half_] = sin(arg);
}
if (dim % 2 != 0 && tpitg.x == 0) {
embed_data[dim] = 0.f;
}
}
// bitonic sort implementation following the CUDA kernels as reference
typedef void (argsort_t)(
ggml : mul_mat_id use the same tensor for all the experts (#6387) * ggml : update mul_mat_id to use the same tensor for all the experts * update cuda * minor * update metal * update test-backend-ops * fix cuda * Update ggml-metal.m Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * update convert.py * update convert-hf-to-gguf.py * update convert.py for mixtral hf models * Update convert-hf-to-gguf.py Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * cuda : support non-pow-2 number of experts * allow quantize to work for split and merged experts models in the same way * cleanup + disable mmap automatically with split tensors models * update imatrix * test-backend-ops : test qwen argsort * update grok model loading * llama : add merged experts tensors to the grok tensor map * minor * gguf : bump version * fix quantizing of merged experts * convert-hf-to-gguf.py : update grok (untested) * make linter happy * cuda/argsort : use shared memory instead of pool memory * convert : fix grok tensor names * metal : add support for non-pow-2 argsort * llama : more loader cleanup, better error checking * cuda : fix warning * llama : still use mmap for loading old models, but copy the data to a host buffer * add review note * llama : remove ffn tensor counting + add sanity check ggml-ci * convert : fix handling of n_experts == None ggml-ci * imatrix : fix ncall counters * llama : produce error if imatrix size does not match * quantize : terminate on errors + trace logs ggml-ci * metal : pad shared memory to 16 bytes --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-04-03 15:07:05 +02:00
device const float * x,
device int32_t * dst,
constant int64_t & ncols,
constant int64_t & ncols_pad,
threadgroup int32_t * shared_values [[threadgroup(0)]],
uint3 tgpig[[threadgroup_position_in_grid]],
uint3 tpitg[[thread_position_in_threadgroup]]);
template<ggml_sort_order order>
kernel void kernel_argsort_f32_i32(
device const float * x,
device int32_t * dst,
constant int64_t & ncols,
ggml : mul_mat_id use the same tensor for all the experts (#6387) * ggml : update mul_mat_id to use the same tensor for all the experts * update cuda * minor * update metal * update test-backend-ops * fix cuda * Update ggml-metal.m Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * update convert.py * update convert-hf-to-gguf.py * update convert.py for mixtral hf models * Update convert-hf-to-gguf.py Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * cuda : support non-pow-2 number of experts * allow quantize to work for split and merged experts models in the same way * cleanup + disable mmap automatically with split tensors models * update imatrix * test-backend-ops : test qwen argsort * update grok model loading * llama : add merged experts tensors to the grok tensor map * minor * gguf : bump version * fix quantizing of merged experts * convert-hf-to-gguf.py : update grok (untested) * make linter happy * cuda/argsort : use shared memory instead of pool memory * convert : fix grok tensor names * metal : add support for non-pow-2 argsort * llama : more loader cleanup, better error checking * cuda : fix warning * llama : still use mmap for loading old models, but copy the data to a host buffer * add review note * llama : remove ffn tensor counting + add sanity check ggml-ci * convert : fix handling of n_experts == None ggml-ci * imatrix : fix ncall counters * llama : produce error if imatrix size does not match * quantize : terminate on errors + trace logs ggml-ci * metal : pad shared memory to 16 bytes --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-04-03 15:07:05 +02:00
constant int64_t & ncols_pad,
threadgroup int32_t * shared_values [[threadgroup(0)]],
uint3 tgpig[[threadgroup_position_in_grid]],
uint3 tpitg[[thread_position_in_threadgroup]]) {
// bitonic sort
int col = tpitg[0];
int row = tgpig[1];
ggml : mul_mat_id use the same tensor for all the experts (#6387) * ggml : update mul_mat_id to use the same tensor for all the experts * update cuda * minor * update metal * update test-backend-ops * fix cuda * Update ggml-metal.m Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * update convert.py * update convert-hf-to-gguf.py * update convert.py for mixtral hf models * Update convert-hf-to-gguf.py Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * cuda : support non-pow-2 number of experts * allow quantize to work for split and merged experts models in the same way * cleanup + disable mmap automatically with split tensors models * update imatrix * test-backend-ops : test qwen argsort * update grok model loading * llama : add merged experts tensors to the grok tensor map * minor * gguf : bump version * fix quantizing of merged experts * convert-hf-to-gguf.py : update grok (untested) * make linter happy * cuda/argsort : use shared memory instead of pool memory * convert : fix grok tensor names * metal : add support for non-pow-2 argsort * llama : more loader cleanup, better error checking * cuda : fix warning * llama : still use mmap for loading old models, but copy the data to a host buffer * add review note * llama : remove ffn tensor counting + add sanity check ggml-ci * convert : fix handling of n_experts == None ggml-ci * imatrix : fix ncall counters * llama : produce error if imatrix size does not match * quantize : terminate on errors + trace logs ggml-ci * metal : pad shared memory to 16 bytes --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-04-03 15:07:05 +02:00
if (col >= ncols_pad) return;
ggml : mul_mat_id use the same tensor for all the experts (#6387) * ggml : update mul_mat_id to use the same tensor for all the experts * update cuda * minor * update metal * update test-backend-ops * fix cuda * Update ggml-metal.m Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * update convert.py * update convert-hf-to-gguf.py * update convert.py for mixtral hf models * Update convert-hf-to-gguf.py Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * cuda : support non-pow-2 number of experts * allow quantize to work for split and merged experts models in the same way * cleanup + disable mmap automatically with split tensors models * update imatrix * test-backend-ops : test qwen argsort * update grok model loading * llama : add merged experts tensors to the grok tensor map * minor * gguf : bump version * fix quantizing of merged experts * convert-hf-to-gguf.py : update grok (untested) * make linter happy * cuda/argsort : use shared memory instead of pool memory * convert : fix grok tensor names * metal : add support for non-pow-2 argsort * llama : more loader cleanup, better error checking * cuda : fix warning * llama : still use mmap for loading old models, but copy the data to a host buffer * add review note * llama : remove ffn tensor counting + add sanity check ggml-ci * convert : fix handling of n_experts == None ggml-ci * imatrix : fix ncall counters * llama : produce error if imatrix size does not match * quantize : terminate on errors + trace logs ggml-ci * metal : pad shared memory to 16 bytes --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-04-03 15:07:05 +02:00
device const float * x_row = x + row * ncols;
threadgroup int32_t * dst_row = shared_values;
// initialize indices
ggml : mul_mat_id use the same tensor for all the experts (#6387) * ggml : update mul_mat_id to use the same tensor for all the experts * update cuda * minor * update metal * update test-backend-ops * fix cuda * Update ggml-metal.m Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * update convert.py * update convert-hf-to-gguf.py * update convert.py for mixtral hf models * Update convert-hf-to-gguf.py Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * cuda : support non-pow-2 number of experts * allow quantize to work for split and merged experts models in the same way * cleanup + disable mmap automatically with split tensors models * update imatrix * test-backend-ops : test qwen argsort * update grok model loading * llama : add merged experts tensors to the grok tensor map * minor * gguf : bump version * fix quantizing of merged experts * convert-hf-to-gguf.py : update grok (untested) * make linter happy * cuda/argsort : use shared memory instead of pool memory * convert : fix grok tensor names * metal : add support for non-pow-2 argsort * llama : more loader cleanup, better error checking * cuda : fix warning * llama : still use mmap for loading old models, but copy the data to a host buffer * add review note * llama : remove ffn tensor counting + add sanity check ggml-ci * convert : fix handling of n_experts == None ggml-ci * imatrix : fix ncall counters * llama : produce error if imatrix size does not match * quantize : terminate on errors + trace logs ggml-ci * metal : pad shared memory to 16 bytes --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-04-03 15:07:05 +02:00
dst_row[col] = col;
threadgroup_barrier(mem_flags::mem_threadgroup);
ggml : mul_mat_id use the same tensor for all the experts (#6387) * ggml : update mul_mat_id to use the same tensor for all the experts * update cuda * minor * update metal * update test-backend-ops * fix cuda * Update ggml-metal.m Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * update convert.py * update convert-hf-to-gguf.py * update convert.py for mixtral hf models * Update convert-hf-to-gguf.py Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * cuda : support non-pow-2 number of experts * allow quantize to work for split and merged experts models in the same way * cleanup + disable mmap automatically with split tensors models * update imatrix * test-backend-ops : test qwen argsort * update grok model loading * llama : add merged experts tensors to the grok tensor map * minor * gguf : bump version * fix quantizing of merged experts * convert-hf-to-gguf.py : update grok (untested) * make linter happy * cuda/argsort : use shared memory instead of pool memory * convert : fix grok tensor names * metal : add support for non-pow-2 argsort * llama : more loader cleanup, better error checking * cuda : fix warning * llama : still use mmap for loading old models, but copy the data to a host buffer * add review note * llama : remove ffn tensor counting + add sanity check ggml-ci * convert : fix handling of n_experts == None ggml-ci * imatrix : fix ncall counters * llama : produce error if imatrix size does not match * quantize : terminate on errors + trace logs ggml-ci * metal : pad shared memory to 16 bytes --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-04-03 15:07:05 +02:00
for (int k = 2; k <= ncols_pad; k *= 2) {
for (int j = k / 2; j > 0; j /= 2) {
int ixj = col ^ j;
if (ixj > col) {
if ((col & k) == 0) {
ggml : mul_mat_id use the same tensor for all the experts (#6387) * ggml : update mul_mat_id to use the same tensor for all the experts * update cuda * minor * update metal * update test-backend-ops * fix cuda * Update ggml-metal.m Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * update convert.py * update convert-hf-to-gguf.py * update convert.py for mixtral hf models * Update convert-hf-to-gguf.py Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * cuda : support non-pow-2 number of experts * allow quantize to work for split and merged experts models in the same way * cleanup + disable mmap automatically with split tensors models * update imatrix * test-backend-ops : test qwen argsort * update grok model loading * llama : add merged experts tensors to the grok tensor map * minor * gguf : bump version * fix quantizing of merged experts * convert-hf-to-gguf.py : update grok (untested) * make linter happy * cuda/argsort : use shared memory instead of pool memory * convert : fix grok tensor names * metal : add support for non-pow-2 argsort * llama : more loader cleanup, better error checking * cuda : fix warning * llama : still use mmap for loading old models, but copy the data to a host buffer * add review note * llama : remove ffn tensor counting + add sanity check ggml-ci * convert : fix handling of n_experts == None ggml-ci * imatrix : fix ncall counters * llama : produce error if imatrix size does not match * quantize : terminate on errors + trace logs ggml-ci * metal : pad shared memory to 16 bytes --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-04-03 15:07:05 +02:00
if (dst_row[col] >= ncols ||
(dst_row[ixj] < ncols && (order == GGML_SORT_ORDER_ASC ?
x_row[dst_row[col]] > x_row[dst_row[ixj]] :
x_row[dst_row[col]] < x_row[dst_row[ixj]]))
) {
SWAP(dst_row[col], dst_row[ixj]);
}
} else {
ggml : mul_mat_id use the same tensor for all the experts (#6387) * ggml : update mul_mat_id to use the same tensor for all the experts * update cuda * minor * update metal * update test-backend-ops * fix cuda * Update ggml-metal.m Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * update convert.py * update convert-hf-to-gguf.py * update convert.py for mixtral hf models * Update convert-hf-to-gguf.py Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * cuda : support non-pow-2 number of experts * allow quantize to work for split and merged experts models in the same way * cleanup + disable mmap automatically with split tensors models * update imatrix * test-backend-ops : test qwen argsort * update grok model loading * llama : add merged experts tensors to the grok tensor map * minor * gguf : bump version * fix quantizing of merged experts * convert-hf-to-gguf.py : update grok (untested) * make linter happy * cuda/argsort : use shared memory instead of pool memory * convert : fix grok tensor names * metal : add support for non-pow-2 argsort * llama : more loader cleanup, better error checking * cuda : fix warning * llama : still use mmap for loading old models, but copy the data to a host buffer * add review note * llama : remove ffn tensor counting + add sanity check ggml-ci * convert : fix handling of n_experts == None ggml-ci * imatrix : fix ncall counters * llama : produce error if imatrix size does not match * quantize : terminate on errors + trace logs ggml-ci * metal : pad shared memory to 16 bytes --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-04-03 15:07:05 +02:00
if (dst_row[ixj] >= ncols ||
(dst_row[col] < ncols && (order == GGML_SORT_ORDER_ASC ?
x_row[dst_row[col]] < x_row[dst_row[ixj]] :
x_row[dst_row[col]] > x_row[dst_row[ixj]]))
) {
SWAP(dst_row[col], dst_row[ixj]);
}
}
}
threadgroup_barrier(mem_flags::mem_threadgroup);
}
}
ggml : mul_mat_id use the same tensor for all the experts (#6387) * ggml : update mul_mat_id to use the same tensor for all the experts * update cuda * minor * update metal * update test-backend-ops * fix cuda * Update ggml-metal.m Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * update convert.py * update convert-hf-to-gguf.py * update convert.py for mixtral hf models * Update convert-hf-to-gguf.py Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * cuda : support non-pow-2 number of experts * allow quantize to work for split and merged experts models in the same way * cleanup + disable mmap automatically with split tensors models * update imatrix * test-backend-ops : test qwen argsort * update grok model loading * llama : add merged experts tensors to the grok tensor map * minor * gguf : bump version * fix quantizing of merged experts * convert-hf-to-gguf.py : update grok (untested) * make linter happy * cuda/argsort : use shared memory instead of pool memory * convert : fix grok tensor names * metal : add support for non-pow-2 argsort * llama : more loader cleanup, better error checking * cuda : fix warning * llama : still use mmap for loading old models, but copy the data to a host buffer * add review note * llama : remove ffn tensor counting + add sanity check ggml-ci * convert : fix handling of n_experts == None ggml-ci * imatrix : fix ncall counters * llama : produce error if imatrix size does not match * quantize : terminate on errors + trace logs ggml-ci * metal : pad shared memory to 16 bytes --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-04-03 15:07:05 +02:00
// copy the result to dst without the padding
if (col < ncols) {
dst[row * ncols + col] = dst_row[col];
}
}
ggml : mul_mat_id use the same tensor for all the experts (#6387) * ggml : update mul_mat_id to use the same tensor for all the experts * update cuda * minor * update metal * update test-backend-ops * fix cuda * Update ggml-metal.m Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * update convert.py * update convert-hf-to-gguf.py * update convert.py for mixtral hf models * Update convert-hf-to-gguf.py Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * cuda : support non-pow-2 number of experts * allow quantize to work for split and merged experts models in the same way * cleanup + disable mmap automatically with split tensors models * update imatrix * test-backend-ops : test qwen argsort * update grok model loading * llama : add merged experts tensors to the grok tensor map * minor * gguf : bump version * fix quantizing of merged experts * convert-hf-to-gguf.py : update grok (untested) * make linter happy * cuda/argsort : use shared memory instead of pool memory * convert : fix grok tensor names * metal : add support for non-pow-2 argsort * llama : more loader cleanup, better error checking * cuda : fix warning * llama : still use mmap for loading old models, but copy the data to a host buffer * add review note * llama : remove ffn tensor counting + add sanity check ggml-ci * convert : fix handling of n_experts == None ggml-ci * imatrix : fix ncall counters * llama : produce error if imatrix size does not match * quantize : terminate on errors + trace logs ggml-ci * metal : pad shared memory to 16 bytes --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-04-03 15:07:05 +02:00
template [[host_name("kernel_argsort_f32_i32_asc")]] kernel argsort_t kernel_argsort_f32_i32<GGML_SORT_ORDER_ASC>;
template [[host_name("kernel_argsort_f32_i32_desc")]] kernel argsort_t kernel_argsort_f32_i32<GGML_SORT_ORDER_DESC>;
kernel void kernel_leaky_relu_f32(
device const float * src0,
device float * dst,
constant float & slope,
uint tpig[[thread_position_in_grid]]) {
dst[tpig] = src0[tpig] > 0.0f ? src0[tpig] : src0[tpig] * slope;
}
ggml : add Flash Attention (#5021) * ggml : add ggml_flash_attn_ext API * ggml : fix GQA support in ggml_flash_attn_ext * ggml : online attention (CPU) * metal : initial implementation * metal : f16 precision * metal : reduce branches * metal : specialize for head size * wip : 8 rows per simd group * wip : 4 rows per simd group * wip : template for rows per warp * metal : parallelize across KV size * metal : parallel reduce across heads * metal : efficient flash_attn_f16 implementation * metal : avoid redundant loads of the attention * metal : scale and mask in matrix form * metal : fix comment * llama : avoid ggml_cast, use F32 query * metal : add parallel reduce version (disabled) * metal : move output into local memory + optimize - the result from each simdgroup now stays in the registers - significantly reduced SRAM usage - more efficient skipping of -INF blocks - avoid simdgroup barrier in hot loop - add comments * metal : add tests, fix scaling, support C > 32 * metal : improve precision * ggml : fix f16 mad * metal : minor * metal : support Q > 8 * tests : add ATTN tests * metal : disable buffer allocation logs * tests : more * metal : faster inner loop for C == 32 * metal : fix array initialization * tests : ifdef * ggml : switch to padded F16 mask for ggml_soft_max, ggml_flash_attn_ext * ggml : fix ggml_soft_max mask requirement * cuda : fix soft_max to use correct mask size * cuda : add flash_attn kernel (wip) * metal : optimize softmax for C > 32 * metal : optimize softmax * tests : minor fix * cuda : avoid zeroing fragments * tests : update dims * cuda : fix __hisinf() result check * cuda : avoid warp_reduce for smax * cuda : use int instead of int64_t Noticeably improves performance (thanks to Johannes) * cuda : make loops use the same loop values Thanks Johannes again for the tip * cuda : unroll some of the loops * cuda : avoid __hisinf branches * cuda : use half2 in softmax * cuda : switch to 1 warp for bs > 16 * cuda : speed-up reduce part of the kernel * cuda : unroll Q*K^T loop * cuda : fix -INF block check * cuda : simplify softmax * cuda : fix matrix names * cuda : minor * llama : adapt to F16 KQ_pos * llama : adapt new models to F16 KQ_mask * ggml : fix F16 store (ARM NEON) * llama : fix type of KQ_mask and KQ_pos * ggml : fix CPU soft_max * tests : add hs=256 * cuda : fix build * metal : improve perf via smaller int registers * cuda : adapt soft_max to F16 mask and pos * CUDA: faster FlashAttention, kernel for bs == 1 * 16 cols for Phi-2 * no vec for hs, no hs==256 ncols==32 for Volta * adjust kernel selection logic * 4 warps, 256 stride for all D * no ncols == 64 * Multiple parallel blocks for batch size 1 * fix compile warnings * fix excessive KQ_b loads * fix cmake build * fix KV cache padding, NaN from INFINITY (#6438) * llama : flash_attn cparam + fix defrag * server: support flash_attn param * server: bench: enable flash_attn param * CUDA: refactor host code, dyn. par. blocks * fix flash_attn_vec_f16 race condition * flush softmax exp below threshold to 0 * store temp KQ in registers * Calculate KQ as FP32 if KQV has GGML_PREC_F32 * Add __hgt2_mask implementation for CUDA 11 * fix KQ FP32 precision fpr parallel_blocks > 1 * llama-bench : add -fa,--flash-attn arg * metal : add BS=1 kernel for flash attention (#6508) * metal : add BS=1 kernel for flash attention (wip) * metal : support more than 1 warps * metal : opts * metal : opt * metal : switch to parallel reduce * metal : reduce registers * metal : simplify * metal : initial FA vec kernel * metal : use F32 attention accumulators * batched-bench : add fattn arg * llama : simplify llama_build_kv_store ggml-ci * llama : adapt build_olmo to changes * ggml : fix arm fp16 store on windows * metal : clean-up * metal : clean-up kernel code * metal : minor * tests : remove benchmarks ggml-ci * ggml : fix avx512 const correctness ggml-ci * ggml : fix soft_max with bias on CPU ggml-ci * common : print --flash-attn in help * ggml : fix num dimensions in ggml_flash_attn_ext * llama : force disable flash attention for incompatible models * ggml : ggml_soft_max support F16/F32 mask/pos ggml-ci * cuda : uint -> uint32_t * cuda : "constexpr dim3" -> "const dim3" ggml-ci * cuda : try to fix __hgt2_mask ggml-ci * ggml : add TODO's for F16/F32 mask/pos support in other backends * llama : replace bool need_kq_pos with use_alibi * llama : prep ALiBi support for BERT models ggml-ci * llama : fix n_batch requirements ggml-ci * cont * server : add help for --flash-attn arg * llama : disable FA for AMD * tests : remove TMP_ATTN_BENCH ggml-ci * llama : support save/load state with FA enabled ggml-ci * ci : add CUDA save-load-state tests ggml-ci * llama : llama_kv_cache_clear zeroes data + fix save-load seq ggml-ci * llama : fix copy-paste errors, add TODO * llama : disallow incompatible states * llama : update llama_state_get_size after v_trans field * metal : remove tmp log * llama : add static reminder for llama_state_get_size * metal : fix max nsg ggml-ci * ci : fix arg order ggml-ci --------- Co-authored-by: Johannes Gäßler <johannesg@5d6.de> Co-authored-by: Pierrick HYMBERT <pierrick.hymbert@gmail.com>
2024-04-30 11:16:08 +02:00
typedef void (flash_attn_ext_f16_t)(
device const char * q,
device const char * k,
device const char * v,
device const char * mask,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
constant int64_t & ne03,
constant uint64_t & nb00,
constant uint64_t & nb01,
constant uint64_t & nb02,
constant uint64_t & nb03,
constant int64_t & ne10,
constant int64_t & ne11,
constant int64_t & ne12,
constant int64_t & ne13,
constant uint64_t & nb10,
constant uint64_t & nb11,
constant uint64_t & nb12,
constant uint64_t & nb13,
constant int64_t & ne31,
constant uint64_t & nb31,
constant int64_t & ne0,
constant int64_t & ne1,
constant int64_t & ne2,
constant int64_t & ne3,
constant float & scale,
threadgroup half * shared,
uint3 tgpig[[threadgroup_position_in_grid]],
uint3 tpitg[[thread_position_in_threadgroup]],
uint3 ntg[[threads_per_threadgroup]],
ushort tiisg[[thread_index_in_simdgroup]],
ushort sgitg[[simdgroup_index_in_threadgroup]]);
// ref: https://arxiv.org/pdf/2307.08691.pdf
template<int64_t D, int64_t Q = 8, int64_t C = 32> // head size, queries per threadgroup, cache items per threadgroup
kernel void kernel_flash_attn_ext_f16(
device const char * q,
device const char * k,
device const char * v,
device const char * mask,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
constant int64_t & ne03,
constant uint64_t & nb00,
constant uint64_t & nb01,
constant uint64_t & nb02,
constant uint64_t & nb03,
constant int64_t & ne10,
constant int64_t & ne11,
constant int64_t & ne12,
constant int64_t & ne13,
constant uint64_t & nb10,
constant uint64_t & nb11,
constant uint64_t & nb12,
constant uint64_t & nb13,
constant int64_t & ne31,
constant uint64_t & nb31,
constant int64_t & ne0,
constant int64_t & ne1,
constant int64_t & ne2,
constant int64_t & ne3,
constant float & scale,
threadgroup half * shared [[threadgroup(0)]],
uint3 tgpig[[threadgroup_position_in_grid]],
uint3 tpitg[[thread_position_in_threadgroup]],
uint3 ntg[[threads_per_threadgroup]],
ushort tiisg[[thread_index_in_simdgroup]],
ushort sgitg[[simdgroup_index_in_threadgroup]]) {
const short nsg = ntg.y; // number of simdgroups
const short iq3 = tgpig[2];
const short iq2 = tgpig[1];
const short iq1 = tgpig[0]*Q;
const short D4 = D/4;
const short D8 = D/8;
2024-05-08 08:14:50 +02:00
//const short Q8 = Q/8;
ggml : add Flash Attention (#5021) * ggml : add ggml_flash_attn_ext API * ggml : fix GQA support in ggml_flash_attn_ext * ggml : online attention (CPU) * metal : initial implementation * metal : f16 precision * metal : reduce branches * metal : specialize for head size * wip : 8 rows per simd group * wip : 4 rows per simd group * wip : template for rows per warp * metal : parallelize across KV size * metal : parallel reduce across heads * metal : efficient flash_attn_f16 implementation * metal : avoid redundant loads of the attention * metal : scale and mask in matrix form * metal : fix comment * llama : avoid ggml_cast, use F32 query * metal : add parallel reduce version (disabled) * metal : move output into local memory + optimize - the result from each simdgroup now stays in the registers - significantly reduced SRAM usage - more efficient skipping of -INF blocks - avoid simdgroup barrier in hot loop - add comments * metal : add tests, fix scaling, support C > 32 * metal : improve precision * ggml : fix f16 mad * metal : minor * metal : support Q > 8 * tests : add ATTN tests * metal : disable buffer allocation logs * tests : more * metal : faster inner loop for C == 32 * metal : fix array initialization * tests : ifdef * ggml : switch to padded F16 mask for ggml_soft_max, ggml_flash_attn_ext * ggml : fix ggml_soft_max mask requirement * cuda : fix soft_max to use correct mask size * cuda : add flash_attn kernel (wip) * metal : optimize softmax for C > 32 * metal : optimize softmax * tests : minor fix * cuda : avoid zeroing fragments * tests : update dims * cuda : fix __hisinf() result check * cuda : avoid warp_reduce for smax * cuda : use int instead of int64_t Noticeably improves performance (thanks to Johannes) * cuda : make loops use the same loop values Thanks Johannes again for the tip * cuda : unroll some of the loops * cuda : avoid __hisinf branches * cuda : use half2 in softmax * cuda : switch to 1 warp for bs > 16 * cuda : speed-up reduce part of the kernel * cuda : unroll Q*K^T loop * cuda : fix -INF block check * cuda : simplify softmax * cuda : fix matrix names * cuda : minor * llama : adapt to F16 KQ_pos * llama : adapt new models to F16 KQ_mask * ggml : fix F16 store (ARM NEON) * llama : fix type of KQ_mask and KQ_pos * ggml : fix CPU soft_max * tests : add hs=256 * cuda : fix build * metal : improve perf via smaller int registers * cuda : adapt soft_max to F16 mask and pos * CUDA: faster FlashAttention, kernel for bs == 1 * 16 cols for Phi-2 * no vec for hs, no hs==256 ncols==32 for Volta * adjust kernel selection logic * 4 warps, 256 stride for all D * no ncols == 64 * Multiple parallel blocks for batch size 1 * fix compile warnings * fix excessive KQ_b loads * fix cmake build * fix KV cache padding, NaN from INFINITY (#6438) * llama : flash_attn cparam + fix defrag * server: support flash_attn param * server: bench: enable flash_attn param * CUDA: refactor host code, dyn. par. blocks * fix flash_attn_vec_f16 race condition * flush softmax exp below threshold to 0 * store temp KQ in registers * Calculate KQ as FP32 if KQV has GGML_PREC_F32 * Add __hgt2_mask implementation for CUDA 11 * fix KQ FP32 precision fpr parallel_blocks > 1 * llama-bench : add -fa,--flash-attn arg * metal : add BS=1 kernel for flash attention (#6508) * metal : add BS=1 kernel for flash attention (wip) * metal : support more than 1 warps * metal : opts * metal : opt * metal : switch to parallel reduce * metal : reduce registers * metal : simplify * metal : initial FA vec kernel * metal : use F32 attention accumulators * batched-bench : add fattn arg * llama : simplify llama_build_kv_store ggml-ci * llama : adapt build_olmo to changes * ggml : fix arm fp16 store on windows * metal : clean-up * metal : clean-up kernel code * metal : minor * tests : remove benchmarks ggml-ci * ggml : fix avx512 const correctness ggml-ci * ggml : fix soft_max with bias on CPU ggml-ci * common : print --flash-attn in help * ggml : fix num dimensions in ggml_flash_attn_ext * llama : force disable flash attention for incompatible models * ggml : ggml_soft_max support F16/F32 mask/pos ggml-ci * cuda : uint -> uint32_t * cuda : "constexpr dim3" -> "const dim3" ggml-ci * cuda : try to fix __hgt2_mask ggml-ci * ggml : add TODO's for F16/F32 mask/pos support in other backends * llama : replace bool need_kq_pos with use_alibi * llama : prep ALiBi support for BERT models ggml-ci * llama : fix n_batch requirements ggml-ci * cont * server : add help for --flash-attn arg * llama : disable FA for AMD * tests : remove TMP_ATTN_BENCH ggml-ci * llama : support save/load state with FA enabled ggml-ci * ci : add CUDA save-load-state tests ggml-ci * llama : llama_kv_cache_clear zeroes data + fix save-load seq ggml-ci * llama : fix copy-paste errors, add TODO * llama : disallow incompatible states * llama : update llama_state_get_size after v_trans field * metal : remove tmp log * llama : add static reminder for llama_state_get_size * metal : fix max nsg ggml-ci * ci : fix arg order ggml-ci --------- Co-authored-by: Johannes Gäßler <johannesg@5d6.de> Co-authored-by: Pierrick HYMBERT <pierrick.hymbert@gmail.com>
2024-04-30 11:16:08 +02:00
const short NW = N_SIMDWIDTH;
const short SH = (C + Q); // shared memory per simdgroup in (half)
const short T = D + 2*nsg*SH; // shared memory size per query in (half)
const short TF = T/2; // shared memory size per query in (float)
const short T4 = T/4; // shared memory size per query in (half4)
threadgroup half * sq = (threadgroup half *) (shared + 0*D); // holds the query data
threadgroup half4 * sq4 = (threadgroup half4 *) (shared + 0*D); // same as above but in half4
threadgroup float * ss = (threadgroup float *) (shared + 2*sgitg*SH + 1*D); // scratch buffer for attention and diagonal matrix
// store the result for all queries in local memory in 8x8 matrices (the O matrix from the paper)
simdgroup_half8x8 lo[D8];
// load heads from Q to shared memory
for (short j = sgitg; j < Q; j += nsg) {
device const float4 * q4 = (device const float4 *) ((device const char *) q + ((iq1 + j)*nb01 + iq2*nb02 + iq3*nb03));
for (short i = tiisg; i < D4; i += NW) {
if (iq1 + j < ne01) {
sq4[j*T4 + i] = (half4) q4[i];
} else {
sq4[j*T4 + i] = 0.0h;
}
}
}
// zero out lo
for (short i = 0; i < D8; ++i) {
lo[i] = make_filled_simdgroup_matrix<half, 8>(0.0h);
}
// zero out shared memory SH
for (short j = 0; j < Q; ++j) {
for (short i = tiisg; i < SH; i += NW) {
ss[j*TF + i] = 0.0f;
}
}
threadgroup_barrier(mem_flags::mem_threadgroup);
{
float S[Q] = { [0 ... Q-1] = 0.0h };
float M[Q] = { [0 ... Q-1] = -FLT_MAX/2 };
// assume K and V are same shape
const short ne22 = ne12;
const short ne23 = ne13;
const uint nb21 = nb11;
const uint nb22 = nb12;
const uint nb23 = nb13;
// broadcast
const short rk2 = ne02/ne12;
const short rk3 = ne03/ne13;
const short rv2 = ne02/ne22;
const short rv3 = ne03/ne23;
// k indices
const short ik2 = iq2/rk2;
const short ik3 = iq3/rk3;
// v indices
const short iv2 = iq2/rv2;
const short iv3 = iq3/rv3;
// load the queries from shared memory into local memory
simdgroup_half8x8 mq[D8];
for (short i = 0; i < D8; ++i) {
simdgroup_load(mq[i], sq + i*8, T);
}
// pointer to the mask
device const half * mp = (device const half *) (mask + iq1*nb31);
// prepare diagonal scale matrix
simdgroup_float8x8 mscale(scale);
// loop over the KV cache
// each simdgroup handles blocks of Q rows and C columns
for (int ic0 = 0; ic0 < ne11; ic0 += C*nsg) {
const int ic = ic0 + C*sgitg;
if (ic >= ne11) {
break;
}
// Q*K^T
{
for (short cc = 0; cc < C/8; ++cc) {
simdgroup_float8x8 mqk = make_filled_simdgroup_matrix<float, 8>(0.h);
device const half * pk = (device const half *) ((device const char *) k + ((ic + 8*cc)*nb11 + ik2*nb12 + ik3*nb13));
for (short i = 0; i < D8; ++i) {
simdgroup_half8x8 mk;
simdgroup_load(mk, pk + i*8, nb11/sizeof(half), 0, true); // transpose
simdgroup_multiply_accumulate(mqk, mq[i], mk, mqk);
}
// mqk = mqk*scale + mask
simdgroup_half8x8 mm;
simdgroup_load(mm, mp + ic + 8*cc, nb31/sizeof(half), 0, false);
simdgroup_multiply_accumulate(mqk, mqk, mscale, mm);
simdgroup_store(mqk, ss + 8*cc, TF, 0, false);
}
}
// used to detect blocks full of -INF
float smax = -INFINITY;
// online softmax
{
float ms[Q];
for (short j = 0; j < Q; ++j) {
const short p = tiisg;
const float m = M[j];
const float s = ss[j*TF + p];
smax = simd_max(max(smax, s));
M[j] = simd_max(max(M[j], s));
ms[j] = exp(m - M[j]);
const float vs = exp(s - M[j]);
S[j] = S[j]*ms[j] + simd_sum(vs);
// the P matrix from the paper (Q rows, C columns)
ss[j*TF + p] = vs;
}
// create a QxQ diagonal matrix for rescaling the output
if (tiisg < Q) {
ss[tiisg*TF + C + tiisg] = ms[tiisg];
}
}
// skip -INF blocks
if (smax == -INFINITY) {
continue;
}
// O = diag(ms)*O
{
simdgroup_float8x8 mm;
simdgroup_load(mm, ss + C, TF, 0, false);
for (short i = 0; i < D8; ++i) {
simdgroup_multiply(lo[i], mm, lo[i]);
}
}
// O = O + (Q*K^T)*V
{
for (short cc = 0; cc < C/8; ++cc) {
device const half * pv = (device const half *) ((device const char *) v + ((ic + 8*cc)*nb21 + iv2*nb22 + iv3*nb23));
for (short i = 0; i < D8; ++i) {
simdgroup_half8x8 mk;
simdgroup_load(mk, pv + i*8, nb21/sizeof(half), 0, false);
simdgroup_float8x8 mv;
simdgroup_load(mv, ss + 8*cc, TF, 0, false);
simdgroup_multiply_accumulate(lo[i], mv, mk, lo[i]);
}
}
}
}
// these are needed for reducing the results from the simdgroups (reuse the ss buffer)
for (short j = 0; j < Q; ++j) {
if (tiisg == 0) {
ss[j*TF + 0] = S[j];
ss[j*TF + 1] = M[j];
}
}
}
// reduce the warps sequentially
for (short sg = 1; sg < nsg; ++sg) {
float S = { 0.0h };
float M = { -FLT_MAX/2 };
threadgroup_barrier(mem_flags::mem_threadgroup);
// each simdgroup stores its output to shared memory, reusing sq
if (sgitg == sg) {
for (short i = 0; i < D8; ++i) {
simdgroup_store(lo[i], sq + i*8, T, 0, false);
}
}
threadgroup_barrier(mem_flags::mem_threadgroup);
// the first simdgroup accumulates the results from the other simdgroups
if (sgitg == 0) {
for (short j = 0; j < Q; ++j) {
const float S0 = ss[j*TF + 0];
const float S1 = ss[j*TF + sg*SH + 0];
const float M0 = ss[j*TF + 1];
const float M1 = ss[j*TF + sg*SH + 1];
M = max(M0, M1);
const float ms0 = exp(M0 - M);
const float ms1 = exp(M1 - M);
S = S0*ms0 + S1*ms1;
if (tiisg == 0) {
ss[j*TF + 0] = S;
ss[j*TF + 1] = M;
ss[j*TF + C + j ] = ms0;
ss[j*TF + C + j + sg*SH] = ms1;
}
}
// O_0 = diag(ms0)*O_0 + diag(ms1)*O_1
{
simdgroup_half8x8 t;
simdgroup_float8x8 ms0;
simdgroup_float8x8 ms1;
simdgroup_load(ms0, ss + C, TF, 0, false);
simdgroup_load(ms1, ss + C + sg*SH, TF, 0, false);
for (short i = 0; i < D8; ++i) {
simdgroup_load (t, sq + i*8, T, 0, false);
simdgroup_multiply(t, ms1, t);
simdgroup_multiply_accumulate(lo[i], ms0, lo[i], t);
}
}
}
}
// store result to shared memory (reuse sq)
if (sgitg == 0) {
for (short i = 0; i < D8; ++i) {
simdgroup_store(lo[i], sq + i*8, T, 0, false);
}
}
device float4 * dst4 = (device float4 *) dst;
// final rescale with 1/S and store to global memory
if (sgitg == 0) {
for (short j = 0; j < Q && iq1 + j < ne01; ++j) {
const float S = ss[j*TF + 0];
for (short i = tiisg; i < D4; i += NW) {
dst4[(iq3*ne2*ne1 + iq2 + (iq1 + j)*ne1)*D4 + i] = (float4) sq4[j*T4 + i]/S;
}
}
}
}
template [[host_name("kernel_flash_attn_ext_f16_h64" )]] kernel flash_attn_ext_f16_t kernel_flash_attn_ext_f16<64>;
template [[host_name("kernel_flash_attn_ext_f16_h80" )]] kernel flash_attn_ext_f16_t kernel_flash_attn_ext_f16<80>;
template [[host_name("kernel_flash_attn_ext_f16_h96" )]] kernel flash_attn_ext_f16_t kernel_flash_attn_ext_f16<96>;
template [[host_name("kernel_flash_attn_ext_f16_h112")]] kernel flash_attn_ext_f16_t kernel_flash_attn_ext_f16<112>;
template [[host_name("kernel_flash_attn_ext_f16_h128")]] kernel flash_attn_ext_f16_t kernel_flash_attn_ext_f16<128>;
template [[host_name("kernel_flash_attn_ext_f16_h256")]] kernel flash_attn_ext_f16_t kernel_flash_attn_ext_f16<256>;
template<int64_t D, int64_t Q = 1, int64_t C = 32> // head size, queries per threadgroup, cache items per threadgroup
kernel void kernel_flash_attn_ext_vec_f16(
device const char * q,
device const char * k,
device const char * v,
device const char * mask,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
constant int64_t & ne03,
constant uint64_t & nb00,
constant uint64_t & nb01,
constant uint64_t & nb02,
constant uint64_t & nb03,
constant int64_t & ne10,
constant int64_t & ne11,
constant int64_t & ne12,
constant int64_t & ne13,
constant uint64_t & nb10,
constant uint64_t & nb11,
constant uint64_t & nb12,
constant uint64_t & nb13,
constant int64_t & ne31,
constant uint64_t & nb31,
constant int64_t & ne0,
constant int64_t & ne1,
constant int64_t & ne2,
constant int64_t & ne3,
constant float & scale,
threadgroup half * shared [[threadgroup(0)]],
uint3 tgpig[[threadgroup_position_in_grid]],
uint3 tpitg[[thread_position_in_threadgroup]],
uint3 ntg[[threads_per_threadgroup]],
ushort tiisg[[thread_index_in_simdgroup]],
ushort sgitg[[simdgroup_index_in_threadgroup]]) {
const short nsg = ntg.y; // number of simdgroups
const short iq3 = tgpig[2];
const short iq2 = tgpig[1];
const short iq1 = tgpig[0];
const short D4 = D/4;
const short NW = N_SIMDWIDTH;
const short SH = (C + Q); // shared memory per simdgroup in (half)
const short T = D + 2*nsg*SH; // shared memory size per query in (half)
//threadgroup half * sq = (threadgroup half *) (shared + 0*D); // holds the query data
threadgroup half4 * sq4 = (threadgroup half4 *) (shared + 0*D); // same as above but in half4
threadgroup float * ss = (threadgroup float *) (shared + 2*sgitg*SH + 1*D); // scratch buffer for attention and diagonal matrix
threadgroup float4 * ss4 = (threadgroup float4 *) (shared + 2*sgitg*SH + 1*D); // same as above but in half4
threadgroup half4 * sr4 = (threadgroup half4 *) (shared + sgitg*D + 1*T); // scratch buffer for the results
// store the result for all queries in local memory in 8x8 matrices (the O matrix from the paper)
half4 lo[D4/NW];
// load heads from Q to shared memory
device const float4 * q4 = (device const float4 *) ((device const char *) q + (iq1*nb01 + iq2*nb02 + iq3*nb03));
for (short i = tiisg; i < D4; i += NW) {
if (iq1 < ne01) {
sq4[i] = (half4) q4[i];
} else {
sq4[i] = 0.0h;
}
}
// zero out lo
for (short i = tiisg; i < D4; i += NW) {
lo[i/NW] = 0.0h;
}
// zero out shared memory SH
for (short i = tiisg; i < SH/4; i += NW) {
ss4[i] = 0.0h;
}
threadgroup_barrier(mem_flags::mem_threadgroup);
{
float S = { 0.0h };
float M = { -FLT_MAX/2 };
// assume K and V are same shape
const short ne22 = ne12;
const short ne23 = ne13;
const uint nb21 = nb11;
const uint nb22 = nb12;
const uint nb23 = nb13;
// broadcast
const short rk2 = ne02/ne12;
const short rk3 = ne03/ne13;
const short rv2 = ne02/ne22;
const short rv3 = ne03/ne23;
// k indices
const short ik2 = iq2 / rk2;
const short ik3 = iq3 / rk3;
// v indices
const short iv2 = iq2 / rv2;
const short iv3 = iq3 / rv3;
// load the queries from shared memory into local memory
half4 mq[D4];
for (short ii = 0; ii < D4; ii += NW) {
short i = ii + tiisg;
mq[i] = sq4[i];
}
// pointer to the mask
device const half4 * mp4 = (device const half4 *) (mask + iq1*nb31);
// loop over the KV cache
// each simdgroup handles blocks of Q rows and C columns
for (int ic0 = 0; ic0 < ne11; ic0 += C*nsg) {
const int ic = ic0 + C*sgitg;
if (ic >= ne11) {
break;
}
// Q*K^T
{
#pragma unroll
for (short cc = 0; cc < C/4; ++cc) {
float4 mqk = { 0.0h };
device const half4 * pk4 = (device const half4 *) ((device const char *) k + ((ic + 4*cc)*nb11 + ik2*nb12 + ik3*nb13));
#pragma unroll
for (short ii = 0; ii < D4; ii += NW) {
const short i = ii + tiisg;
half4x4 mk;
mk[0] = pk4[i + 0*(nb11/8)];
mk[1] = pk4[i + 1*(nb11/8)];
mk[2] = pk4[i + 2*(nb11/8)];
mk[3] = pk4[i + 3*(nb11/8)];
mqk += (float4) (mq[i] * mk);
}
// reduce the results from the threads in the simdgroup
mqk += simd_shuffle_down(mqk, 16);
mqk += simd_shuffle_down(mqk, 8);
mqk += simd_shuffle_down(mqk, 4);
mqk += simd_shuffle_down(mqk, 2);
mqk += simd_shuffle_down(mqk, 1);
// mqk = mqk*scale + mask
if (tiisg == 0) {
float4 mm = (float4) mp4[ic/4 + cc];
mqk = mqk*scale + mm;
ss4[cc] = mqk;
}
}
}
// online softmax
{
const short p = tiisg;
const float m = M;
const float s = ss[p];
M = simd_max(max(M, s));
const float ms = exp(m - M);
const float vs = exp(s - M);
S = S*ms + simd_sum(vs);
// the P matrix from the paper (Q rows, C columns)
ss[p] = vs;
// O = diag(ms)*O
#pragma unroll
for (short ii = 0; ii < D4; ii += NW) {
const short i = ii + tiisg;
lo[i/NW] *= ms;
}
}
// O = O + (Q*K^T)*V
{
#pragma unroll
for (short cc = 0; cc < C/4; ++cc) {
device const half4 * pv4 = (device const half4 *) ((device const char *) v + ((ic + 4*cc)*nb21 + iv2*nb22 + iv3*nb23));
#pragma unroll
for (short ii = 0; ii < D4; ii += NW) {
const short i = ii + tiisg;
lo[i/NW] += pv4[i + 0*(nb21/8)] * ss[4*cc + 0];
lo[i/NW] += pv4[i + 1*(nb21/8)] * ss[4*cc + 1];
lo[i/NW] += pv4[i + 2*(nb21/8)] * ss[4*cc + 2];
lo[i/NW] += pv4[i + 3*(nb21/8)] * ss[4*cc + 3];
}
}
}
}
// these are needed for reducing the results from the simdgroups (reuse the ss buffer)
if (tiisg == 0) {
ss[0] = S;
ss[1] = M;
}
}
// store results to shared memory
for (short ii = 0; ii < D4; ii += NW) {
short i = ii + tiisg;
sr4[i] = lo[ii/NW];
}
threadgroup_barrier(mem_flags::mem_threadgroup);
// parallel reduce
for (short r = nsg/2; r > 0; r >>= 1) {
if (sgitg < r) {
const float S0 = ss[ 0];
const float S1 = ss[r*SH + 0];
const float M0 = ss[ 1];
const float M1 = ss[r*SH + 1];
const float M = max(M0, M1);
const float ms0 = exp(M0 - M);
const float ms1 = exp(M1 - M);
const float S = S0*ms0 + S1*ms1;
if (tiisg == 0) {
ss[0] = S;
ss[1] = M;
}
// O_0 = diag(ms0)*O_0 + diag(ms1)*O_1
for (short ii = 0; ii < D4; ii += NW) {
short i = ii + tiisg;
sr4[i] = sr4[i]*ms0 + sr4[i + r*D4]*ms1;
}
}
threadgroup_barrier(mem_flags::mem_threadgroup);
}
device float4 * dst4 = (device float4 *) dst;
// final rescale with 1/S and store to global memory
if (sgitg == 0) {
const float S = ss[0];
for (short ii = 0; ii < D4; ii += NW) {
short i = ii + tiisg;
dst4[(iq3*ne2*ne1 + iq2 + (iq1)*ne1)*D4 + i] = (float4) sr4[i]/S;
}
}
}
template [[host_name("kernel_flash_attn_ext_vec_f16_h128")]] kernel flash_attn_ext_f16_t kernel_flash_attn_ext_vec_f16<128>;
template [[host_name("kernel_flash_attn_ext_vec_f16_h256")]] kernel flash_attn_ext_f16_t kernel_flash_attn_ext_vec_f16<256>;
kernel void kernel_cpy_f16_f16(
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 13:04:25 +01:00
device const half * src0,
device half * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
constant int64_t & ne03,
constant uint64_t & nb00,
constant uint64_t & nb01,
constant uint64_t & nb02,
constant uint64_t & nb03,
constant int64_t & ne0,
constant int64_t & ne1,
constant int64_t & ne2,
constant int64_t & ne3,
constant uint64_t & nb0,
constant uint64_t & nb1,
constant uint64_t & nb2,
constant uint64_t & nb3,
uint3 tgpig[[threadgroup_position_in_grid]],
uint3 tpitg[[thread_position_in_threadgroup]],
uint3 ntg[[threads_per_threadgroup]]) {
const int64_t i03 = tgpig[2];
const int64_t i02 = tgpig[1];
const int64_t i01 = tgpig[0];
const int64_t n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
const int64_t i3 = n / (ne2*ne1*ne0);
const int64_t i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0);
const int64_t i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0;
const int64_t i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0);
device half * dst_data = (device half *) ((device char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
for (int64_t i00 = tpitg.x; i00 < ne00; i00 += ntg.x) {
device const half * src = (device half *)((device char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
dst_data[i00] = src[0];
}
}
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 13:04:25 +01:00
kernel void kernel_cpy_f16_f32(
device const half * src0,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
constant int64_t & ne03,
constant uint64_t & nb00,
constant uint64_t & nb01,
constant uint64_t & nb02,
constant uint64_t & nb03,
constant int64_t & ne0,
constant int64_t & ne1,
constant int64_t & ne2,
constant int64_t & ne3,
constant uint64_t & nb0,
constant uint64_t & nb1,
constant uint64_t & nb2,
constant uint64_t & nb3,
uint3 tgpig[[threadgroup_position_in_grid]],
uint3 tpitg[[thread_position_in_threadgroup]],
uint3 ntg[[threads_per_threadgroup]]) {
const int64_t i03 = tgpig[2];
const int64_t i02 = tgpig[1];
const int64_t i01 = tgpig[0];
const int64_t n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
const int64_t i3 = n / (ne2*ne1*ne0);
const int64_t i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0);
const int64_t i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0;
const int64_t i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0);
device float * dst_data = (device float *) ((device char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
for (int64_t i00 = tpitg.x; i00 < ne00; i00 += ntg.x) {
device const half * src = (device half *)((device char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
dst_data[i00] = src[0];
}
}
llama : Metal inference (#1642) * mtl : export the LLaMA computation graph * ci : disable temporary * mtl : adapt the MNIST example as starter * mtl : no need for mtl-export tool, add cli arg for main instead * mtl : export just a small part of the graph for now to make it easier * mtl : move MSL code into separate file for easy editing * mtl : initial get_rows_q4_0 kernel * mtl : confirmed get_rows_q4_0 is working correctly * mtl : add rms_norm kernel + confirm working * mtl : add mul kernel + confirm working * mtl : initial mul_mat Q4 kernel (wrong results) * mtl : mul_mat fixes (still wrong) * mtl : another mul_mat Q4 (still does not work) * mtl : working mul_mat q4 * ggml : fix handling of "view" ops in ggml_graph_import() * mtl : add rope kernel * mtl : add reshape and transpose handling * ggml : store offset as opt arg for ggml_view_xd() operators * mtl : add cpy kernel + handle view ops * mtl : confirm f16 x f32 attention mul mat * mtl : add scale kernel * mtl : add diag_mask_inf kernel * mtl : fix soft_max kernel * ggml : update ggml_nbytes() to handle non-contiguous tensors * mtl : verify V tensor contents * mtl : add f32 -> f32 cpy kernel * mtl : add silu kernel * mtl : add non-broadcast mul kernel * mtl : full GPU inference of the computation graph * mtl : optimize rms_norm and soft_max kernels * mtl : add f16 mat x f32 vec multiplication kernel * mtl : fix bug in f16 x f32 mul mat + speed-up computation * mtl : faster mul_mat_q4_0_f32 kernel * mtl : fix kernel signature + roll inner loop * mtl : more threads for rms_norm + better timing * mtl : remove printfs from inner loop * mtl : simplify implementation * mtl : add save/load vocab to ggml file * mtl : plug Metal inference into llama.cpp (very quick-n-dirty) * mtl : make it work with main example Lots of hacks but at least now it generates text * mtl : preparing for merge * mtl : clean-up ggml mtl interface + suport scratch / inplace * mtl : remove temp / debug code * metal : final refactoring and simplification * Revert "ci : disable temporary" This reverts commit 98c267fc77fe811082f672538fc91bcfc9072d63. * metal : add comments * metal : clean-up stuff, fix typos * readme : add Metal instructions * readme : add example for main
2023-06-04 22:34:30 +02:00
kernel void kernel_cpy_f32_f16(
device const float * src0,
device half * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
constant int64_t & ne03,
constant uint64_t & nb00,
constant uint64_t & nb01,
constant uint64_t & nb02,
constant uint64_t & nb03,
constant int64_t & ne0,
constant int64_t & ne1,
constant int64_t & ne2,
constant int64_t & ne3,
constant uint64_t & nb0,
constant uint64_t & nb1,
constant uint64_t & nb2,
constant uint64_t & nb3,
uint3 tgpig[[threadgroup_position_in_grid]],
uint3 tpitg[[thread_position_in_threadgroup]],
uint3 ntg[[threads_per_threadgroup]]) {
const int64_t i03 = tgpig[2];
const int64_t i02 = tgpig[1];
const int64_t i01 = tgpig[0];
const int64_t n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
const int64_t i3 = n / (ne2*ne1*ne0);
const int64_t i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0);
const int64_t i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0;
const int64_t i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0);
device half * dst_data = (device half *) ((device char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
for (int64_t i00 = tpitg.x; i00 < ne00; i00 += ntg.x) {
device const float * src = (device float *)((device char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
dst_data[i00] = src[0];
}
}
kernel void kernel_cpy_f32_f32(
device const float * src0,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
constant int64_t & ne03,
constant uint64_t & nb00,
constant uint64_t & nb01,
constant uint64_t & nb02,
constant uint64_t & nb03,
constant int64_t & ne0,
constant int64_t & ne1,
constant int64_t & ne2,
constant int64_t & ne3,
constant uint64_t & nb0,
constant uint64_t & nb1,
constant uint64_t & nb2,
constant uint64_t & nb3,
uint3 tgpig[[threadgroup_position_in_grid]],
uint3 tpitg[[thread_position_in_threadgroup]],
uint3 ntg[[threads_per_threadgroup]]) {
const int64_t i03 = tgpig[2];
const int64_t i02 = tgpig[1];
const int64_t i01 = tgpig[0];
const int64_t n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
const int64_t i3 = n / (ne2*ne1*ne0);
const int64_t i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0);
const int64_t i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0;
const int64_t i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0);
device float * dst_data = (device float *) ((device char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
for (int64_t i00 = tpitg.x; i00 < ne00; i00 += ntg.x) {
device const float * src = (device float *)((device char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
dst_data[i00] = src[0];
}
}
kernel void kernel_cpy_f32_q8_0(
device const float * src0,
device void * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
constant int64_t & ne03,
constant uint64_t & nb00,
constant uint64_t & nb01,
constant uint64_t & nb02,
constant uint64_t & nb03,
constant int64_t & ne0,
constant int64_t & ne1,
constant int64_t & ne2,
constant int64_t & ne3,
constant uint64_t & nb0,
constant uint64_t & nb1,
constant uint64_t & nb2,
constant uint64_t & nb3,
uint3 tgpig[[threadgroup_position_in_grid]],
uint3 tpitg[[thread_position_in_threadgroup]],
uint3 ntg[[threads_per_threadgroup]]) {
const int64_t i03 = tgpig[2];
const int64_t i02 = tgpig[1];
const int64_t i01 = tgpig[0];
const int64_t n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
const int64_t i3 = n / (ne2*ne1*ne0);
const int64_t i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0);
const int64_t i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0;
const int64_t i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0)/QK8_0;
device block_q8_0 * dst_data = (device block_q8_0 *) ((device char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
for (int64_t i00 = tpitg.x*QK8_0; i00 < ne00; i00 += ntg.x*QK8_0) {
device const float * src = (device float *)((device char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
float amax = 0.0f; // absolute max
for (int j = 0; j < QK8_0; j++) {
const float v = src[j];
amax = MAX(amax, fabs(v));
}
const float d = amax / ((1 << 7) - 1);
const float id = d ? 1.0f/d : 0.0f;
dst_data[i00/QK8_0].d = d;
for (int j = 0; j < QK8_0; ++j) {
const float x0 = src[j]*id;
dst_data[i00/QK8_0].qs[j] = round(x0);
}
}
}
kernel void kernel_cpy_f32_q4_0(
device const float * src0,
device void * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
constant int64_t & ne03,
constant uint64_t & nb00,
constant uint64_t & nb01,
constant uint64_t & nb02,
constant uint64_t & nb03,
constant int64_t & ne0,
constant int64_t & ne1,
constant int64_t & ne2,
constant int64_t & ne3,
constant uint64_t & nb0,
constant uint64_t & nb1,
constant uint64_t & nb2,
constant uint64_t & nb3,
uint3 tgpig[[threadgroup_position_in_grid]],
uint3 tpitg[[thread_position_in_threadgroup]],
uint3 ntg[[threads_per_threadgroup]]) {
const int64_t i03 = tgpig[2];
const int64_t i02 = tgpig[1];
const int64_t i01 = tgpig[0];
const int64_t n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
const int64_t i3 = n / (ne2*ne1*ne0);
const int64_t i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0);
const int64_t i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0;
const int64_t i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0)/QK4_0;
device block_q4_0 * dst_data = (device block_q4_0 *) ((device char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
for (int64_t i00 = tpitg.x*QK4_0; i00 < ne00; i00 += ntg.x*QK4_0) {
device const float * src = (device float *)((device char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
float amax = 0.0f; // absolute max
float max = 0.0f;
for (int j = 0; j < QK4_0; j++) {
const float v = src[j];
if (amax < fabs(v)) {
amax = fabs(v);
max = v;
}
}
const float d = max / -8;
const float id = d ? 1.0f/d : 0.0f;
dst_data[i00/QK4_0].d = d;
for (int j = 0; j < QK4_0/2; ++j) {
const float x0 = src[0 + j]*id;
const float x1 = src[QK4_0/2 + j]*id;
const uint8_t xi0 = MIN(15, (int8_t)(x0 + 8.5f));
const uint8_t xi1 = MIN(15, (int8_t)(x1 + 8.5f));
dst_data[i00/QK4_0].qs[j] = xi0;
dst_data[i00/QK4_0].qs[j] |= xi1 << 4;
}
}
}
kernel void kernel_cpy_f32_q4_1(
device const float * src0,
device void * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
constant int64_t & ne03,
constant uint64_t & nb00,
constant uint64_t & nb01,
constant uint64_t & nb02,
constant uint64_t & nb03,
constant int64_t & ne0,
constant int64_t & ne1,
constant int64_t & ne2,
constant int64_t & ne3,
constant uint64_t & nb0,
constant uint64_t & nb1,
constant uint64_t & nb2,
constant uint64_t & nb3,
uint3 tgpig[[threadgroup_position_in_grid]],
uint3 tpitg[[thread_position_in_threadgroup]],
uint3 ntg[[threads_per_threadgroup]]) {
const int64_t i03 = tgpig[2];
const int64_t i02 = tgpig[1];
const int64_t i01 = tgpig[0];
const int64_t n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
const int64_t i3 = n / (ne2*ne1*ne0);
const int64_t i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0);
const int64_t i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0;
const int64_t i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0)/QK4_1;
device block_q4_1 * dst_data = (device block_q4_1 *) ((device char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
for (int64_t i00 = tpitg.x*QK4_1; i00 < ne00; i00 += ntg.x*QK4_1) {
device const float * src = (device float *)((device char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
float min = FLT_MAX;
float max = -FLT_MAX;
for (int j = 0; j < QK4_1; j++) {
const float v = src[j];
if (min > v) min = v;
if (max < v) max = v;
}
const float d = (max - min) / ((1 << 4) - 1);
const float id = d ? 1.0f/d : 0.0f;
dst_data[i00/QK4_1].d = d;
dst_data[i00/QK4_1].m = min;
for (int j = 0; j < QK4_1/2; ++j) {
const float x0 = (src[0 + j] - min)*id;
const float x1 = (src[QK4_1/2 + j] - min)*id;
const uint8_t xi0 = MIN(15, (int8_t)(x0 + 0.5f));
const uint8_t xi1 = MIN(15, (int8_t)(x1 + 0.5f));
dst_data[i00/QK4_1].qs[j] = xi0;
dst_data[i00/QK4_1].qs[j] |= xi1 << 4;
}
}
}
kernel void kernel_cpy_f32_q5_0(
device const float * src0,
device void * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
constant int64_t & ne03,
constant uint64_t & nb00,
constant uint64_t & nb01,
constant uint64_t & nb02,
constant uint64_t & nb03,
constant int64_t & ne0,
constant int64_t & ne1,
constant int64_t & ne2,
constant int64_t & ne3,
constant uint64_t & nb0,
constant uint64_t & nb1,
constant uint64_t & nb2,
constant uint64_t & nb3,
uint3 tgpig[[threadgroup_position_in_grid]],
uint3 tpitg[[thread_position_in_threadgroup]],
uint3 ntg[[threads_per_threadgroup]]) {
const int64_t i03 = tgpig[2];
const int64_t i02 = tgpig[1];
const int64_t i01 = tgpig[0];
const int64_t n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
const int64_t i3 = n / (ne2*ne1*ne0);
const int64_t i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0);
const int64_t i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0;
const int64_t i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0)/QK5_0;
device block_q5_0 * dst_data = (device block_q5_0 *) ((device char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
for (int64_t i00 = tpitg.x*QK5_0; i00 < ne00; i00 += ntg.x*QK5_0) {
device const float * src = (device float *)((device char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
float amax = 0.0f; // absolute max
float max = 0.0f;
for (int j = 0; j < QK5_0; j++) {
const float v = src[j];
if (amax < fabs(v)) {
amax = fabs(v);
max = v;
}
}
const float d = max / -16;
const float id = d ? 1.0f/d : 0.0f;
dst_data[i00/QK5_0].d = d;
uint32_t qh = 0;
for (int j = 0; j < QK5_0/2; ++j) {
const float x0 = src[0 + j]*id;
const float x1 = src[QK5_0/2 + j]*id;
const uint8_t xi0 = MIN(31, (int8_t)(x0 + 16.5f));
const uint8_t xi1 = MIN(31, (int8_t)(x1 + 16.5f));
dst_data[i00/QK5_0].qs[j] = (xi0 & 0xf) | ((xi1 & 0xf) << 4);
qh |= ((xi0 & 0x10u) >> 4) << (j + 0);
qh |= ((xi1 & 0x10u) >> 4) << (j + QK5_0/2);
}
thread const uint8_t * qh8 = (thread const uint8_t *)&qh;
for (int j = 0; j < 4; ++j) {
dst_data[i00/QK5_0].qh[j] = qh8[j];
}
}
}
kernel void kernel_cpy_f32_q5_1(
device const float * src0,
device void * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
constant int64_t & ne03,
constant uint64_t & nb00,
constant uint64_t & nb01,
constant uint64_t & nb02,
constant uint64_t & nb03,
constant int64_t & ne0,
constant int64_t & ne1,
constant int64_t & ne2,
constant int64_t & ne3,
constant uint64_t & nb0,
constant uint64_t & nb1,
constant uint64_t & nb2,
constant uint64_t & nb3,
uint3 tgpig[[threadgroup_position_in_grid]],
uint3 tpitg[[thread_position_in_threadgroup]],
uint3 ntg[[threads_per_threadgroup]]) {
const int64_t i03 = tgpig[2];
const int64_t i02 = tgpig[1];
const int64_t i01 = tgpig[0];
const int64_t n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
const int64_t i3 = n / (ne2*ne1*ne0);
const int64_t i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0);
const int64_t i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0;
const int64_t i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0)/QK5_1;
device block_q5_1 * dst_data = (device block_q5_1 *) ((device char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
for (int64_t i00 = tpitg.x*QK5_1; i00 < ne00; i00 += ntg.x*QK5_1) {
device const float * src = (device float *)((device char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
float max = src[0];
float min = src[0];
for (int j = 1; j < QK5_1; j++) {
const float v = src[j];
min = v < min ? v : min;
max = v > max ? v : max;
}
const float d = (max - min) / 31;
const float id = d ? 1.0f/d : 0.0f;
dst_data[i00/QK5_1].d = d;
dst_data[i00/QK5_1].m = min;
uint32_t qh = 0;
for (int j = 0; j < QK5_1/2; ++j) {
const float x0 = (src[0 + j] - min)*id;
const float x1 = (src[QK5_1/2 + j] - min)*id;
const uint8_t xi0 = (uint8_t)(x0 + 0.5f);
const uint8_t xi1 = (uint8_t)(x1 + 0.5f);
dst_data[i00/QK5_1].qs[j] = (xi0 & 0xf) | ((xi1 & 0xf) << 4);
qh |= ((xi0 & 0x10u) >> 4) << (j + 0);
qh |= ((xi1 & 0x10u) >> 4) << (j + QK5_1/2);
}
thread const uint8_t * qh8 = (thread const uint8_t *)&qh;
for (int j = 0; j < 4; ++j) {
dst_data[i00/QK5_1].qh[j] = qh8[j];
}
}
}
static inline int best_index_int8(int n, constant float * val, float x) {
if (x <= val[0]) return 0;
if (x >= val[n-1]) return n-1;
int ml = 0, mu = n-1;
while (mu-ml > 1) {
int mav = (ml+mu)/2;
if (x < val[mav]) mu = mav; else ml = mav;
}
return x - val[mu-1] < val[mu] - x ? mu-1 : mu;
}
constexpr constant static float kvalues_iq4nl_f[16] = {
-127.f, -104.f, -83.f, -65.f, -49.f, -35.f, -22.f, -10.f, 1.f, 13.f, 25.f, 38.f, 53.f, 69.f, 89.f, 113.f
};
kernel void kernel_cpy_f32_iq4_nl(
device const float * src0,
device void * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
constant int64_t & ne03,
constant uint64_t & nb00,
constant uint64_t & nb01,
constant uint64_t & nb02,
constant uint64_t & nb03,
constant int64_t & ne0,
constant int64_t & ne1,
constant int64_t & ne2,
constant int64_t & ne3,
constant uint64_t & nb0,
constant uint64_t & nb1,
constant uint64_t & nb2,
constant uint64_t & nb3,
uint3 tgpig[[threadgroup_position_in_grid]],
uint3 tpitg[[thread_position_in_threadgroup]],
uint3 ntg[[threads_per_threadgroup]]) {
const int64_t i03 = tgpig[2];
const int64_t i02 = tgpig[1];
const int64_t i01 = tgpig[0];
const int64_t n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
const int64_t i3 = n / (ne2*ne1*ne0);
const int64_t i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0);
const int64_t i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0;
const int64_t i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0)/QK4_NL;
device block_iq4_nl * dst_data = (device block_iq4_nl *) ((device char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
for (int64_t i00 = tpitg.x*QK4_NL; i00 < ne00; i00 += ntg.x*QK4_NL) {
device const float * src = (device float *)((device char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
float amax = 0.0f; // absolute max
float max = 0.0f;
for (int j = 0; j < QK4_0; j++) {
const float v = src[j];
if (amax < fabs(v)) {
amax = fabs(v);
max = v;
}
}
const float d = max / kvalues_iq4nl_f[0];
const float id = d ? 1.0f/d : 0.0f;
float sumqx = 0, sumq2 = 0;
for (int j = 0; j < QK4_NL/2; ++j) {
const float x0 = src[0 + j]*id;
const float x1 = src[QK4_NL/2 + j]*id;
const uint8_t xi0 = best_index_int8(16, kvalues_iq4nl_f, x0);
const uint8_t xi1 = best_index_int8(16, kvalues_iq4nl_f, x1);
dst_data[i00/QK4_NL].qs[j] = xi0 | (xi1 << 4);
const float v0 = kvalues_iq4nl_f[xi0];
const float v1 = kvalues_iq4nl_f[xi1];
const float w0 = src[0 + j]*src[0 + j];
const float w1 = src[QK4_NL/2 + j]*src[QK4_NL/2 + j];
sumqx += w0*v0*src[j] + w1*v1*src[QK4_NL/2 + j];
sumq2 += w0*v0*v0 + w1*v1*v1;
}
dst_data[i00/QK4_NL].d = sumq2 > 0 ? sumqx/sumq2 : d;
}
}
kernel void kernel_concat(
device const char * src0,
device const char * src1,
device char * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
constant int64_t & ne03,
constant uint64_t & nb00,
constant uint64_t & nb01,
constant uint64_t & nb02,
constant uint64_t & nb03,
constant int64_t & ne10,
constant int64_t & ne11,
constant int64_t & ne12,
constant int64_t & ne13,
constant uint64_t & nb10,
constant uint64_t & nb11,
constant uint64_t & nb12,
constant uint64_t & nb13,
constant int64_t & ne0,
constant int64_t & ne1,
constant int64_t & ne2,
constant int64_t & ne3,
constant uint64_t & nb0,
constant uint64_t & nb1,
constant uint64_t & nb2,
constant uint64_t & nb3,
uint3 tgpig[[threadgroup_position_in_grid]],
uint3 tpitg[[thread_position_in_threadgroup]],
uint3 ntg[[threads_per_threadgroup]]) {
const int64_t i03 = tgpig.z;
const int64_t i02 = tgpig.y;
const int64_t i01 = tgpig.x;
const int64_t i13 = i03 % ne13;
const int64_t i12 = i02 % ne12;
const int64_t i11 = i01 % ne11;
device const char * src0_ptr = src0 + i03*nb03 + i02*nb02 + i01*nb01 + tpitg.x*nb00;
device const char * src1_ptr = src1 + i13*nb13 + i12*nb12 + i11*nb11 + tpitg.x*nb10;
device char * dst_ptr = dst + i03*nb3 + i02*nb2 + i01*nb1 + tpitg.x*nb0;
for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) {
if (i02 < ne02) {
((device float *)dst_ptr)[0] = ((device float *)src0_ptr)[0];
src0_ptr += ntg.x*nb00;
} else {
((device float *)dst_ptr)[0] = ((device float *)src1_ptr)[0];
src1_ptr += ntg.x*nb10;
}
dst_ptr += ntg.x*nb0;
}
}
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 13:04:25 +01:00
void kernel_mul_mv_q2_K_f32_impl(
device const void * src0,
device const float * src1,
device float * dst,
int64_t ne00,
int64_t ne01,
int64_t ne02,
int64_t ne10,
int64_t ne12,
int64_t ne0,
int64_t ne1,
uint r2,
uint r3,
threadgroup int8_t * shared_values,
uint3 tgpig,
uint tiisg,
uint sgitg) {
const int nb = ne00/QK_K;
const int r0 = tgpig.x;
const int r1 = tgpig.y;
const int im = tgpig.z;
const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST;
const int ib_row = first_row * nb;
const uint i12 = im%ne12;
const uint i13 = im/ne12;
const uint offset0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
device const block_q2_K * x = (device const block_q2_K *) src0 + ib_row + offset0;
device const float * y = (device const float *) src1 + r1*ne10 + im*ne00*ne1;
float yl[32];
float sumf[N_DST]={0.f}, all_sum;
const int step = sizeof(block_q2_K) * nb;
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
#if QK_K == 256
const int ix = tiisg/8; // 0...3
const int it = tiisg%8; // 0...7
const int iq = it/4; // 0 or 1
const int ir = it%4; // 0...3
const int is = (8*ir)/16;// 0 or 1
device const float * y4 = y + ix * QK_K + 128 * iq + 8 * ir;
for (int ib = ix; ib < nb; ib += 4) {
float4 sumy = {0.f, 0.f, 0.f, 0.f};
for (int i = 0; i < 8; ++i) {
yl[i+ 0] = y4[i+ 0]; sumy[0] += yl[i+ 0];
yl[i+ 8] = y4[i+32]; sumy[1] += yl[i+ 8];
yl[i+16] = y4[i+64]; sumy[2] += yl[i+16];
yl[i+24] = y4[i+96]; sumy[3] += yl[i+24];
}
device const uint8_t * sc = (device const uint8_t *)x[ib].scales + 8*iq + is;
device const uint16_t * qs = (device const uint16_t *)x[ib].qs + 16 * iq + 4 * ir;
device const half * dh = &x[ib].d;
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
for (int row = 0; row < N_DST; row++) {
float4 acc1 = {0.f, 0.f, 0.f, 0.f};
float4 acc2 = {0.f, 0.f, 0.f, 0.f};
for (int i = 0; i < 8; i += 2) {
acc1[0] += yl[i+ 0] * (qs[i/2] & 0x0003);
acc2[0] += yl[i+ 1] * (qs[i/2] & 0x0300);
acc1[1] += yl[i+ 8] * (qs[i/2] & 0x000c);
acc2[1] += yl[i+ 9] * (qs[i/2] & 0x0c00);
acc1[2] += yl[i+16] * (qs[i/2] & 0x0030);
acc2[2] += yl[i+17] * (qs[i/2] & 0x3000);
acc1[3] += yl[i+24] * (qs[i/2] & 0x00c0);
acc2[3] += yl[i+25] * (qs[i/2] & 0xc000);
}
float dall = dh[0];
float dmin = dh[1] * 1.f/16.f;
sumf[row] += dall * ((acc1[0] + 1.f/256.f * acc2[0]) * (sc[0] & 0xF) * 1.f/ 1.f +
(acc1[1] + 1.f/256.f * acc2[1]) * (sc[2] & 0xF) * 1.f/ 4.f +
(acc1[2] + 1.f/256.f * acc2[2]) * (sc[4] & 0xF) * 1.f/16.f +
(acc1[3] + 1.f/256.f * acc2[3]) * (sc[6] & 0xF) * 1.f/64.f) -
dmin * (sumy[0] * (sc[0] & 0xF0) + sumy[1] * (sc[2] & 0xF0) + sumy[2] * (sc[4] & 0xF0) + sumy[3] * (sc[6] & 0xF0));
qs += step/2;
sc += step;
dh += step/2;
}
y4 += 4 * QK_K;
}
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
#else
const int ix = tiisg/2; // 0...15
const int it = tiisg%2; // 0...1
device const float * y4 = y + ix * QK_K + 8 * it;
for (int ib = ix; ib < nb; ib += 16) {
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
float4 sumy = {0.f, 0.f, 0.f, 0.f};
for (int i = 0; i < 8; ++i) {
yl[i+ 0] = y4[i+ 0]; sumy[0] += yl[i+ 0];
yl[i+ 8] = y4[i+16]; sumy[1] += yl[i+ 8];
yl[i+16] = y4[i+32]; sumy[2] += yl[i+16];
yl[i+24] = y4[i+48]; sumy[3] += yl[i+24];
}
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
device const uint8_t * sc = (device const uint8_t *)x[ib].scales;
device const uint16_t * qs = (device const uint16_t *)x[ib].qs + 4 * it;
device const half * dh = &x[ib].d;
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
for (int row = 0; row < N_DST; row++) {
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
float4 acc1 = {0.f, 0.f, 0.f, 0.f};
float4 acc2 = {0.f, 0.f, 0.f, 0.f};
for (int i = 0; i < 8; i += 2) {
acc1[0] += yl[i+ 0] * (qs[i/2] & 0x0003);
acc2[0] += yl[i+ 1] * (qs[i/2] & 0x0300);
acc1[1] += yl[i+ 8] * (qs[i/2] & 0x000c);
acc2[1] += yl[i+ 9] * (qs[i/2] & 0x0c00);
acc1[2] += yl[i+16] * (qs[i/2] & 0x0030);
acc2[2] += yl[i+17] * (qs[i/2] & 0x3000);
acc1[3] += yl[i+24] * (qs[i/2] & 0x00c0);
acc2[3] += yl[i+25] * (qs[i/2] & 0xc000);
}
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
float dall = dh[0];
float dmin = dh[1];
sumf[row] += dall * ((acc1[0] + 1.f/256.f * acc2[0]) * (sc[0] & 0xF) * 1.f/ 1.f +
(acc1[1] + 1.f/256.f * acc2[1]) * (sc[1] & 0xF) * 1.f/ 4.f +
(acc1[2] + 1.f/256.f * acc2[2]) * (sc[2] & 0xF) * 1.f/16.f +
(acc1[3] + 1.f/256.f * acc2[3]) * (sc[3] & 0xF) * 1.f/64.f) -
dmin * (sumy[0] * (sc[0] >> 4) + sumy[1] * (sc[1] >> 4) + sumy[2] * (sc[2] >> 4) + sumy[3] * (sc[3] >> 4));
qs += step/2;
sc += step;
dh += step/2;
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
}
y4 += 16 * QK_K;
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
}
#endif
for (int row = 0; row < N_DST; ++row) {
all_sum = simd_sum(sumf[row]);
if (tiisg == 0) {
dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum;
}
}
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
}
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 13:04:25 +01:00
[[host_name("kernel_mul_mv_q2_K_f32")]]
kernel void kernel_mul_mv_q2_K_f32(
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
device const void * src0,
device const float * src1,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
constant uint64_t & nb00,
constant uint64_t & nb01,
constant uint64_t & nb02,
constant int64_t & ne10,
constant int64_t & ne11,
constant int64_t & ne12,
constant uint64_t & nb10,
constant uint64_t & nb11,
constant uint64_t & nb12,
constant int64_t & ne0,
constant int64_t & ne1,
constant uint & r2,
constant uint & r3,
uint3 tgpig[[threadgroup_position_in_grid]],
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 13:04:25 +01:00
uint tiisg[[thread_index_in_simdgroup]],
uint sgitg[[simdgroup_index_in_threadgroup]]) {
kernel_mul_mv_q2_K_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, nullptr, tgpig, tiisg, sgitg);
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 13:04:25 +01:00
}
#if QK_K == 256
void kernel_mul_mv_q3_K_f32_impl(
device const void * src0,
device const float * src1,
device float * dst,
int64_t ne00,
int64_t ne01,
int64_t ne02,
int64_t ne10,
int64_t ne12,
int64_t ne0,
int64_t ne1,
uint r2,
uint r3,
threadgroup int8_t * shared_values,
uint3 tgpig,
uint tiisg,
uint sgitg) {
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
const int nb = ne00/QK_K;
const int64_t r0 = tgpig.x;
const int64_t r1 = tgpig.y;
const int64_t im = tgpig.z;
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
const int first_row = (r0 * N_SIMDGROUP + sgitg) * 2;
const uint i12 = im%ne12;
const uint i13 = im/ne12;
const uint offset0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
device const block_q3_K * x = (device const block_q3_K *) src0 + first_row*nb + offset0;
device const float * yy = (device const float *) src1 + r1*ne10 + im*ne00*ne1;
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
float yl[32];
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
llama : custom attention mask + parallel decoding + no context swaps (#3228) * tests : verify that RoPE is "additive" * llama : replace ggml_diag_mask_inf with ggml_add (custom -inf mask) * ggml : ggml_rope now takes a vector with positions instead of n_past * metal : add rope_f16 kernel + optimize cpy kernels * llama : unified KV cache + batch inference API * llama : add new llama_decode() API that works with llama_batch * llama : add cell_max heuristic for more efficient kv_cache * llama : extend llama_kv_cache API * llama : more robust cell_max heuristic + wip shift * metal : disable concurrency optimization * llama : add llama_kv_cache_shift_seq + no more context swaps * llama : apply K-cache roping for Falcon and Baichuan * speculative : fix KV cache management * parallel : example for serving multiple users in parallel * parallel : disable hot-plug to avoid cache fragmentation * fixes : speculative KV cache + llama worst-case graph * llama : extend batch API to select which logits to output * llama : fix worst case graph build * ggml-cuda : update rope implementation for parallel decoding (#3254) * ggml-cuda : update rope implementation for parallel decoding * better solution for p0 computation * fix rope * simpler rope implementation --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * make : add parallel to build + fix static functions in llama.cpp * simple : fix token counting * parallel : various improvements * llama : fix cell_max logic + rename functions * parallel : try smaller batches when the KV cache is fragmented * parallel : fix sequence termination criteria * llama : silence errors KV cache errors * parallel : remove new line from prompt * parallel : process system prompt once + configurable paramters + llama API * parallel : remove question with short answers * parallel : count cache misses * parallel : print misses on each request * parallel : minor * llama : fix n_kv to never become 0 * parallel : rename hot-plug to continuous-batching * llama : improve llama_batch API + simplify parallel example * simple : add parallel decoding support * simple : improve comments + free batch * ggml-cuda : add rope f16, restore performance with parallel decoding (#3272) * ggml-cuda : add rope f16, restore performance * offload KQ_mask with all models * fix rope shift --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * llama : disable MPI for now ggml-ci * train : make KQ_pos memory buffer permanent via dummy scale op * ggml : revert change to ggml_cpy, add ggml_cont_Nd instead (#3275) ggml-ci * parallel : fix bug (extra BOS) + smaller token_prev array * parallel : fix cases where the input prompts can overflow the batch * parallel : add disabled experimental batch chunking in powers of two * llama : llama.h formatting + comments * simple : add README.md * llama : fix kv cache heuristic when context is less than 32 * parallel : fix crash when `-n -1` * llama : simplify returns if/else branches * metal : use mm kernels for batch size > 2 * examples : utilize new llama_get_logits_ith() * examples : add example for batched decoding * examples : do not eval prompt 2 times (close #3348) * server : clear the KV cache beyond n_past before llama_decode * server : avoid context swaps by shifting the KV cache --------- Co-authored-by: slaren <slarengh@gmail.com>
2023-09-28 18:04:36 +02:00
//const uint16_t kmask1 = 0x3030;
//const uint16_t kmask2 = 0x0f0f;
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
const int tid = tiisg/4;
const int ix = tiisg%4;
const int ip = tid/4; // 0 or 1
const int il = 2*((tid%4)/2); // 0 or 2
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
const int ir = tid%2;
const int n = 8;
const int l0 = n*ir;
// One would think that the Metal compiler would figure out that ip and il can only have
// 4 possible states, and optimize accordingly. Well, no. It needs help, and we do it
// with these two tales.
//
// Possible masks for the high bit
const ushort4 mm[4] = {{0x0001, 0x0100, 0x0002, 0x0200}, // ip = 0, il = 0
{0x0004, 0x0400, 0x0008, 0x0800}, // ip = 0, il = 2
{0x0010, 0x1000, 0x0020, 0x2000}, // ip = 1, il = 0
{0x0040, 0x4000, 0x0080, 0x8000}}; // ip = 1, il = 2
// Possible masks for the low 2 bits
const int4 qm[2] = {{0x0003, 0x0300, 0x000c, 0x0c00}, {0x0030, 0x3000, 0x00c0, 0xc000}};
const ushort4 hm = mm[2*ip + il/2];
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
const int shift = 2*il;
const float v1 = il == 0 ? 4.f : 64.f;
const float v2 = 4.f * v1;
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
const uint16_t s_shift1 = 4*ip;
const uint16_t s_shift2 = s_shift1 + il;
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
const int q_offset = 32*ip + l0;
const int y_offset = 128*ip + 32*il + l0;
const int step = sizeof(block_q3_K) * nb / 2;
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
device const float * y1 = yy + ix*QK_K + y_offset;
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
uint32_t scales32, aux32;
thread uint16_t * scales16 = (thread uint16_t *)&scales32;
thread const int8_t * scales = (thread const int8_t *)&scales32;
float sumf1[2] = {0.f};
float sumf2[2] = {0.f};
for (int i = ix; i < nb; i += 4) {
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
for (int l = 0; l < 8; ++l) {
yl[l+ 0] = y1[l+ 0];
yl[l+ 8] = y1[l+16];
yl[l+16] = y1[l+32];
yl[l+24] = y1[l+48];
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
}
device const uint16_t * q = (device const uint16_t *)(x[i].qs + q_offset);
device const uint16_t * h = (device const uint16_t *)(x[i].hmask + l0);
device const uint16_t * a = (device const uint16_t *)(x[i].scales);
device const half * dh = &x[i].d;
for (int row = 0; row < 2; ++row) {
const float d_all = (float)dh[0];
scales16[0] = a[4];
scales16[1] = a[5];
aux32 = ((scales32 >> s_shift2) << 4) & 0x30303030;
scales16[0] = a[il+0];
scales16[1] = a[il+1];
scales32 = ((scales32 >> s_shift1) & 0x0f0f0f0f) | aux32;
float s1 = 0, s2 = 0, s3 = 0, s4 = 0, s5 = 0, s6 = 0;
for (int l = 0; l < n; l += 2) {
const int32_t qs = q[l/2];
s1 += yl[l+0] * (qs & qm[il/2][0]);
s2 += yl[l+1] * (qs & qm[il/2][1]);
s3 += ((h[l/2] & hm[0]) ? 0.f : yl[l+0]) + ((h[l/2] & hm[1]) ? 0.f : yl[l+1]);
s4 += yl[l+16] * (qs & qm[il/2][2]);
s5 += yl[l+17] * (qs & qm[il/2][3]);
s6 += ((h[l/2] & hm[2]) ? 0.f : yl[l+16]) + ((h[l/2] & hm[3]) ? 0.f : yl[l+17]);
}
float d1 = d_all * (s1 + 1.f/256.f * s2 - s3*v1);
float d2 = d_all * (s4 + 1.f/256.f * s5 - s6*v2);
sumf1[row] += d1 * (scales[0] - 32);
sumf2[row] += d2 * (scales[2] - 32);
s1 = s2 = s3 = s4 = s5 = s6 = 0;
for (int l = 0; l < n; l += 2) {
const int32_t qs = q[l/2+8];
s1 += yl[l+8] * (qs & qm[il/2][0]);
s2 += yl[l+9] * (qs & qm[il/2][1]);
s3 += ((h[l/2+8] & hm[0]) ? 0.f : yl[l+8]) + ((h[l/2+8] & hm[1]) ? 0.f : yl[l+9]);
s4 += yl[l+24] * (qs & qm[il/2][2]);
s5 += yl[l+25] * (qs & qm[il/2][3]);
s6 += ((h[l/2+8] & hm[2]) ? 0.f : yl[l+24]) + ((h[l/2+8] & hm[3]) ? 0.f : yl[l+25]);
}
d1 = d_all * (s1 + 1.f/256.f * s2 - s3*v1);
d2 = d_all * (s4 + 1.f/256.f * s5 - s6*v2);
sumf1[row] += d1 * (scales[1] - 32);
sumf2[row] += d2 * (scales[3] - 32);
q += step;
h += step;
a += step;
dh += step;
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
}
y1 += 4 * QK_K;
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
}
for (int row = 0; row < 2; ++row) {
const float sumf = (sumf1[row] + 0.25f * sumf2[row]) / (1 << shift);
sumf1[row] = simd_sum(sumf);
}
if (tiisg == 0) {
for (int row = 0; row < 2; ++row) {
dst[r1*ne0 + im*ne0*ne1 + first_row + row] = sumf1[row];
}
}
}
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
#else
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 13:04:25 +01:00
void kernel_mul_mv_q3_K_f32_impl(
device const void * src0,
device const float * src1,
device float * dst,
constant int64_t & ne00,
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 13:04:25 +01:00
constant int64_t & ne01,
constant int64_t & ne02,
constant int64_t & ne10,
constant int64_t & ne12,
constant int64_t & ne0,
constant int64_t & ne1,
constant uint & r2,
constant uint & r3,
threadgroup int8_t * shared_values [[threadgroup(0)]],
uint3 tgpig[[threadgroup_position_in_grid]],
uint tiisg[[thread_index_in_simdgroup]],
uint sgitg[[simdgroup_index_in_threadgroup]]) {
const int nb = ne00/QK_K;
const int64_t r0 = tgpig.x;
const int64_t r1 = tgpig.y;
const int64_t im = tgpig.z;
const int row = 2 * r0 + sgitg;
const uint i12 = im%ne12;
const uint i13 = im/ne12;
const uint offset0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
device const block_q3_K * x = (device const block_q3_K *) src0 + row*nb + offset0;
device const float * yy = (device const float *) src1 + r1*ne10 + im*ne00*ne1;
const int ix = tiisg/4;
const int il = 4 * (tiisg%4);// 0, 4, 8, 12
const int iq = il/8; // 0, 0, 1, 1
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
const int in = il%8; // 0, 4, 0, 4
float2 sum = {0.f, 0.f};
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
for (int i = ix; i < nb; i += 8) {
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
const float d_all = (float)(x[i].d);
device const uint16_t * q = (device const uint16_t *)(x[i].qs + il);
device const uint16_t * h = (device const uint16_t *)(x[i].hmask + in);
device const uint16_t * s = (device const uint16_t *)(x[i].scales);
device const float * y = yy + i * QK_K + il;
const float d1 = d_all * ((int32_t)(s[0] & 0x000F) - 8);
const float d2 = d_all * ((int32_t)(s[0] & 0x00F0) - 128) * 1.f/64.f;
const float d3 = d_all * ((int32_t)(s[0] & 0x0F00) - 2048) * 1.f/4096.f;
const float d4 = d_all * ((int32_t)(s[0] & 0xF000) - 32768) * 1.f/262144.f;
for (int l = 0; l < 4; l += 2) {
const uint16_t hm = h[l/2] >> iq;
sum[0] += y[l+ 0] * d1 * ((int32_t)(q[l/2] & 0x0003) - ((hm & 0x0001) ? 0 : 4))
+ y[l+16] * d2 * ((int32_t)(q[l/2] & 0x000c) - ((hm & 0x0004) ? 0 : 16))
+ y[l+32] * d3 * ((int32_t)(q[l/2] & 0x0030) - ((hm & 0x0010) ? 0 : 64))
+ y[l+48] * d4 * ((int32_t)(q[l/2] & 0x00c0) - ((hm & 0x0040) ? 0 : 256));
sum[1] += y[l+ 1] * d1 * ((int32_t)(q[l/2] & 0x0300) - ((hm & 0x0100) ? 0 : 1024))
+ y[l+17] * d2 * ((int32_t)(q[l/2] & 0x0c00) - ((hm & 0x0400) ? 0 : 4096))
+ y[l+33] * d3 * ((int32_t)(q[l/2] & 0x3000) - ((hm & 0x1000) ? 0 : 16384))
+ y[l+49] * d4 * ((int32_t)(q[l/2] & 0xc000) - ((hm & 0x4000) ? 0 : 65536));
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
}
}
const float sumf = sum[0] + sum[1] * 1.f/256.f;
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
const float tot = simd_sum(sumf);
if (tiisg == 0) {
dst[r1*ne0 + im*ne0*ne1 + row] = tot;
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
}
}
#endif
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 13:04:25 +01:00
[[host_name("kernel_mul_mv_q3_K_f32")]]
kernel void kernel_mul_mv_q3_K_f32(
device const void * src0,
device const float * src1,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
constant uint64_t & nb00,
constant uint64_t & nb01,
constant uint64_t & nb02,
constant int64_t & ne10,
constant int64_t & ne11,
constant int64_t & ne12,
constant uint64_t & nb10,
constant uint64_t & nb11,
constant uint64_t & nb12,
constant int64_t & ne0,
constant int64_t & ne1,
constant uint & r2,
constant uint & r3,
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 13:04:25 +01:00
uint3 tgpig[[threadgroup_position_in_grid]],
uint tiisg[[thread_index_in_simdgroup]],
uint sgitg[[simdgroup_index_in_threadgroup]]) {
kernel_mul_mv_q3_K_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, nullptr, tgpig, tiisg, sgitg);
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 13:04:25 +01:00
}
#if QK_K == 256
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 13:04:25 +01:00
void kernel_mul_mv_q4_K_f32_impl(
device const void * src0,
device const float * src1,
device float * dst,
int64_t ne00,
int64_t ne01,
int64_t ne02,
int64_t ne10,
int64_t ne12,
int64_t ne0,
int64_t ne1,
uint r2,
uint r3,
threadgroup int8_t * shared_values,
uint3 tgpig,
uint tiisg,
uint sgitg) {
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
const uint16_t kmask1 = 0x3f3f;
const uint16_t kmask2 = 0x0f0f;
const uint16_t kmask3 = 0xc0c0;
const int ix = tiisg/8; // 0...3
const int it = tiisg%8; // 0...7
const int iq = it/4; // 0 or 1
const int ir = it%4; // 0...3
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
const int nb = ne00/QK_K;
const int r0 = tgpig.x;
const int r1 = tgpig.y;
const int im = tgpig.z;
//const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST;
const int first_row = r0 * N_DST;
const int ib_row = first_row * nb;
const uint i12 = im%ne12;
const uint i13 = im/ne12;
const uint offset0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
device const block_q4_K * x = (device const block_q4_K *) src0 + ib_row + offset0;
device const float * y = (device const float *) src1 + r1*ne10 + im*ne00*ne1;
float yl[16];
float yh[16];
float sumf[N_DST]={0.f}, all_sum;
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
const int step = sizeof(block_q4_K) * nb / 2;
device const float * y4 = y + ix * QK_K + 64 * iq + 8 * ir;
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
uint16_t sc16[4];
thread const uint8_t * sc8 = (thread const uint8_t *)sc16;
for (int ib = ix; ib < nb; ib += 4) {
float4 sumy = {0.f, 0.f, 0.f, 0.f};
for (int i = 0; i < 8; ++i) {
yl[i+0] = y4[i+ 0]; sumy[0] += yl[i+0];
yl[i+8] = y4[i+ 32]; sumy[1] += yl[i+8];
yh[i+0] = y4[i+128]; sumy[2] += yh[i+0];
yh[i+8] = y4[i+160]; sumy[3] += yh[i+8];
}
device const uint16_t * sc = (device const uint16_t *)x[ib].scales + iq;
device const uint16_t * q1 = (device const uint16_t *)x[ib].qs + 16 * iq + 4 * ir;
device const half * dh = &x[ib].d;
for (int row = 0; row < N_DST; row++) {
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
sc16[0] = sc[0] & kmask1;
sc16[1] = sc[2] & kmask1;
sc16[2] = ((sc[4] >> 0) & kmask2) | ((sc[0] & kmask3) >> 2);
sc16[3] = ((sc[4] >> 4) & kmask2) | ((sc[2] & kmask3) >> 2);
device const uint16_t * q2 = q1 + 32;
float4 acc1 = {0.f, 0.f, 0.f, 0.f};
float4 acc2 = {0.f, 0.f, 0.f, 0.f};
for (int i = 0; i < 8; i += 2) {
acc1[0] += yl[i+0] * (q1[i/2] & 0x000F);
acc1[1] += yl[i+1] * (q1[i/2] & 0x0F00);
acc1[2] += yl[i+8] * (q1[i/2] & 0x00F0);
acc1[3] += yl[i+9] * (q1[i/2] & 0xF000);
acc2[0] += yh[i+0] * (q2[i/2] & 0x000F);
acc2[1] += yh[i+1] * (q2[i/2] & 0x0F00);
acc2[2] += yh[i+8] * (q2[i/2] & 0x00F0);
acc2[3] += yh[i+9] * (q2[i/2] & 0xF000);
}
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
float dall = dh[0];
float dmin = dh[1];
sumf[row] += dall * ((acc1[0] + 1.f/256.f * acc1[1]) * sc8[0] +
(acc1[2] + 1.f/256.f * acc1[3]) * sc8[1] * 1.f/16.f +
(acc2[0] + 1.f/256.f * acc2[1]) * sc8[4] +
(acc2[2] + 1.f/256.f * acc2[3]) * sc8[5] * 1.f/16.f) -
dmin * (sumy[0] * sc8[2] + sumy[1] * sc8[3] + sumy[2] * sc8[6] + sumy[3] * sc8[7]);
q1 += step;
sc += step;
dh += step;
}
y4 += 4 * QK_K;
}
for (int row = 0; row < N_DST; ++row) {
all_sum = simd_sum(sumf[row]);
if (tiisg == 0) {
dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum;
}
}
}
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
#else
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 13:04:25 +01:00
void kernel_mul_mv_q4_K_f32_impl(
device const void * src0,
device const float * src1,
device float * dst,
constant int64_t & ne00,
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 13:04:25 +01:00
constant int64_t & ne01,
constant int64_t & ne02,
constant int64_t & ne10,
constant int64_t & ne12,
constant int64_t & ne0,
constant int64_t & ne1,
constant uint & r2,
constant uint & r3,
threadgroup int8_t * shared_values [[threadgroup(0)]],
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 13:04:25 +01:00
uint3 tgpig[[threadgroup_position_in_grid]],
uint tiisg[[thread_index_in_simdgroup]],
uint sgitg[[simdgroup_index_in_threadgroup]]) {
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
const int ix = tiisg/4; // 0...7
const int it = tiisg%4; // 0...3
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
const int nb = ne00/QK_K;
const int r0 = tgpig.x;
const int r1 = tgpig.y;
const int im = tgpig.z;
const int first_row = r0 * N_DST;
const int ib_row = first_row * nb;
const uint i12 = im%ne12;
const uint i13 = im/ne12;
const uint offset0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
device const block_q4_K * x = (device const block_q4_K *) src0 + ib_row + offset0;
device const float * y = (device const float *) src1 + r1*ne10 + im*ne00*ne1;
float yl[8];
float yh[8];
float sumf[N_DST]={0.f}, all_sum;
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
const int step = sizeof(block_q4_K) * nb / 2;
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
device const float * y4 = y + ix * QK_K + 8 * it;
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
uint16_t sc16[4];
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
for (int ib = ix; ib < nb; ib += 8) {
float2 sumy = {0.f, 0.f};
for (int i = 0; i < 8; ++i) {
yl[i] = y4[i+ 0]; sumy[0] += yl[i];
yh[i] = y4[i+32]; sumy[1] += yh[i];
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
}
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
device const uint16_t * sc = (device const uint16_t *)x[ib].scales;
device const uint16_t * qs = (device const uint16_t *)x[ib].qs + 4 * it;
device const half * dh = x[ib].d;
for (int row = 0; row < N_DST; row++) {
sc16[0] = sc[0] & 0x000f;
sc16[1] = sc[0] & 0x0f00;
sc16[2] = sc[0] & 0x00f0;
sc16[3] = sc[0] & 0xf000;
float2 acc1 = {0.f, 0.f};
float2 acc2 = {0.f, 0.f};
for (int i = 0; i < 8; i += 2) {
acc1[0] += yl[i+0] * (qs[i/2] & 0x000F);
acc1[1] += yl[i+1] * (qs[i/2] & 0x0F00);
acc2[0] += yh[i+0] * (qs[i/2] & 0x00F0);
acc2[1] += yh[i+1] * (qs[i/2] & 0xF000);
}
float dall = dh[0];
float dmin = dh[1];
sumf[row] += dall * ((acc1[0] + 1.f/256.f * acc1[1]) * sc16[0] +
(acc2[0] + 1.f/256.f * acc2[1]) * sc16[1] * 1.f/4096.f) -
dmin * 1.f/16.f * (sumy[0] * sc16[2] + sumy[1] * sc16[3] * 1.f/256.f);
qs += step;
sc += step;
dh += step;
}
y4 += 8 * QK_K;
}
for (int row = 0; row < N_DST; ++row) {
all_sum = simd_sum(sumf[row]);
if (tiisg == 0) {
dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum;
}
}
}
#endif
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 13:04:25 +01:00
[[host_name("kernel_mul_mv_q4_K_f32")]]
kernel void kernel_mul_mv_q4_K_f32(
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
device const void * src0,
device const float * src1,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
constant uint64_t & nb00,
constant uint64_t & nb01,
constant uint64_t & nb02,
constant int64_t & ne10,
constant int64_t & ne11,
constant int64_t & ne12,
constant uint64_t & nb10,
constant uint64_t & nb11,
constant uint64_t & nb12,
constant int64_t & ne0,
constant int64_t & ne1,
constant uint & r2,
constant uint & r3,
uint3 tgpig[[threadgroup_position_in_grid]],
uint tiisg[[thread_index_in_simdgroup]],
uint sgitg[[simdgroup_index_in_threadgroup]]) {
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
kernel_mul_mv_q4_K_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, nullptr, tgpig, tiisg, sgitg);
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 13:04:25 +01:00
}
void kernel_mul_mv_q5_K_f32_impl(
device const void * src0,
device const float * src1,
device float * dst,
int64_t ne00,
int64_t ne01,
int64_t ne02,
int64_t ne10,
int64_t ne12,
int64_t ne0,
int64_t ne1,
uint r2,
uint r3,
threadgroup int8_t * shared_values,
uint3 tgpig,
uint tiisg,
uint sgitg) {
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 13:04:25 +01:00
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
const int nb = ne00/QK_K;
const int64_t r0 = tgpig.x;
const int64_t r1 = tgpig.y;
const int im = tgpig.z;
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
const int first_row = (r0 * N_SIMDGROUP + sgitg) * 2;
const uint i12 = im%ne12;
const uint i13 = im/ne12;
const uint offset0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
device const block_q5_K * x = (device const block_q5_K *) src0 + first_row*nb + offset0;
device const float * yy = (device const float *) src1 + r1*ne10 + im*ne00*ne1;
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
float sumf[2]={0.f};
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
const int step = sizeof(block_q5_K) * nb;
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
#if QK_K == 256
#
float yl[16], yh[16];
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
const uint16_t kmask1 = 0x3f3f;
const uint16_t kmask2 = 0x0f0f;
const uint16_t kmask3 = 0xc0c0;
const int tid = tiisg/4;
const int ix = tiisg%4;
const int iq = tid/4;
const int ir = tid%4;
const int n = 8;
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
const int l0 = n*ir;
const int q_offset = 32*iq + l0;
const int y_offset = 64*iq + l0;
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
const uint8_t hm1 = 1u << (2*iq);
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
const uint8_t hm2 = hm1 << 1;
const uint8_t hm3 = hm1 << 4;
const uint8_t hm4 = hm2 << 4;
uint16_t sc16[4];
thread const uint8_t * sc8 = (thread const uint8_t *)sc16;
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
device const float * y1 = yy + ix*QK_K + y_offset;
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
for (int i = ix; i < nb; i += 4) {
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
device const uint8_t * q1 = x[i].qs + q_offset;
device const uint8_t * qh = x[i].qh + l0;
device const half * dh = &x[i].d;
device const uint16_t * a = (device const uint16_t *)x[i].scales + iq;
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
device const float * y2 = y1 + 128;
float4 sumy = {0.f, 0.f, 0.f, 0.f};
for (int l = 0; l < 8; ++l) {
yl[l+0] = y1[l+ 0]; sumy[0] += yl[l+0];
yl[l+8] = y1[l+32]; sumy[1] += yl[l+8];
yh[l+0] = y2[l+ 0]; sumy[2] += yh[l+0];
yh[l+8] = y2[l+32]; sumy[3] += yh[l+8];
}
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
for (int row = 0; row < 2; ++row) {
device const uint8_t * q2 = q1 + 64;
sc16[0] = a[0] & kmask1;
sc16[1] = a[2] & kmask1;
sc16[2] = ((a[4] >> 0) & kmask2) | ((a[0] & kmask3) >> 2);
sc16[3] = ((a[4] >> 4) & kmask2) | ((a[2] & kmask3) >> 2);
float4 acc1 = {0.f};
float4 acc2 = {0.f};
for (int l = 0; l < n; ++l) {
uint8_t h = qh[l];
acc1[0] += yl[l+0] * (q1[l] & 0x0F);
acc1[1] += yl[l+8] * (q1[l] & 0xF0);
acc1[2] += yh[l+0] * (q2[l] & 0x0F);
acc1[3] += yh[l+8] * (q2[l] & 0xF0);
acc2[0] += h & hm1 ? yl[l+0] : 0.f;
acc2[1] += h & hm2 ? yl[l+8] : 0.f;
acc2[2] += h & hm3 ? yh[l+0] : 0.f;
acc2[3] += h & hm4 ? yh[l+8] : 0.f;
}
const float dall = dh[0];
const float dmin = dh[1];
sumf[row] += dall * (sc8[0] * (acc1[0] + 16.f*acc2[0]) +
sc8[1] * (acc1[1]/16.f + 16.f*acc2[1]) +
sc8[4] * (acc1[2] + 16.f*acc2[2]) +
sc8[5] * (acc1[3]/16.f + 16.f*acc2[3])) -
dmin * (sumy[0] * sc8[2] + sumy[1] * sc8[3] + sumy[2] * sc8[6] + sumy[3] * sc8[7]);
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
q1 += step;
qh += step;
dh += step/2;
a += step/2;
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
}
y1 += 4 * QK_K;
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
}
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
#else
float yl[8], yh[8];
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
const int il = 4 * (tiisg/8); // 0, 4, 8, 12
const int ix = tiisg%8;
const int iq = il/8; // 0, 0, 1, 1
const int in = il%8; // 0, 4, 0, 4
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
device const float * y = yy + ix*QK_K + il;
for (int i = ix; i < nb; i += 8) {
for (int l = 0; l < 4; ++l) {
yl[l+0] = y[l+ 0];
yl[l+4] = y[l+16];
yh[l+0] = y[l+32];
yh[l+4] = y[l+48];
}
device const half * dh = &x[i].d;
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
device const uint8_t * q = x[i].qs + il;
device const uint8_t * h = x[i].qh + in;
device const int8_t * s = x[i].scales;
for (int row = 0; row < 2; ++row) {
const float d = dh[0];
float2 acc = {0.f, 0.f};
for (int l = 0; l < 4; ++l) {
const uint8_t hl = h[l] >> iq;
acc[0] += yl[l+0] * s[0] * ((int16_t)(q[l+ 0] & 0x0F) - (hl & 0x01 ? 0 : 16))
+ yl[l+4] * s[1] * ((int16_t)(q[l+16] & 0x0F) - (hl & 0x04 ? 0 : 16));
acc[1] += yh[l+0] * s[2] * ((int16_t)(q[l+ 0] & 0xF0) - (hl & 0x10 ? 0 : 256))
+ yh[l+4] * s[3] * ((int16_t)(q[l+16] & 0xF0) - (hl & 0x40 ? 0 : 256));
}
sumf[row] += d * (acc[0] + 1.f/16.f * acc[1]);
q += step;
h += step;
s += step;
dh += step/2;
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
}
y += 8 * QK_K;
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
}
#endif
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
for (int row = 0; row < 2; ++row) {
const float tot = simd_sum(sumf[row]);
if (tiisg == 0) {
dst[r1*ne0 + im*ne0*ne1 + first_row + row] = tot;
}
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
}
}
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 13:04:25 +01:00
[[host_name("kernel_mul_mv_q5_K_f32")]]
kernel void kernel_mul_mv_q5_K_f32(
device const void * src0,
device const float * src1,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
constant uint64_t & nb00,
constant uint64_t & nb01,
constant uint64_t & nb02,
constant int64_t & ne10,
constant int64_t & ne11,
constant int64_t & ne12,
constant uint64_t & nb10,
constant uint64_t & nb11,
constant uint64_t & nb12,
constant int64_t & ne0,
constant int64_t & ne1,
constant uint & r2,
constant uint & r3,
uint3 tgpig[[threadgroup_position_in_grid]],
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 13:04:25 +01:00
uint tiisg[[thread_index_in_simdgroup]],
uint sgitg[[simdgroup_index_in_threadgroup]]) {
kernel_mul_mv_q5_K_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, nullptr, tgpig, tiisg, sgitg);
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 13:04:25 +01:00
}
void kernel_mul_mv_q6_K_f32_impl(
device const void * src0,
device const float * src1,
device float * dst,
int64_t ne00,
int64_t ne01,
int64_t ne02,
int64_t ne10,
int64_t ne12,
int64_t ne0,
int64_t ne1,
uint r2,
uint r3,
threadgroup int8_t * shared_values,
uint3 tgpig,
uint tiisg,
uint sgitg) {
const uint8_t kmask1 = 0x03;
const uint8_t kmask2 = 0x0C;
const uint8_t kmask3 = 0x30;
const uint8_t kmask4 = 0xC0;
const int nb = ne00/QK_K;
const int64_t r0 = tgpig.x;
const int64_t r1 = tgpig.y;
const int im = tgpig.z;
const int row = 2 * r0 + sgitg;
const uint i12 = im%ne12;
const uint i13 = im/ne12;
const uint offset0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
device const block_q6_K * x = (device const block_q6_K *) src0 + row * nb + offset0;
device const float * yy = (device const float *) src1 + r1*ne10 + im*ne00*ne1;
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
float sumf = 0;
#if QK_K == 256
const int tid = tiisg/2;
const int ix = tiisg%2;
const int ip = tid/8; // 0 or 1
const int il = tid%8;
const int n = 4;
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
const int l0 = n*il;
const int is = 8*ip + l0/16;
const int y_offset = 128*ip + l0;
const int q_offset_l = 64*ip + l0;
const int q_offset_h = 32*ip + l0;
for (int i = ix; i < nb; i += 2) {
device const uint8_t * q1 = x[i].ql + q_offset_l;
device const uint8_t * q2 = q1 + 32;
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
device const uint8_t * qh = x[i].qh + q_offset_h;
device const int8_t * sc = x[i].scales + is;
Metal implementation for all k_quants (#1807) * metal : improve q4_K 28.3 -> 26.0 ms/token by avoiding a branch in the calculation of the scales. * metal : small improvement for Q4_K * metal : still optimizing Q4_K This commit pushes it down to 25.3 ms / token. The crazy idea of using 6 bits for the scales is really costly on Metal: if I remove the bit fiddling necessary to make the block scales, time goes almost to the Q4_0 23 ms/token. Before pushing the k-quants upstream I had a Q4_K variant that had used 8-bit scales. It wasn't more accurate, used 0.125 bits more per weight, was running slightly slower on the CPU (due to the larger model size and being memory bound there), and the difference was entirely negligible under CUDA. So, I decided to publish the version with 6-bit scales. Perhaps I should re-consider and change to 8-bit scales? * metal : some more optimizations Q2_K: 25.4 ms/token Q6_K: 27.3 ms/token Q4_0: 22.8 ms/token Q4_1: 23.1 ms/token * metal : Q3_K support Something is not quite right yet. * metal : Q5_K support Initial version achieves 31.2 ms/token, 210 GB/s * metal : still not able to figure out why q3_K does not work * Minor * metal : yet another failed attempt to make q3_K work * metal : optimize Q5_K 31.2 ms -> 27.8 ms. 250 GB/s. * metal : q3_K still not working Adding a heavily commented q3_K metal kernel to explain my obviously faulty logic. Perhaps someone could spot the issue? * metal : q3_K finally working Not optimized at all. What was the issue? The scales are not 4-bytes aligned, and I was accessing them with a uint32_t pointer. When I tried that on CUDA, I got an error (illegal memory access) and added a memcpy to a local array of 3 uint32_t's. But on Metal it told me there is no memcpy, so I tried accessing directly. There is no error, just garbage results. At some point I did try accessing the scales with an uint16_t pointer (the scales are for sure 2-byte aligned), but was still getting garbage. I guess, there must have been another bug. No access to scales is via a uint16_t pointer and, after starting from scratch from the C dequantize function, it finally works. * metal : Q3_K 1st optimization pass * metal : Q3_K second optimization pass - 29.6 ms/token * metal : Q3_K cleanup * metal : fixed accidentally broken Q2_K --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-12 21:39:21 +02:00
device const float * y = yy + i * QK_K + y_offset;
const float dall = x[i].d;
float4 sums = {0.f, 0.f, 0.f, 0.f};
for (int l = 0; l < n; ++l) {
sums[0] += y[l+ 0] * ((int8_t)((q1[l] & 0xF) | ((qh[l] & kmask1) << 4)) - 32);
sums[1] += y[l+32] * ((int8_t)((q2[l] & 0xF) | ((qh[l] & kmask2) << 2)) - 32);
sums[2] += y[l+64] * ((int8_t)((q1[l] >> 4) | ((qh[l] & kmask3) << 0)) - 32);
sums[3] += y[l+96] * ((int8_t)((q2[l] >> 4) | ((qh[l] & kmask4) >> 2)) - 32);
}
sumf += dall * (sums[0] * sc[0] + sums[1] * sc[2] + sums[2] * sc[4] + sums[3] * sc[6]);
}
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
#else
const int ix = tiisg/4;
const int il = 4*(tiisg%4);
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
for (int i = ix; i < nb; i += 8) {
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 18:43:07 +02:00
device const float * y = yy + i * QK_K + il;
device const uint8_t * ql = x[i].ql + il;
device const uint8_t * qh = x[i].qh + il;
device const int8_t * s = x[i].scales;
const float d = x[i].d;
float4 sums = {0.f, 0.f, 0.f, 0.f};
for (int l = 0; l < 4; ++l) {
sums[0] += y[l+ 0] * ((int8_t)((ql[l+ 0] & 0xF) | ((qh[l] & kmask1) << 4)) - 32);
sums[1] += y[l+16] * ((int8_t)((ql[l+16] & 0xF) | ((qh[l] & kmask2) << 2)) - 32);
sums[2] += y[l+32] * ((int8_t)((ql[l+ 0] >> 4) | ((qh[l] & kmask3) >> 0)) - 32);
sums[3] += y[l+48] * ((int8_t)((ql[l+16] >> 4) | ((qh[l] & kmask4) >> 2)) - 32);
}
sumf += d * (sums[0] * s[0] + sums[1] * s[1] + sums[2] * s[2] + sums[3] * s[3]);
}
#endif
const float tot = simd_sum(sumf);
if (tiisg == 0) {
dst[r1*ne0 + im*ne0*ne1 + row] = tot;
}
}
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 13:04:25 +01:00
[[host_name("kernel_mul_mv_q6_K_f32")]]
kernel void kernel_mul_mv_q6_K_f32(
device const void * src0,
device const float * src1,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
constant uint64_t & nb00,
constant uint64_t & nb01,
constant uint64_t & nb02,
constant int64_t & ne10,
constant int64_t & ne11,
constant int64_t & ne12,
constant uint64_t & nb10,
constant uint64_t & nb11,
constant uint64_t & nb12,
constant int64_t & ne0,
constant int64_t & ne1,
constant uint & r2,
constant uint & r3,
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 13:04:25 +01:00
uint3 tgpig[[threadgroup_position_in_grid]],
uint tiisg[[thread_index_in_simdgroup]],
uint sgitg[[simdgroup_index_in_threadgroup]]) {
kernel_mul_mv_q6_K_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, nullptr, tgpig, tiisg, sgitg);
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 13:04:25 +01:00
}
2024-01-08 16:02:32 +01:00
// ======================= "True" 2-bit
void kernel_mul_mv_iq2_xxs_f32_impl(
device const void * src0,
device const float * src1,
device float * dst,
int64_t ne00,
int64_t ne01,
int64_t ne02,
int64_t ne10,
int64_t ne12,
int64_t ne0,
int64_t ne1,
uint r2,
uint r3,
threadgroup int8_t * shared_values,
uint3 tgpig,
uint tiisg,
uint sgitg) {
2024-01-08 16:02:32 +01:00
const int nb = ne00/QK_K;
const int r0 = tgpig.x;
const int r1 = tgpig.y;
const int im = tgpig.z;
const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST;
const int ib_row = first_row * nb;
const uint i12 = im%ne12;
const uint i13 = im/ne12;
const uint offset0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
device const block_iq2_xxs * x = (device const block_iq2_xxs *) src0 + ib_row + offset0;
device const float * y = (device const float *) src1 + r1*ne10 + im*ne00*ne1;
float yl[32];
float sumf[N_DST]={0.f}, all_sum;
const int nb32 = nb * (QK_K / 32);
threadgroup uint64_t * values = (threadgroup uint64_t *)shared_values;
threadgroup uint8_t * shared_signs = (threadgroup uint8_t *)(values + 256);
{
int nval = 4;
int pos = (32*sgitg + tiisg)*nval;
for (int i = 0; i < nval; ++i) values[pos + i] = iq2xxs_grid[pos + i];
2024-01-08 16:02:32 +01:00
nval = 2;
pos = (32*sgitg + tiisg)*nval;
for (int i = 0; i < nval; ++i) shared_signs[pos+i] = ksigns_iq2xs[pos+i];
threadgroup_barrier(mem_flags::mem_threadgroup);
}
const int ix = tiisg;
device const float * y4 = y + 32 * ix;
for (int ib32 = ix; ib32 < nb32; ib32 += 32) {
for (int i = 0; i < 32; ++i) {
yl[i] = y4[i];
}
const int ibl = ib32 / (QK_K / 32);
const int ib = ib32 % (QK_K / 32);
device const block_iq2_xxs * xr = x + ibl;
device const uint16_t * q2 = xr->qs + 4 * ib;
device const half * dh = &xr->d;
for (int row = 0; row < N_DST; row++) {
const float db = dh[0];
device const uint8_t * aux8 = (device const uint8_t *)q2;
const uint32_t aux32 = q2[2] | (q2[3] << 16);
const float d = db * (0.5f + (aux32 >> 28));
float sum = 0;
for (int l = 0; l < 4; ++l) {
const threadgroup uint8_t * grid = (const threadgroup uint8_t *)(values + aux8[l]);
const uint8_t signs = shared_signs[(aux32 >> 7*l) & 127];
for (int j = 0; j < 8; ++j) {
sum += yl[8*l + j] * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
}
}
sumf[row] += d * sum;
dh += nb*sizeof(block_iq2_xxs)/2;
q2 += nb*sizeof(block_iq2_xxs)/2;
}
y4 += 32 * 32;
}
for (int row = 0; row < N_DST; ++row) {
all_sum = simd_sum(sumf[row]);
if (tiisg == 0) {
dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum * 0.25f;
}
}
}
[[host_name("kernel_mul_mv_iq2_xxs_f32")]]
kernel void kernel_mul_mv_iq2_xxs_f32(
device const void * src0,
device const float * src1,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
constant uint64_t & nb00,
constant uint64_t & nb01,
constant uint64_t & nb02,
constant int64_t & ne10,
constant int64_t & ne11,
constant int64_t & ne12,
constant uint64_t & nb10,
constant uint64_t & nb11,
constant uint64_t & nb12,
constant int64_t & ne0,
constant int64_t & ne1,
constant uint & r2,
constant uint & r3,
threadgroup int8_t * shared_values [[threadgroup(0)]],
uint3 tgpig[[threadgroup_position_in_grid]],
uint tiisg[[thread_index_in_simdgroup]],
uint sgitg[[simdgroup_index_in_threadgroup]]) {
kernel_mul_mv_iq2_xxs_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, shared_values, tgpig, tiisg, sgitg);
}
void kernel_mul_mv_iq2_xs_f32_impl(
device const void * src0,
device const float * src1,
device float * dst,
int64_t ne00,
int64_t ne01,
int64_t ne02,
int64_t ne10,
int64_t ne12,
int64_t ne0,
int64_t ne1,
uint r2,
uint r3,
threadgroup int8_t * shared_values,
uint3 tgpig,
uint tiisg,
uint sgitg) {
const int nb = ne00/QK_K;
const int r0 = tgpig.x;
const int r1 = tgpig.y;
const int im = tgpig.z;
const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST;
const int ib_row = first_row * nb;
const uint i12 = im%ne12;
const uint i13 = im/ne12;
const uint offset0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
device const block_iq2_xs * x = (device const block_iq2_xs *) src0 + ib_row + offset0;
device const float * y = (device const float *) src1 + r1*ne10 + im*ne00*ne1;
float yl[32];
float sumf[N_DST]={0.f}, all_sum;
const int nb32 = nb * (QK_K / 32);
threadgroup uint64_t * values = (threadgroup uint64_t *)shared_values;
threadgroup uint8_t * shared_signs = (threadgroup uint8_t *)(values + 512);
{
int nval = 8;
int pos = (32*sgitg + tiisg)*nval;
for (int i = 0; i < nval; ++i) values[pos + i] = iq2xs_grid[pos + i];
nval = 2;
pos = (32*sgitg + tiisg)*nval;
for (int i = 0; i < nval; ++i) shared_signs[pos+i] = ksigns_iq2xs[pos+i];
threadgroup_barrier(mem_flags::mem_threadgroup);
}
const int ix = tiisg;
device const float * y4 = y + 32 * ix;
for (int ib32 = ix; ib32 < nb32; ib32 += 32) {
for (int i = 0; i < 32; ++i) {
yl[i] = y4[i];
}
const int ibl = ib32 / (QK_K / 32);
const int ib = ib32 % (QK_K / 32);
device const block_iq2_xs * xr = x + ibl;
device const uint16_t * q2 = xr->qs + 4 * ib;
device const uint8_t * sc = xr->scales + ib;
device const half * dh = &xr->d;
for (int row = 0; row < N_DST; row++) {
const float db = dh[0];
const uint8_t ls1 = sc[0] & 0xf;
const uint8_t ls2 = sc[0] >> 4;
const float d1 = db * (0.5f + ls1);
const float d2 = db * (0.5f + ls2);
float sum1 = 0, sum2 = 0;
for (int l = 0; l < 2; ++l) {
const threadgroup uint8_t * grid = (const threadgroup uint8_t *)(values + (q2[l] & 511));
const uint8_t signs = shared_signs[(q2[l] >> 9)];
for (int j = 0; j < 8; ++j) {
sum1 += yl[8*l + j] * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
}
}
for (int l = 2; l < 4; ++l) {
const threadgroup uint8_t * grid = (const threadgroup uint8_t *)(values + (q2[l] & 511));
const uint8_t signs = shared_signs[(q2[l] >> 9)];
for (int j = 0; j < 8; ++j) {
sum2 += yl[8*l + j] * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
}
}
sumf[row] += d1 * sum1 + d2 * sum2;
dh += nb*sizeof(block_iq2_xs)/2;
q2 += nb*sizeof(block_iq2_xs)/2;
sc += nb*sizeof(block_iq2_xs);
}
y4 += 32 * 32;
}
for (int row = 0; row < N_DST; ++row) {
all_sum = simd_sum(sumf[row]);
if (tiisg == 0) {
dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum * 0.25f;
}
}
}
[[host_name("kernel_mul_mv_iq2_xs_f32")]]
kernel void kernel_mul_mv_iq2_xs_f32(
device const void * src0,
device const float * src1,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
constant uint64_t & nb00,
constant uint64_t & nb01,
constant uint64_t & nb02,
constant int64_t & ne10,
constant int64_t & ne11,
constant int64_t & ne12,
constant uint64_t & nb10,
constant uint64_t & nb11,
constant uint64_t & nb12,
constant int64_t & ne0,
constant int64_t & ne1,
constant uint & r2,
constant uint & r3,
threadgroup int8_t * shared_values [[threadgroup(0)]],
uint3 tgpig[[threadgroup_position_in_grid]],
uint tiisg[[thread_index_in_simdgroup]],
uint sgitg[[simdgroup_index_in_threadgroup]]) {
kernel_mul_mv_iq2_xs_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, shared_values, tgpig, tiisg, sgitg);
}
void kernel_mul_mv_iq3_xxs_f32_impl(
device const void * src0,
device const float * src1,
device float * dst,
int64_t ne00,
int64_t ne01,
int64_t ne02,
int64_t ne10,
int64_t ne12,
int64_t ne0,
int64_t ne1,
uint r2,
uint r3,
threadgroup int8_t * shared_values,
uint3 tgpig,
uint tiisg,
uint sgitg) {
const int nb = ne00/QK_K;
const int r0 = tgpig.x;
const int r1 = tgpig.y;
const int im = tgpig.z;
const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST;
const int ib_row = first_row * nb;
const uint i12 = im%ne12;
const uint i13 = im/ne12;
const uint offset0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
device const block_iq3_xxs * x = (device const block_iq3_xxs *) src0 + ib_row + offset0;
device const float * y = (device const float *) src1 + r1*ne10 + im*ne00*ne1;
float yl[32];
float sumf[N_DST]={0.f}, all_sum;
const int nb32 = nb * (QK_K / 32);
threadgroup uint32_t * values = (threadgroup uint32_t *)shared_values;
threadgroup uint8_t * shared_signs = (threadgroup uint8_t *)(values + 256);
{
int nval = 4;
int pos = (32*sgitg + tiisg)*nval;
for (int i = 0; i < nval; ++i) values[pos + i] = iq3xxs_grid[pos + i];
nval = 2;
pos = (32*sgitg + tiisg)*nval;
for (int i = 0; i < nval; ++i) shared_signs[pos+i] = ksigns_iq2xs[pos+i];
threadgroup_barrier(mem_flags::mem_threadgroup);
}
const int ix = tiisg;
device const float * y4 = y + 32 * ix;
for (int ib32 = ix; ib32 < nb32; ib32 += 32) {
for (int i = 0; i < 32; ++i) {
yl[i] = y4[i];
}
const int ibl = ib32 / (QK_K / 32);
const int ib = ib32 % (QK_K / 32);
device const block_iq3_xxs * xr = x + ibl;
device const uint8_t * q3 = xr->qs + 8 * ib;
device const uint16_t * gas = (device const uint16_t *)(xr->qs + QK_K/4) + 2 * ib;
device const half * dh = &xr->d;
for (int row = 0; row < N_DST; row++) {
const float db = dh[0];
const uint32_t aux32 = gas[0] | (gas[1] << 16);
const float d = db * (0.5f + (aux32 >> 28));
float2 sum = {0};
for (int l = 0; l < 4; ++l) {
const threadgroup uint8_t * grid1 = (const threadgroup uint8_t *)(values + q3[2*l+0]);
const threadgroup uint8_t * grid2 = (const threadgroup uint8_t *)(values + q3[2*l+1]);
const uint8_t signs = shared_signs[(aux32 >> 7*l) & 127];
for (int j = 0; j < 4; ++j) {
sum[0] += yl[8*l + j + 0] * grid1[j] * (signs & kmask_iq2xs[j+0] ? -1.f : 1.f);
sum[1] += yl[8*l + j + 4] * grid2[j] * (signs & kmask_iq2xs[j+4] ? -1.f : 1.f);
}
}
sumf[row] += d * (sum[0] + sum[1]);
dh += nb*sizeof(block_iq3_xxs)/2;
q3 += nb*sizeof(block_iq3_xxs);
gas += nb*sizeof(block_iq3_xxs)/2;
}
y4 += 32 * 32;
}
for (int row = 0; row < N_DST; ++row) {
all_sum = simd_sum(sumf[row]);
if (tiisg == 0) {
dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum * 0.5f;
}
}
}
[[host_name("kernel_mul_mv_iq3_xxs_f32")]]
kernel void kernel_mul_mv_iq3_xxs_f32(
device const void * src0,
device const float * src1,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
constant uint64_t & nb00,
constant uint64_t & nb01,
constant uint64_t & nb02,
constant int64_t & ne10,
constant int64_t & ne11,
constant int64_t & ne12,
constant uint64_t & nb10,
constant uint64_t & nb11,
constant uint64_t & nb12,
constant int64_t & ne0,
constant int64_t & ne1,
constant uint & r2,
constant uint & r3,
threadgroup int8_t * shared_values [[threadgroup(0)]],
uint3 tgpig[[threadgroup_position_in_grid]],
uint tiisg[[thread_index_in_simdgroup]],
uint sgitg[[simdgroup_index_in_threadgroup]]) {
kernel_mul_mv_iq3_xxs_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, shared_values, tgpig, tiisg, sgitg);
}
IQ3_S: a much better alternative to Q3_K (#5676) * iq4_nl: squash commits for easier rebase * Basics (quantize, dequantize) * CUDA dequantize and dot product * Slightly faster CUDA dot product (120 t/s) * Switch to 6-bit scales * Scalar dot product * AVX2 dot product * ARM_NEON dot product * Works on metal, but still slow * Slightly better Metal dot product * Another small Metal improvement * Metal dot product is getting there * Faster CUDA dot product * Add 1/8 ffn_down layers as Q5_K when no imatrix has been provided * Report the actual bpw * Add _xs mix that is 4.05 bpw for non-MoE models * Remove IQ4_XS for now, slightly adjust kvalues_iq4nl * AVX2 dot product uses Q8_0 instead of Q8_K * Add to test-backend-ops * Minor fix * Also use use Q5_K for attn_output in MoE models * Fixes after merging latest master * Switching to blocks of 32 * AVX2 for blocks of 32 * Scaler dot product for blocks of 32 * ARM_NEON dot product for blocks of 32 * Metal kernels for blocks of 32 * Slightly faster Metal kernels * Resurrecting iq3_xs After all the experimentation, nothing was better than this. * Minor PPL improvement via a block scale fudge factor * Minor improvement via 3 neighbours * iq3_xs: working scalar and AVX2 dot products * iq3_xs: ARM_NEON dot product - works but extremely slow (10 t/s) * iq3_xs: working Metal implementation * Adding IQ3_M - IQ3_XS mix with mostly Q4_K * iiq3_xs: a 3.4375 bpw variant * iq3_xs: make CUDA work for new version * iq3_xs: make scalar and AVX2 work for new version * iq3_s: make ARM_NEON work with new version * iq3_xs: make new version work on metal Performance is very similar to Q3_K_S * iq3_xs: tiny Metal speed improvement * iq3_xs: tiny Metal speed improvement * Fix stupid warning * Q3_K_XS now uses a mix of IQ3_XS and IQ3_XXS * iq3_xs: rename to iq3_s * iq3_s: make tests pass * Move Q3_K_XS mix to 3.25 bpw * Attempt to fix failing tests * Another attempt to fix the Windows builds * Attempt to fix ROCm * ROCm again * iq3_s: partial fix for QK_K = 64 * iq3_s: make it work on metal for QK_K = 64 Pleasent surprise: the coding was super-block size independent, so all it took was to delete some QK_K == 256 guards. * Will this fix ROCm? --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2024-02-24 15:23:52 +01:00
void kernel_mul_mv_iq3_s_f32_impl(
device const void * src0,
device const float * src1,
device float * dst,
int64_t ne00,
int64_t ne01,
int64_t ne02,
int64_t ne10,
int64_t ne12,
int64_t ne0,
int64_t ne1,
uint r2,
uint r3,
threadgroup int8_t * shared_values,
uint3 tgpig,
uint tiisg,
uint sgitg) {
IQ3_S: a much better alternative to Q3_K (#5676) * iq4_nl: squash commits for easier rebase * Basics (quantize, dequantize) * CUDA dequantize and dot product * Slightly faster CUDA dot product (120 t/s) * Switch to 6-bit scales * Scalar dot product * AVX2 dot product * ARM_NEON dot product * Works on metal, but still slow * Slightly better Metal dot product * Another small Metal improvement * Metal dot product is getting there * Faster CUDA dot product * Add 1/8 ffn_down layers as Q5_K when no imatrix has been provided * Report the actual bpw * Add _xs mix that is 4.05 bpw for non-MoE models * Remove IQ4_XS for now, slightly adjust kvalues_iq4nl * AVX2 dot product uses Q8_0 instead of Q8_K * Add to test-backend-ops * Minor fix * Also use use Q5_K for attn_output in MoE models * Fixes after merging latest master * Switching to blocks of 32 * AVX2 for blocks of 32 * Scaler dot product for blocks of 32 * ARM_NEON dot product for blocks of 32 * Metal kernels for blocks of 32 * Slightly faster Metal kernels * Resurrecting iq3_xs After all the experimentation, nothing was better than this. * Minor PPL improvement via a block scale fudge factor * Minor improvement via 3 neighbours * iq3_xs: working scalar and AVX2 dot products * iq3_xs: ARM_NEON dot product - works but extremely slow (10 t/s) * iq3_xs: working Metal implementation * Adding IQ3_M - IQ3_XS mix with mostly Q4_K * iiq3_xs: a 3.4375 bpw variant * iq3_xs: make CUDA work for new version * iq3_xs: make scalar and AVX2 work for new version * iq3_s: make ARM_NEON work with new version * iq3_xs: make new version work on metal Performance is very similar to Q3_K_S * iq3_xs: tiny Metal speed improvement * iq3_xs: tiny Metal speed improvement * Fix stupid warning * Q3_K_XS now uses a mix of IQ3_XS and IQ3_XXS * iq3_xs: rename to iq3_s * iq3_s: make tests pass * Move Q3_K_XS mix to 3.25 bpw * Attempt to fix failing tests * Another attempt to fix the Windows builds * Attempt to fix ROCm * ROCm again * iq3_s: partial fix for QK_K = 64 * iq3_s: make it work on metal for QK_K = 64 Pleasent surprise: the coding was super-block size independent, so all it took was to delete some QK_K == 256 guards. * Will this fix ROCm? --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2024-02-24 15:23:52 +01:00
const int nb = ne00/QK_K;
const int r0 = tgpig.x;
const int r1 = tgpig.y;
const int im = tgpig.z;
const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST;
const int ib_row = first_row * nb;
const uint i12 = im%ne12;
const uint i13 = im/ne12;
const uint offset0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
device const block_iq3_s * x = (device const block_iq3_s *) src0 + ib_row + offset0;
device const float * y = (device const float *) src1 + r1*ne10 + im*ne00*ne1;
float yl[32];
float sumf[N_DST]={0.f}, all_sum;
const int nb32 = nb * (QK_K / 32);
threadgroup uint32_t * values = (threadgroup uint32_t *)shared_values;
{
int nval = 8;
int pos = (32*sgitg + tiisg)*nval;
for (int i = 0; i < nval; ++i) values[pos + i] = iq3s_grid[pos + i];
IQ3_S: a much better alternative to Q3_K (#5676) * iq4_nl: squash commits for easier rebase * Basics (quantize, dequantize) * CUDA dequantize and dot product * Slightly faster CUDA dot product (120 t/s) * Switch to 6-bit scales * Scalar dot product * AVX2 dot product * ARM_NEON dot product * Works on metal, but still slow * Slightly better Metal dot product * Another small Metal improvement * Metal dot product is getting there * Faster CUDA dot product * Add 1/8 ffn_down layers as Q5_K when no imatrix has been provided * Report the actual bpw * Add _xs mix that is 4.05 bpw for non-MoE models * Remove IQ4_XS for now, slightly adjust kvalues_iq4nl * AVX2 dot product uses Q8_0 instead of Q8_K * Add to test-backend-ops * Minor fix * Also use use Q5_K for attn_output in MoE models * Fixes after merging latest master * Switching to blocks of 32 * AVX2 for blocks of 32 * Scaler dot product for blocks of 32 * ARM_NEON dot product for blocks of 32 * Metal kernels for blocks of 32 * Slightly faster Metal kernels * Resurrecting iq3_xs After all the experimentation, nothing was better than this. * Minor PPL improvement via a block scale fudge factor * Minor improvement via 3 neighbours * iq3_xs: working scalar and AVX2 dot products * iq3_xs: ARM_NEON dot product - works but extremely slow (10 t/s) * iq3_xs: working Metal implementation * Adding IQ3_M - IQ3_XS mix with mostly Q4_K * iiq3_xs: a 3.4375 bpw variant * iq3_xs: make CUDA work for new version * iq3_xs: make scalar and AVX2 work for new version * iq3_s: make ARM_NEON work with new version * iq3_xs: make new version work on metal Performance is very similar to Q3_K_S * iq3_xs: tiny Metal speed improvement * iq3_xs: tiny Metal speed improvement * Fix stupid warning * Q3_K_XS now uses a mix of IQ3_XS and IQ3_XXS * iq3_xs: rename to iq3_s * iq3_s: make tests pass * Move Q3_K_XS mix to 3.25 bpw * Attempt to fix failing tests * Another attempt to fix the Windows builds * Attempt to fix ROCm * ROCm again * iq3_s: partial fix for QK_K = 64 * iq3_s: make it work on metal for QK_K = 64 Pleasent surprise: the coding was super-block size independent, so all it took was to delete some QK_K == 256 guards. * Will this fix ROCm? --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2024-02-24 15:23:52 +01:00
threadgroup_barrier(mem_flags::mem_threadgroup);
}
const int ix = tiisg;
device const float * y4 = y + 32 * ix;
for (int ib32 = ix; ib32 < nb32; ib32 += 32) {
for (int i = 0; i < 32; ++i) {
yl[i] = y4[i];
}
const int ibl = ib32 / (QK_K / 32);
const int ib = ib32 % (QK_K / 32);
device const block_iq3_s * xr = x + ibl;
device const uint8_t * qs = xr->qs + 8 * ib;
device const uint8_t * qh = xr->qh + ib;
device const uint8_t * sc = xr->scales + (ib/2);
device const uint8_t * signs = xr->signs + 4 * ib;
device const half * dh = &xr->d;
for (int row = 0; row < N_DST; row++) {
const float db = dh[0];
const float d = db * (1 + 2*((sc[0] >> 4*(ib%2)) & 0xf));
IQ3_S: a much better alternative to Q3_K (#5676) * iq4_nl: squash commits for easier rebase * Basics (quantize, dequantize) * CUDA dequantize and dot product * Slightly faster CUDA dot product (120 t/s) * Switch to 6-bit scales * Scalar dot product * AVX2 dot product * ARM_NEON dot product * Works on metal, but still slow * Slightly better Metal dot product * Another small Metal improvement * Metal dot product is getting there * Faster CUDA dot product * Add 1/8 ffn_down layers as Q5_K when no imatrix has been provided * Report the actual bpw * Add _xs mix that is 4.05 bpw for non-MoE models * Remove IQ4_XS for now, slightly adjust kvalues_iq4nl * AVX2 dot product uses Q8_0 instead of Q8_K * Add to test-backend-ops * Minor fix * Also use use Q5_K for attn_output in MoE models * Fixes after merging latest master * Switching to blocks of 32 * AVX2 for blocks of 32 * Scaler dot product for blocks of 32 * ARM_NEON dot product for blocks of 32 * Metal kernels for blocks of 32 * Slightly faster Metal kernels * Resurrecting iq3_xs After all the experimentation, nothing was better than this. * Minor PPL improvement via a block scale fudge factor * Minor improvement via 3 neighbours * iq3_xs: working scalar and AVX2 dot products * iq3_xs: ARM_NEON dot product - works but extremely slow (10 t/s) * iq3_xs: working Metal implementation * Adding IQ3_M - IQ3_XS mix with mostly Q4_K * iiq3_xs: a 3.4375 bpw variant * iq3_xs: make CUDA work for new version * iq3_xs: make scalar and AVX2 work for new version * iq3_s: make ARM_NEON work with new version * iq3_xs: make new version work on metal Performance is very similar to Q3_K_S * iq3_xs: tiny Metal speed improvement * iq3_xs: tiny Metal speed improvement * Fix stupid warning * Q3_K_XS now uses a mix of IQ3_XS and IQ3_XXS * iq3_xs: rename to iq3_s * iq3_s: make tests pass * Move Q3_K_XS mix to 3.25 bpw * Attempt to fix failing tests * Another attempt to fix the Windows builds * Attempt to fix ROCm * ROCm again * iq3_s: partial fix for QK_K = 64 * iq3_s: make it work on metal for QK_K = 64 Pleasent surprise: the coding was super-block size independent, so all it took was to delete some QK_K == 256 guards. * Will this fix ROCm? --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2024-02-24 15:23:52 +01:00
float2 sum = {0};
for (int l = 0; l < 4; ++l) {
const threadgroup uint32_t * table1 = qh[0] & kmask_iq2xs[2*l+0] ? values + 256 : values;
const threadgroup uint32_t * table2 = qh[0] & kmask_iq2xs[2*l+1] ? values + 256 : values;
const threadgroup uint8_t * grid1 = (const threadgroup uint8_t *)(table1 + qs[2*l+0]);
const threadgroup uint8_t * grid2 = (const threadgroup uint8_t *)(table2 + qs[2*l+1]);
IQ3_S: a much better alternative to Q3_K (#5676) * iq4_nl: squash commits for easier rebase * Basics (quantize, dequantize) * CUDA dequantize and dot product * Slightly faster CUDA dot product (120 t/s) * Switch to 6-bit scales * Scalar dot product * AVX2 dot product * ARM_NEON dot product * Works on metal, but still slow * Slightly better Metal dot product * Another small Metal improvement * Metal dot product is getting there * Faster CUDA dot product * Add 1/8 ffn_down layers as Q5_K when no imatrix has been provided * Report the actual bpw * Add _xs mix that is 4.05 bpw for non-MoE models * Remove IQ4_XS for now, slightly adjust kvalues_iq4nl * AVX2 dot product uses Q8_0 instead of Q8_K * Add to test-backend-ops * Minor fix * Also use use Q5_K for attn_output in MoE models * Fixes after merging latest master * Switching to blocks of 32 * AVX2 for blocks of 32 * Scaler dot product for blocks of 32 * ARM_NEON dot product for blocks of 32 * Metal kernels for blocks of 32 * Slightly faster Metal kernels * Resurrecting iq3_xs After all the experimentation, nothing was better than this. * Minor PPL improvement via a block scale fudge factor * Minor improvement via 3 neighbours * iq3_xs: working scalar and AVX2 dot products * iq3_xs: ARM_NEON dot product - works but extremely slow (10 t/s) * iq3_xs: working Metal implementation * Adding IQ3_M - IQ3_XS mix with mostly Q4_K * iiq3_xs: a 3.4375 bpw variant * iq3_xs: make CUDA work for new version * iq3_xs: make scalar and AVX2 work for new version * iq3_s: make ARM_NEON work with new version * iq3_xs: make new version work on metal Performance is very similar to Q3_K_S * iq3_xs: tiny Metal speed improvement * iq3_xs: tiny Metal speed improvement * Fix stupid warning * Q3_K_XS now uses a mix of IQ3_XS and IQ3_XXS * iq3_xs: rename to iq3_s * iq3_s: make tests pass * Move Q3_K_XS mix to 3.25 bpw * Attempt to fix failing tests * Another attempt to fix the Windows builds * Attempt to fix ROCm * ROCm again * iq3_s: partial fix for QK_K = 64 * iq3_s: make it work on metal for QK_K = 64 Pleasent surprise: the coding was super-block size independent, so all it took was to delete some QK_K == 256 guards. * Will this fix ROCm? --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2024-02-24 15:23:52 +01:00
for (int j = 0; j < 4; ++j) {
sum[0] += yl[8*l + j + 0] * grid1[j] * select(1, -1, signs[l] & kmask_iq2xs[j+0]);
sum[1] += yl[8*l + j + 4] * grid2[j] * select(1, -1, signs[l] & kmask_iq2xs[j+4]);
}
}
sumf[row] += d * (sum[0] + sum[1]);
dh += nb*sizeof(block_iq3_s)/2;
qs += nb*sizeof(block_iq3_s);
qh += nb*sizeof(block_iq3_s);
sc += nb*sizeof(block_iq3_s);
signs += nb*sizeof(block_iq3_s);
}
y4 += 32 * 32;
}
for (int row = 0; row < N_DST; ++row) {
all_sum = simd_sum(sumf[row]);
if (tiisg == 0) {
dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum;
IQ3_S: a much better alternative to Q3_K (#5676) * iq4_nl: squash commits for easier rebase * Basics (quantize, dequantize) * CUDA dequantize and dot product * Slightly faster CUDA dot product (120 t/s) * Switch to 6-bit scales * Scalar dot product * AVX2 dot product * ARM_NEON dot product * Works on metal, but still slow * Slightly better Metal dot product * Another small Metal improvement * Metal dot product is getting there * Faster CUDA dot product * Add 1/8 ffn_down layers as Q5_K when no imatrix has been provided * Report the actual bpw * Add _xs mix that is 4.05 bpw for non-MoE models * Remove IQ4_XS for now, slightly adjust kvalues_iq4nl * AVX2 dot product uses Q8_0 instead of Q8_K * Add to test-backend-ops * Minor fix * Also use use Q5_K for attn_output in MoE models * Fixes after merging latest master * Switching to blocks of 32 * AVX2 for blocks of 32 * Scaler dot product for blocks of 32 * ARM_NEON dot product for blocks of 32 * Metal kernels for blocks of 32 * Slightly faster Metal kernels * Resurrecting iq3_xs After all the experimentation, nothing was better than this. * Minor PPL improvement via a block scale fudge factor * Minor improvement via 3 neighbours * iq3_xs: working scalar and AVX2 dot products * iq3_xs: ARM_NEON dot product - works but extremely slow (10 t/s) * iq3_xs: working Metal implementation * Adding IQ3_M - IQ3_XS mix with mostly Q4_K * iiq3_xs: a 3.4375 bpw variant * iq3_xs: make CUDA work for new version * iq3_xs: make scalar and AVX2 work for new version * iq3_s: make ARM_NEON work with new version * iq3_xs: make new version work on metal Performance is very similar to Q3_K_S * iq3_xs: tiny Metal speed improvement * iq3_xs: tiny Metal speed improvement * Fix stupid warning * Q3_K_XS now uses a mix of IQ3_XS and IQ3_XXS * iq3_xs: rename to iq3_s * iq3_s: make tests pass * Move Q3_K_XS mix to 3.25 bpw * Attempt to fix failing tests * Another attempt to fix the Windows builds * Attempt to fix ROCm * ROCm again * iq3_s: partial fix for QK_K = 64 * iq3_s: make it work on metal for QK_K = 64 Pleasent surprise: the coding was super-block size independent, so all it took was to delete some QK_K == 256 guards. * Will this fix ROCm? --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2024-02-24 15:23:52 +01:00
}
}
}
[[host_name("kernel_mul_mv_iq3_s_f32")]]
kernel void kernel_mul_mv_iq3_s_f32(
device const void * src0,
device const float * src1,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
constant uint64_t & nb00,
constant uint64_t & nb01,
constant uint64_t & nb02,
constant int64_t & ne10,
constant int64_t & ne11,
constant int64_t & ne12,
constant uint64_t & nb10,
constant uint64_t & nb11,
constant uint64_t & nb12,
constant int64_t & ne0,
constant int64_t & ne1,
constant uint & r2,
constant uint & r3,
threadgroup int8_t * shared_values [[threadgroup(0)]],
uint3 tgpig[[threadgroup_position_in_grid]],
uint tiisg[[thread_index_in_simdgroup]],
uint sgitg[[simdgroup_index_in_threadgroup]]) {
kernel_mul_mv_iq3_s_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, shared_values, tgpig, tiisg, sgitg);
}
void kernel_mul_mv_iq2_s_f32_impl(
device const void * src0,
device const float * src1,
device float * dst,
int64_t ne00,
int64_t ne01,
int64_t ne02,
int64_t ne10,
int64_t ne12,
int64_t ne0,
int64_t ne1,
uint r2,
uint r3,
threadgroup int8_t * shared_values,
uint3 tgpig,
uint tiisg,
uint sgitg) {
const int nb = ne00/QK_K;
const int r0 = tgpig.x;
const int r1 = tgpig.y;
const int im = tgpig.z;
const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST;
const int ib_row = first_row * nb;
const uint i12 = im%ne12;
const uint i13 = im/ne12;
const uint offset0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
device const block_iq2_s * x = (device const block_iq2_s *) src0 + ib_row + offset0;
device const float * y = (device const float *) src1 + r1*ne10 + im*ne00*ne1;
float yl[32];
float sumf[N_DST]={0.f}, all_sum;
const int nb32 = nb * (QK_K / 32);
//threadgroup uint64_t * values = (threadgroup uint64_t *)shared_values;
//{
// int nval = 32;
// int pos = (32*sgitg + tiisg)*nval;
// for (int i = 0; i < nval; ++i) values[pos + i] = iq2s_grid[pos + i];
// threadgroup_barrier(mem_flags::mem_threadgroup);
//}
const int ix = tiisg;
device const float * y4 = y + 32 * ix;
for (int ib32 = ix; ib32 < nb32; ib32 += 32) {
for (int i = 0; i < 32; ++i) {
yl[i] = y4[i];
}
const int ibl = ib32 / (QK_K / 32);
const int ib = ib32 % (QK_K / 32);
device const block_iq2_s * xr = x + ibl;
device const uint8_t * qs = xr->qs + 4 * ib;
device const uint8_t * qh = xr->qh + ib;
device const uint8_t * sc = xr->scales + ib;
device const uint8_t * signs = qs + QK_K/8;
device const half * dh = &xr->d;
for (int row = 0; row < N_DST; row++) {
const float db = dh[0];
const float d1 = db * (0.5f + (sc[0] & 0xf));
const float d2 = db * (0.5f + (sc[0] >> 4));
float2 sum = {0};
for (int l = 0; l < 2; ++l) {
//const threadgroup uint8_t * grid1 = (const threadgroup uint8_t *)(values + (qs[l+0] | ((qh[0] << (8-2*l)) & 0x300)));
//const threadgroup uint8_t * grid2 = (const threadgroup uint8_t *)(values + (qs[l+2] | ((qh[0] << (4-2*l)) & 0x300)));
constant uint8_t * grid1 = (constant uint8_t *)(iq2s_grid + (qs[l+0] | ((qh[0] << (8-2*l)) & 0x300)));
constant uint8_t * grid2 = (constant uint8_t *)(iq2s_grid + (qs[l+2] | ((qh[0] << (4-2*l)) & 0x300)));
for (int j = 0; j < 8; ++j) {
sum[0] += yl[8*l + j + 0] * grid1[j] * select(1, -1, signs[l+0] & kmask_iq2xs[j]);
sum[1] += yl[8*l + j + 16] * grid2[j] * select(1, -1, signs[l+2] & kmask_iq2xs[j]);
}
}
sumf[row] += d1 * sum[0] + d2 * sum[1];
dh += nb*sizeof(block_iq2_s)/2;
qs += nb*sizeof(block_iq2_s);
qh += nb*sizeof(block_iq2_s);
sc += nb*sizeof(block_iq2_s);
signs += nb*sizeof(block_iq2_s);
}
y4 += 32 * 32;
}
for (int row = 0; row < N_DST; ++row) {
all_sum = simd_sum(sumf[row]);
if (tiisg == 0) {
dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum * 0.25f;
}
}
}
[[host_name("kernel_mul_mv_iq2_s_f32")]]
kernel void kernel_mul_mv_iq2_s_f32(
device const void * src0,
device const float * src1,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
constant uint64_t & nb00,
constant uint64_t & nb01,
constant uint64_t & nb02,
constant int64_t & ne10,
constant int64_t & ne11,
constant int64_t & ne12,
constant uint64_t & nb10,
constant uint64_t & nb11,
constant uint64_t & nb12,
constant int64_t & ne0,
constant int64_t & ne1,
constant uint & r2,
constant uint & r3,
threadgroup int8_t * shared_values [[threadgroup(0)]],
uint3 tgpig[[threadgroup_position_in_grid]],
uint tiisg[[thread_index_in_simdgroup]],
uint sgitg[[simdgroup_index_in_threadgroup]]) {
kernel_mul_mv_iq2_s_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, shared_values, tgpig, tiisg, sgitg);
}
void kernel_mul_mv_iq1_s_f32_impl(
device const void * src0,
device const float * src1,
device float * dst,
int64_t ne00,
int64_t ne01,
int64_t ne02,
int64_t ne10,
int64_t ne12,
int64_t ne0,
int64_t ne1,
uint r2,
uint r3,
threadgroup int8_t * shared_value,
uint3 tgpig,
uint tiisg,
uint sgitg) {
const int nb = ne00/QK_K;
const int r0 = tgpig.x;
const int r1 = tgpig.y;
const int im = tgpig.z;
const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST;
const int ib_row = first_row * nb;
const uint i12 = im%ne12;
const uint i13 = im/ne12;
const uint offset0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
device const block_iq1_s * x = (device const block_iq1_s *) src0 + ib_row + offset0;
device const float * y = (device const float *) src1 + r1*ne10 + im*ne00*ne1;
float yl[32];
float sumf[N_DST]={0.f}, all_sum;
const int nb32 = nb * (QK_K / 32);
const int ix = tiisg;
device const float * y4 = y + 32 * ix;
for (int ib32 = ix; ib32 < nb32; ib32 += 32) {
float sumy = 0;
for (int i = 0; i < 32; ++i) {
yl[i] = y4[i];
sumy += yl[i];
}
const int ibl = ib32 / (QK_K / 32);
const int ib = ib32 % (QK_K / 32);
device const block_iq1_s * xr = x + ibl;
device const uint8_t * qs = xr->qs + 4 * ib;
device const uint16_t * qh = xr->qh + ib;
device const half * dh = &xr->d;
for (int row = 0; row < N_DST; row++) {
constant uint8_t * grid1 = (constant uint8_t *)(iq1s_grid_gpu + (qs[0] | ((qh[0] << 8) & 0x700)));
constant uint8_t * grid2 = (constant uint8_t *)(iq1s_grid_gpu + (qs[1] | ((qh[0] << 5) & 0x700)));
constant uint8_t * grid3 = (constant uint8_t *)(iq1s_grid_gpu + (qs[2] | ((qh[0] << 2) & 0x700)));
constant uint8_t * grid4 = (constant uint8_t *)(iq1s_grid_gpu + (qs[3] | ((qh[0] >> 1) & 0x700)));
float sum = 0;
for (int j = 0; j < 4; ++j) {
sum += yl[j+ 0] * (grid1[j] & 0xf) + yl[j+ 4] * (grid1[j] >> 4)
+ yl[j+ 8] * (grid2[j] & 0xf) + yl[j+12] * (grid2[j] >> 4)
+ yl[j+16] * (grid3[j] & 0xf) + yl[j+20] * (grid3[j] >> 4)
+ yl[j+24] * (grid4[j] & 0xf) + yl[j+28] * (grid4[j] >> 4);
}
sumf[row] += (float)dh[0] * (sum + sumy * (qh[0] & 0x8000 ? -1 - IQ1S_DELTA : -1 + IQ1S_DELTA)) * (2*((qh[0] >> 12) & 7) + 1);
dh += nb*sizeof(block_iq1_s)/2;
qs += nb*sizeof(block_iq1_s);
qh += nb*sizeof(block_iq1_s)/2;
}
y4 += 32 * 32;
}
for (int row = 0; row < N_DST; ++row) {
all_sum = simd_sum(sumf[row]);
if (tiisg == 0) {
dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum;
}
}
}
IQ1_M: 1.75 bpw quantization (#6302) * iq1_m: basics * iq1_m: basics-2 * iq1_m: CUDA dequantize works Very 1st shot I get PPL = 9.76 for LLaMA-v2-7B. * iq1_m: separate shifts for each group of 8 in a block We get PPL(LLaMA-v2-7B ) = 9.2810 PPL(LLaMA-v2-13B) = 6.8105 Not bad, but slightly higher than sqrt(PPL(IQ1_S) * PPL(IQ2_XXS)) which is the expected outcome given that IQ1_M is halfway between IQ1_S and IQ2_XXS in terms of bpw. From this, we would expect PPL = 9.14 for LLaMA-v2-7B PPL = 6.63 for LLaMA-v2-13B * iq1_m: go to 3-bit scales There is slight increase in PPL, but the 0.0625 bpw reduction in size is totally worth it. We now have PPL(LLaMA-v2-7B ) = 9.4469 at 1.96 bpw PPL(LLaMA-v2-13B) = 6.8717 at 1.93 bpw PPL(LLaMA-v2-70B) = 4.8568 at 1.85 bpw * iq1_m: scalar dot product * iq1_m: AVX2 dot product * iq1_m: very slightly faster AVX2 dot product * iq1_m: ARM_NEON dot product Works, but very slow (10.5 t/s) * iq1_m: Metal - dequantize works, dot product does not * iq1_m: Metal now works About the same performance as iq1_s. * iq1_m: minor * iq1_m: checking pure iq1_m quantization It is pretty bad: PPL(LLaMA-v2-7B) = 34 if we quantize output.weight with Q4_K. * iiq1_m: slightly faster ARM_NEON dot product 10.5 t/s -> 11.65 t/s * iq1_m: faster ARM_NEON dot product 11.65 t/s -> 14.9 t/s * iq1_m: another minor ARM_NEON dot product improvement 14.9 -> 15.0 t/s * iq1_m: small PPL improvement via super-block scale adjustment After quantizing block scales redo the super-block scale fit. PPL(LLaMA-v2-7B ) = 9.3346 PPL(LLaMA-v2-13B) = 6.8419 PPL(LLaMA-v2-70B) = 4.8294 PPL(Mistral-7B ) = 8.1624 * iq1_m: adapt to CUDA refactoring * iq1_m: remove unused variable We have progressed to warnings being errors. * iq1_m: add to backend-ops tests * iq1_m: fix Windows ARM * iq1_m: use common definition of iq1m_scale_t * cuda: assert -> NO_DEVICE_CODE * iq1_M: PR comments --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2024-03-26 15:21:27 +01:00
void kernel_mul_mv_iq1_m_f32_impl(
device const void * src0,
device const float * src1,
device float * dst,
int64_t ne00,
int64_t ne01,
int64_t ne02,
int64_t ne10,
int64_t ne12,
int64_t ne0,
int64_t ne1,
uint r2,
uint r3,
threadgroup int8_t * shared_value,
uint3 tgpig,
uint tiisg,
uint sgitg) {
IQ1_M: 1.75 bpw quantization (#6302) * iq1_m: basics * iq1_m: basics-2 * iq1_m: CUDA dequantize works Very 1st shot I get PPL = 9.76 for LLaMA-v2-7B. * iq1_m: separate shifts for each group of 8 in a block We get PPL(LLaMA-v2-7B ) = 9.2810 PPL(LLaMA-v2-13B) = 6.8105 Not bad, but slightly higher than sqrt(PPL(IQ1_S) * PPL(IQ2_XXS)) which is the expected outcome given that IQ1_M is halfway between IQ1_S and IQ2_XXS in terms of bpw. From this, we would expect PPL = 9.14 for LLaMA-v2-7B PPL = 6.63 for LLaMA-v2-13B * iq1_m: go to 3-bit scales There is slight increase in PPL, but the 0.0625 bpw reduction in size is totally worth it. We now have PPL(LLaMA-v2-7B ) = 9.4469 at 1.96 bpw PPL(LLaMA-v2-13B) = 6.8717 at 1.93 bpw PPL(LLaMA-v2-70B) = 4.8568 at 1.85 bpw * iq1_m: scalar dot product * iq1_m: AVX2 dot product * iq1_m: very slightly faster AVX2 dot product * iq1_m: ARM_NEON dot product Works, but very slow (10.5 t/s) * iq1_m: Metal - dequantize works, dot product does not * iq1_m: Metal now works About the same performance as iq1_s. * iq1_m: minor * iq1_m: checking pure iq1_m quantization It is pretty bad: PPL(LLaMA-v2-7B) = 34 if we quantize output.weight with Q4_K. * iiq1_m: slightly faster ARM_NEON dot product 10.5 t/s -> 11.65 t/s * iq1_m: faster ARM_NEON dot product 11.65 t/s -> 14.9 t/s * iq1_m: another minor ARM_NEON dot product improvement 14.9 -> 15.0 t/s * iq1_m: small PPL improvement via super-block scale adjustment After quantizing block scales redo the super-block scale fit. PPL(LLaMA-v2-7B ) = 9.3346 PPL(LLaMA-v2-13B) = 6.8419 PPL(LLaMA-v2-70B) = 4.8294 PPL(Mistral-7B ) = 8.1624 * iq1_m: adapt to CUDA refactoring * iq1_m: remove unused variable We have progressed to warnings being errors. * iq1_m: add to backend-ops tests * iq1_m: fix Windows ARM * iq1_m: use common definition of iq1m_scale_t * cuda: assert -> NO_DEVICE_CODE * iq1_M: PR comments --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2024-03-26 15:21:27 +01:00
const int nb = ne00/QK_K;
const int r0 = tgpig.x;
const int r1 = tgpig.y;
const int im = tgpig.z;
const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST;
const int ib_row = first_row * nb;
const uint i12 = im%ne12;
const uint i13 = im/ne12;
const uint offset0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
device const block_iq1_m * x = (device const block_iq1_m *) src0 + ib_row + offset0;
device const float * y = (device const float *) src1 + r1*ne10 + im*ne00*ne1;
float yl[32];
float sumf[N_DST]={0.f}, all_sum;
const int nb32 = nb * (QK_K / 32);
const int ix = tiisg;
device const float * y4 = y + 32 * ix;
#if QK_K != 64
IQ1_M: 1.75 bpw quantization (#6302) * iq1_m: basics * iq1_m: basics-2 * iq1_m: CUDA dequantize works Very 1st shot I get PPL = 9.76 for LLaMA-v2-7B. * iq1_m: separate shifts for each group of 8 in a block We get PPL(LLaMA-v2-7B ) = 9.2810 PPL(LLaMA-v2-13B) = 6.8105 Not bad, but slightly higher than sqrt(PPL(IQ1_S) * PPL(IQ2_XXS)) which is the expected outcome given that IQ1_M is halfway between IQ1_S and IQ2_XXS in terms of bpw. From this, we would expect PPL = 9.14 for LLaMA-v2-7B PPL = 6.63 for LLaMA-v2-13B * iq1_m: go to 3-bit scales There is slight increase in PPL, but the 0.0625 bpw reduction in size is totally worth it. We now have PPL(LLaMA-v2-7B ) = 9.4469 at 1.96 bpw PPL(LLaMA-v2-13B) = 6.8717 at 1.93 bpw PPL(LLaMA-v2-70B) = 4.8568 at 1.85 bpw * iq1_m: scalar dot product * iq1_m: AVX2 dot product * iq1_m: very slightly faster AVX2 dot product * iq1_m: ARM_NEON dot product Works, but very slow (10.5 t/s) * iq1_m: Metal - dequantize works, dot product does not * iq1_m: Metal now works About the same performance as iq1_s. * iq1_m: minor * iq1_m: checking pure iq1_m quantization It is pretty bad: PPL(LLaMA-v2-7B) = 34 if we quantize output.weight with Q4_K. * iiq1_m: slightly faster ARM_NEON dot product 10.5 t/s -> 11.65 t/s * iq1_m: faster ARM_NEON dot product 11.65 t/s -> 14.9 t/s * iq1_m: another minor ARM_NEON dot product improvement 14.9 -> 15.0 t/s * iq1_m: small PPL improvement via super-block scale adjustment After quantizing block scales redo the super-block scale fit. PPL(LLaMA-v2-7B ) = 9.3346 PPL(LLaMA-v2-13B) = 6.8419 PPL(LLaMA-v2-70B) = 4.8294 PPL(Mistral-7B ) = 8.1624 * iq1_m: adapt to CUDA refactoring * iq1_m: remove unused variable We have progressed to warnings being errors. * iq1_m: add to backend-ops tests * iq1_m: fix Windows ARM * iq1_m: use common definition of iq1m_scale_t * cuda: assert -> NO_DEVICE_CODE * iq1_M: PR comments --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2024-03-26 15:21:27 +01:00
iq1m_scale_t scale;
#endif
IQ1_M: 1.75 bpw quantization (#6302) * iq1_m: basics * iq1_m: basics-2 * iq1_m: CUDA dequantize works Very 1st shot I get PPL = 9.76 for LLaMA-v2-7B. * iq1_m: separate shifts for each group of 8 in a block We get PPL(LLaMA-v2-7B ) = 9.2810 PPL(LLaMA-v2-13B) = 6.8105 Not bad, but slightly higher than sqrt(PPL(IQ1_S) * PPL(IQ2_XXS)) which is the expected outcome given that IQ1_M is halfway between IQ1_S and IQ2_XXS in terms of bpw. From this, we would expect PPL = 9.14 for LLaMA-v2-7B PPL = 6.63 for LLaMA-v2-13B * iq1_m: go to 3-bit scales There is slight increase in PPL, but the 0.0625 bpw reduction in size is totally worth it. We now have PPL(LLaMA-v2-7B ) = 9.4469 at 1.96 bpw PPL(LLaMA-v2-13B) = 6.8717 at 1.93 bpw PPL(LLaMA-v2-70B) = 4.8568 at 1.85 bpw * iq1_m: scalar dot product * iq1_m: AVX2 dot product * iq1_m: very slightly faster AVX2 dot product * iq1_m: ARM_NEON dot product Works, but very slow (10.5 t/s) * iq1_m: Metal - dequantize works, dot product does not * iq1_m: Metal now works About the same performance as iq1_s. * iq1_m: minor * iq1_m: checking pure iq1_m quantization It is pretty bad: PPL(LLaMA-v2-7B) = 34 if we quantize output.weight with Q4_K. * iiq1_m: slightly faster ARM_NEON dot product 10.5 t/s -> 11.65 t/s * iq1_m: faster ARM_NEON dot product 11.65 t/s -> 14.9 t/s * iq1_m: another minor ARM_NEON dot product improvement 14.9 -> 15.0 t/s * iq1_m: small PPL improvement via super-block scale adjustment After quantizing block scales redo the super-block scale fit. PPL(LLaMA-v2-7B ) = 9.3346 PPL(LLaMA-v2-13B) = 6.8419 PPL(LLaMA-v2-70B) = 4.8294 PPL(Mistral-7B ) = 8.1624 * iq1_m: adapt to CUDA refactoring * iq1_m: remove unused variable We have progressed to warnings being errors. * iq1_m: add to backend-ops tests * iq1_m: fix Windows ARM * iq1_m: use common definition of iq1m_scale_t * cuda: assert -> NO_DEVICE_CODE * iq1_M: PR comments --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2024-03-26 15:21:27 +01:00
for (int ib32 = ix; ib32 < nb32; ib32 += 32) {
float4 sumy = {0.f};
for (int i = 0; i < 8; ++i) {
yl[i+ 0] = y4[i+ 0]; sumy[0] += yl[i+ 0];
yl[i+ 8] = y4[i+ 8]; sumy[1] += yl[i+ 8];
yl[i+16] = y4[i+16]; sumy[2] += yl[i+16];
yl[i+24] = y4[i+24]; sumy[3] += yl[i+24];
}
const int ibl = ib32 / (QK_K / 32);
const int ib = ib32 % (QK_K / 32);
device const block_iq1_m * xr = x + ibl;
device const uint8_t * qs = xr->qs + 4 * ib;
device const uint8_t * qh = xr->qh + 2 * ib;
device const uint16_t * sc = (device const uint16_t *)xr->scales;
for (int row = 0; row < N_DST; row++) {
#if QK_K != 64
IQ1_M: 1.75 bpw quantization (#6302) * iq1_m: basics * iq1_m: basics-2 * iq1_m: CUDA dequantize works Very 1st shot I get PPL = 9.76 for LLaMA-v2-7B. * iq1_m: separate shifts for each group of 8 in a block We get PPL(LLaMA-v2-7B ) = 9.2810 PPL(LLaMA-v2-13B) = 6.8105 Not bad, but slightly higher than sqrt(PPL(IQ1_S) * PPL(IQ2_XXS)) which is the expected outcome given that IQ1_M is halfway between IQ1_S and IQ2_XXS in terms of bpw. From this, we would expect PPL = 9.14 for LLaMA-v2-7B PPL = 6.63 for LLaMA-v2-13B * iq1_m: go to 3-bit scales There is slight increase in PPL, but the 0.0625 bpw reduction in size is totally worth it. We now have PPL(LLaMA-v2-7B ) = 9.4469 at 1.96 bpw PPL(LLaMA-v2-13B) = 6.8717 at 1.93 bpw PPL(LLaMA-v2-70B) = 4.8568 at 1.85 bpw * iq1_m: scalar dot product * iq1_m: AVX2 dot product * iq1_m: very slightly faster AVX2 dot product * iq1_m: ARM_NEON dot product Works, but very slow (10.5 t/s) * iq1_m: Metal - dequantize works, dot product does not * iq1_m: Metal now works About the same performance as iq1_s. * iq1_m: minor * iq1_m: checking pure iq1_m quantization It is pretty bad: PPL(LLaMA-v2-7B) = 34 if we quantize output.weight with Q4_K. * iiq1_m: slightly faster ARM_NEON dot product 10.5 t/s -> 11.65 t/s * iq1_m: faster ARM_NEON dot product 11.65 t/s -> 14.9 t/s * iq1_m: another minor ARM_NEON dot product improvement 14.9 -> 15.0 t/s * iq1_m: small PPL improvement via super-block scale adjustment After quantizing block scales redo the super-block scale fit. PPL(LLaMA-v2-7B ) = 9.3346 PPL(LLaMA-v2-13B) = 6.8419 PPL(LLaMA-v2-70B) = 4.8294 PPL(Mistral-7B ) = 8.1624 * iq1_m: adapt to CUDA refactoring * iq1_m: remove unused variable We have progressed to warnings being errors. * iq1_m: add to backend-ops tests * iq1_m: fix Windows ARM * iq1_m: use common definition of iq1m_scale_t * cuda: assert -> NO_DEVICE_CODE * iq1_M: PR comments --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2024-03-26 15:21:27 +01:00
scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000);
#endif
IQ1_M: 1.75 bpw quantization (#6302) * iq1_m: basics * iq1_m: basics-2 * iq1_m: CUDA dequantize works Very 1st shot I get PPL = 9.76 for LLaMA-v2-7B. * iq1_m: separate shifts for each group of 8 in a block We get PPL(LLaMA-v2-7B ) = 9.2810 PPL(LLaMA-v2-13B) = 6.8105 Not bad, but slightly higher than sqrt(PPL(IQ1_S) * PPL(IQ2_XXS)) which is the expected outcome given that IQ1_M is halfway between IQ1_S and IQ2_XXS in terms of bpw. From this, we would expect PPL = 9.14 for LLaMA-v2-7B PPL = 6.63 for LLaMA-v2-13B * iq1_m: go to 3-bit scales There is slight increase in PPL, but the 0.0625 bpw reduction in size is totally worth it. We now have PPL(LLaMA-v2-7B ) = 9.4469 at 1.96 bpw PPL(LLaMA-v2-13B) = 6.8717 at 1.93 bpw PPL(LLaMA-v2-70B) = 4.8568 at 1.85 bpw * iq1_m: scalar dot product * iq1_m: AVX2 dot product * iq1_m: very slightly faster AVX2 dot product * iq1_m: ARM_NEON dot product Works, but very slow (10.5 t/s) * iq1_m: Metal - dequantize works, dot product does not * iq1_m: Metal now works About the same performance as iq1_s. * iq1_m: minor * iq1_m: checking pure iq1_m quantization It is pretty bad: PPL(LLaMA-v2-7B) = 34 if we quantize output.weight with Q4_K. * iiq1_m: slightly faster ARM_NEON dot product 10.5 t/s -> 11.65 t/s * iq1_m: faster ARM_NEON dot product 11.65 t/s -> 14.9 t/s * iq1_m: another minor ARM_NEON dot product improvement 14.9 -> 15.0 t/s * iq1_m: small PPL improvement via super-block scale adjustment After quantizing block scales redo the super-block scale fit. PPL(LLaMA-v2-7B ) = 9.3346 PPL(LLaMA-v2-13B) = 6.8419 PPL(LLaMA-v2-70B) = 4.8294 PPL(Mistral-7B ) = 8.1624 * iq1_m: adapt to CUDA refactoring * iq1_m: remove unused variable We have progressed to warnings being errors. * iq1_m: add to backend-ops tests * iq1_m: fix Windows ARM * iq1_m: use common definition of iq1m_scale_t * cuda: assert -> NO_DEVICE_CODE * iq1_M: PR comments --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2024-03-26 15:21:27 +01:00
constant uint8_t * grid1 = (constant uint8_t *)(iq1s_grid_gpu + (qs[0] | ((qh[0] << 8) & 0x700)));
constant uint8_t * grid2 = (constant uint8_t *)(iq1s_grid_gpu + (qs[1] | ((qh[0] << 4) & 0x700)));
constant uint8_t * grid3 = (constant uint8_t *)(iq1s_grid_gpu + (qs[2] | ((qh[1] << 8) & 0x700)));
constant uint8_t * grid4 = (constant uint8_t *)(iq1s_grid_gpu + (qs[3] | ((qh[1] << 4) & 0x700)));
float2 sum = {0.f};
for (int j = 0; j < 4; ++j) {
sum[0] += yl[j+ 0] * (grid1[j] & 0xf) + yl[j+ 4] * (grid1[j] >> 4)
+ yl[j+ 8] * (grid2[j] & 0xf) + yl[j+12] * (grid2[j] >> 4);
sum[1] += yl[j+16] * (grid3[j] & 0xf) + yl[j+20] * (grid3[j] >> 4)
+ yl[j+24] * (grid4[j] & 0xf) + yl[j+28] * (grid4[j] >> 4);
}
const float delta1 = sumy[0] * (qh[0] & 0x08 ? -1 - IQ1M_DELTA : -1 + IQ1M_DELTA) + sumy[1] * (qh[0] & 0x80 ? -1 - IQ1M_DELTA : -1 + IQ1M_DELTA);
const float delta2 = sumy[2] * (qh[1] & 0x08 ? -1 - IQ1M_DELTA : -1 + IQ1M_DELTA) + sumy[3] * (qh[1] & 0x80 ? -1 - IQ1M_DELTA : -1 + IQ1M_DELTA);
#if QK_K == 64
const float d = (float) *((device const half *)(sc - 1));
sumf[row] += d * ((sum[0] + delta1) * (2*((sc[0] >> (8*(ib%2)+0)) & 0xf) + 1) +
(sum[1] + delta2) * (2*((sc[0] >> (8*(ib%2)+4)) & 0xf) + 1));
#else
IQ1_M: 1.75 bpw quantization (#6302) * iq1_m: basics * iq1_m: basics-2 * iq1_m: CUDA dequantize works Very 1st shot I get PPL = 9.76 for LLaMA-v2-7B. * iq1_m: separate shifts for each group of 8 in a block We get PPL(LLaMA-v2-7B ) = 9.2810 PPL(LLaMA-v2-13B) = 6.8105 Not bad, but slightly higher than sqrt(PPL(IQ1_S) * PPL(IQ2_XXS)) which is the expected outcome given that IQ1_M is halfway between IQ1_S and IQ2_XXS in terms of bpw. From this, we would expect PPL = 9.14 for LLaMA-v2-7B PPL = 6.63 for LLaMA-v2-13B * iq1_m: go to 3-bit scales There is slight increase in PPL, but the 0.0625 bpw reduction in size is totally worth it. We now have PPL(LLaMA-v2-7B ) = 9.4469 at 1.96 bpw PPL(LLaMA-v2-13B) = 6.8717 at 1.93 bpw PPL(LLaMA-v2-70B) = 4.8568 at 1.85 bpw * iq1_m: scalar dot product * iq1_m: AVX2 dot product * iq1_m: very slightly faster AVX2 dot product * iq1_m: ARM_NEON dot product Works, but very slow (10.5 t/s) * iq1_m: Metal - dequantize works, dot product does not * iq1_m: Metal now works About the same performance as iq1_s. * iq1_m: minor * iq1_m: checking pure iq1_m quantization It is pretty bad: PPL(LLaMA-v2-7B) = 34 if we quantize output.weight with Q4_K. * iiq1_m: slightly faster ARM_NEON dot product 10.5 t/s -> 11.65 t/s * iq1_m: faster ARM_NEON dot product 11.65 t/s -> 14.9 t/s * iq1_m: another minor ARM_NEON dot product improvement 14.9 -> 15.0 t/s * iq1_m: small PPL improvement via super-block scale adjustment After quantizing block scales redo the super-block scale fit. PPL(LLaMA-v2-7B ) = 9.3346 PPL(LLaMA-v2-13B) = 6.8419 PPL(LLaMA-v2-70B) = 4.8294 PPL(Mistral-7B ) = 8.1624 * iq1_m: adapt to CUDA refactoring * iq1_m: remove unused variable We have progressed to warnings being errors. * iq1_m: add to backend-ops tests * iq1_m: fix Windows ARM * iq1_m: use common definition of iq1m_scale_t * cuda: assert -> NO_DEVICE_CODE * iq1_M: PR comments --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2024-03-26 15:21:27 +01:00
sumf[row] += (float)scale.f16 * ((sum[0] + delta1) * (2*((sc[ib/2] >> (6*(ib%2)+0)) & 7) + 1) +
(sum[1] + delta2) * (2*((sc[ib/2] >> (6*(ib%2)+3)) & 7) + 1));
#endif
IQ1_M: 1.75 bpw quantization (#6302) * iq1_m: basics * iq1_m: basics-2 * iq1_m: CUDA dequantize works Very 1st shot I get PPL = 9.76 for LLaMA-v2-7B. * iq1_m: separate shifts for each group of 8 in a block We get PPL(LLaMA-v2-7B ) = 9.2810 PPL(LLaMA-v2-13B) = 6.8105 Not bad, but slightly higher than sqrt(PPL(IQ1_S) * PPL(IQ2_XXS)) which is the expected outcome given that IQ1_M is halfway between IQ1_S and IQ2_XXS in terms of bpw. From this, we would expect PPL = 9.14 for LLaMA-v2-7B PPL = 6.63 for LLaMA-v2-13B * iq1_m: go to 3-bit scales There is slight increase in PPL, but the 0.0625 bpw reduction in size is totally worth it. We now have PPL(LLaMA-v2-7B ) = 9.4469 at 1.96 bpw PPL(LLaMA-v2-13B) = 6.8717 at 1.93 bpw PPL(LLaMA-v2-70B) = 4.8568 at 1.85 bpw * iq1_m: scalar dot product * iq1_m: AVX2 dot product * iq1_m: very slightly faster AVX2 dot product * iq1_m: ARM_NEON dot product Works, but very slow (10.5 t/s) * iq1_m: Metal - dequantize works, dot product does not * iq1_m: Metal now works About the same performance as iq1_s. * iq1_m: minor * iq1_m: checking pure iq1_m quantization It is pretty bad: PPL(LLaMA-v2-7B) = 34 if we quantize output.weight with Q4_K. * iiq1_m: slightly faster ARM_NEON dot product 10.5 t/s -> 11.65 t/s * iq1_m: faster ARM_NEON dot product 11.65 t/s -> 14.9 t/s * iq1_m: another minor ARM_NEON dot product improvement 14.9 -> 15.0 t/s * iq1_m: small PPL improvement via super-block scale adjustment After quantizing block scales redo the super-block scale fit. PPL(LLaMA-v2-7B ) = 9.3346 PPL(LLaMA-v2-13B) = 6.8419 PPL(LLaMA-v2-70B) = 4.8294 PPL(Mistral-7B ) = 8.1624 * iq1_m: adapt to CUDA refactoring * iq1_m: remove unused variable We have progressed to warnings being errors. * iq1_m: add to backend-ops tests * iq1_m: fix Windows ARM * iq1_m: use common definition of iq1m_scale_t * cuda: assert -> NO_DEVICE_CODE * iq1_M: PR comments --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2024-03-26 15:21:27 +01:00
sc += nb*sizeof(block_iq1_m)/2;
qs += nb*sizeof(block_iq1_m);
qh += nb*sizeof(block_iq1_m);
}
y4 += 32 * 32;
}
for (int row = 0; row < N_DST; ++row) {
all_sum = simd_sum(sumf[row]);
if (tiisg == 0) {
dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum;
}
}
}
void kernel_mul_mv_iq4_nl_f32_impl(
device const void * src0,
device const float * src1,
device float * dst,
int64_t ne00,
int64_t ne01,
int64_t ne02,
int64_t ne10,
int64_t ne12,
int64_t ne0,
int64_t ne1,
uint r2,
uint r3,
threadgroup int8_t * shared_values_i8,
uint3 tgpig,
uint tiisg,
uint sgitg) {
threadgroup float * shared_values = (threadgroup float *)shared_values_i8;
const int nb = ne00/QK4_NL;
const int r0 = tgpig.x;
const int r1 = tgpig.y;
const int im = tgpig.z;
const int first_row = (r0 * 2 + sgitg) * 2;
const int ib_row = first_row * nb;
const uint i12 = im%ne12;
const uint i13 = im/ne12;
const uint offset0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
device const block_iq4_nl * x = (device const block_iq4_nl *) src0 + ib_row + offset0;
device const float * y = (device const float *) src1 + r1*ne10 + im*ne00*ne1;
const int ix = tiisg/2; // 0...15
const int it = tiisg%2; // 0 or 1
shared_values[tiisg] = kvalues_iq4nl_f[tiisg%16];
threadgroup_barrier(mem_flags::mem_threadgroup);
float4 yl[4];
float sumf[2]={0.f}, all_sum;
device const float * yb = y + ix * QK4_NL + it * 8;
uint32_t aux32[2];
thread const uint8_t * q8 = (thread const uint8_t *)aux32;
float4 qf1, qf2;
for (int ib = ix; ib < nb; ib += 16) {
device const float4 * y4 = (device const float4 *)yb;
yl[0] = y4[0]; yl[1] = y4[4]; yl[2] = y4[1]; yl[3] = y4[5];
for (int row = 0; row < 2; ++row) {
device const block_iq4_nl & xb = x[row*nb + ib];
device const uint16_t * q4 = (device const uint16_t *)(xb.qs + 8*it);
float4 acc1 = {0.f}, acc2 = {0.f};
aux32[0] = q4[0] | (q4[1] << 16);
aux32[1] = (aux32[0] >> 4) & 0x0f0f0f0f;
aux32[0] &= 0x0f0f0f0f;
qf1 = {shared_values[q8[0]], shared_values[q8[1]], shared_values[q8[2]], shared_values[q8[3]]};
qf2 = {shared_values[q8[4]], shared_values[q8[5]], shared_values[q8[6]], shared_values[q8[7]]};
acc1 += yl[0] * qf1;
acc2 += yl[1] * qf2;
aux32[0] = q4[2] | (q4[3] << 16);
aux32[1] = (aux32[0] >> 4) & 0x0f0f0f0f;
aux32[0] &= 0x0f0f0f0f;
qf1 = {shared_values[q8[0]], shared_values[q8[1]], shared_values[q8[2]], shared_values[q8[3]]};
qf2 = {shared_values[q8[4]], shared_values[q8[5]], shared_values[q8[6]], shared_values[q8[7]]};
acc1 += yl[2] * qf1;
acc2 += yl[3] * qf2;
acc1 += acc2;
sumf[row] += (float)xb.d * (acc1[0] + acc1[1] + acc1[2] + acc1[3]);
}
yb += 16 * QK4_NL;
}
for (int row = 0; row < 2; ++row) {
all_sum = simd_sum(sumf[row]);
if (tiisg == 0) {
dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum;
}
}
}
#if QK_K != 64
void kernel_mul_mv_iq4_xs_f32_impl(
device const void * src0,
device const float * src1,
device float * dst,
int64_t ne00,
int64_t ne01,
int64_t ne02,
int64_t ne10,
int64_t ne12,
int64_t ne0,
int64_t ne1,
uint r2,
uint r3,
threadgroup int8_t * shared_values_i8,
uint3 tgpig,
uint tiisg,
uint sgitg) {
threadgroup float * shared_values = (threadgroup float *)shared_values_i8;
const int nb = ne00/QK_K;
const int r0 = tgpig.x;
const int r1 = tgpig.y;
const int im = tgpig.z;
const int first_row = (r0 * 2 + sgitg) * 2;
const int ib_row = first_row * nb;
const uint i12 = im%ne12;
const uint i13 = im/ne12;
const uint offset0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
device const block_iq4_xs * x = (device const block_iq4_xs *) src0 + ib_row + offset0;
device const float * y = (device const float *) src1 + r1*ne10 + im*ne00*ne1;
const int ix = tiisg/16; // 0 or 1
const int it = tiisg%16; // 0...15
const int ib = it/2;
const int il = it%2;
shared_values[tiisg] = kvalues_iq4nl_f[tiisg%16];
threadgroup_barrier(mem_flags::mem_threadgroup);
float4 yl[4];
float sumf[2]={0.f}, all_sum;
device const float * yb = y + ix * QK_K + ib * 32 + il * 8;
uint32_t aux32[2];
thread const uint8_t * q8 = (thread const uint8_t *)aux32;
float4 qf1, qf2;
for (int ibl = ix; ibl < nb; ibl += 2) {
device const float4 * y4 = (device const float4 *)yb;
yl[0] = y4[0]; yl[1] = y4[4]; yl[2] = y4[1]; yl[3] = y4[5];
for (int row = 0; row < 2; ++row) {
device const block_iq4_xs & xb = x[row*nb + ibl];
device const uint32_t * q4 = (device const uint32_t *)(xb.qs + 16*ib + 8*il);
float4 acc1 = {0.f}, acc2 = {0.f};
aux32[0] = q4[0] & 0x0f0f0f0f;
aux32[1] = (q4[0] >> 4) & 0x0f0f0f0f;
qf1 = {shared_values[q8[0]], shared_values[q8[1]], shared_values[q8[2]], shared_values[q8[3]]};
qf2 = {shared_values[q8[4]], shared_values[q8[5]], shared_values[q8[6]], shared_values[q8[7]]};
acc1 += yl[0] * qf1;
acc2 += yl[1] * qf2;
aux32[0] = q4[1] & 0x0f0f0f0f;
aux32[1] = (q4[1] >> 4) & 0x0f0f0f0f;
qf1 = {shared_values[q8[0]], shared_values[q8[1]], shared_values[q8[2]], shared_values[q8[3]]};
qf2 = {shared_values[q8[4]], shared_values[q8[5]], shared_values[q8[6]], shared_values[q8[7]]};
acc1 += yl[2] * qf1;
acc2 += yl[3] * qf2;
acc1 += acc2;
const int ls = (((xb.scales_l[ib/2] >> 4*(ib%2)) & 0xf) | (((xb.scales_h >> 2*ib) & 3) << 4)) - 32;
sumf[row] += (float)xb.d * ls * (acc1[0] + acc1[1] + acc1[2] + acc1[3]);
}
yb += 2 * QK_K;
}
for (int row = 0; row < 2; ++row) {
all_sum = simd_sum(sumf[row]);
if (tiisg == 0) {
dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum;
}
}
}
#endif
[[host_name("kernel_mul_mv_iq1_s_f32")]]
kernel void kernel_mul_mv_iq1_s_f32(
device const void * src0,
device const float * src1,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
constant uint64_t & nb00,
constant uint64_t & nb01,
constant uint64_t & nb02,
constant int64_t & ne10,
constant int64_t & ne11,
constant int64_t & ne12,
constant uint64_t & nb10,
constant uint64_t & nb11,
constant uint64_t & nb12,
constant int64_t & ne0,
constant int64_t & ne1,
constant uint & r2,
constant uint & r3,
uint3 tgpig[[threadgroup_position_in_grid]],
uint tiisg[[thread_index_in_simdgroup]],
uint sgitg[[simdgroup_index_in_threadgroup]]) {
kernel_mul_mv_iq1_s_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, nullptr, tgpig, tiisg, sgitg);
}
IQ1_M: 1.75 bpw quantization (#6302) * iq1_m: basics * iq1_m: basics-2 * iq1_m: CUDA dequantize works Very 1st shot I get PPL = 9.76 for LLaMA-v2-7B. * iq1_m: separate shifts for each group of 8 in a block We get PPL(LLaMA-v2-7B ) = 9.2810 PPL(LLaMA-v2-13B) = 6.8105 Not bad, but slightly higher than sqrt(PPL(IQ1_S) * PPL(IQ2_XXS)) which is the expected outcome given that IQ1_M is halfway between IQ1_S and IQ2_XXS in terms of bpw. From this, we would expect PPL = 9.14 for LLaMA-v2-7B PPL = 6.63 for LLaMA-v2-13B * iq1_m: go to 3-bit scales There is slight increase in PPL, but the 0.0625 bpw reduction in size is totally worth it. We now have PPL(LLaMA-v2-7B ) = 9.4469 at 1.96 bpw PPL(LLaMA-v2-13B) = 6.8717 at 1.93 bpw PPL(LLaMA-v2-70B) = 4.8568 at 1.85 bpw * iq1_m: scalar dot product * iq1_m: AVX2 dot product * iq1_m: very slightly faster AVX2 dot product * iq1_m: ARM_NEON dot product Works, but very slow (10.5 t/s) * iq1_m: Metal - dequantize works, dot product does not * iq1_m: Metal now works About the same performance as iq1_s. * iq1_m: minor * iq1_m: checking pure iq1_m quantization It is pretty bad: PPL(LLaMA-v2-7B) = 34 if we quantize output.weight with Q4_K. * iiq1_m: slightly faster ARM_NEON dot product 10.5 t/s -> 11.65 t/s * iq1_m: faster ARM_NEON dot product 11.65 t/s -> 14.9 t/s * iq1_m: another minor ARM_NEON dot product improvement 14.9 -> 15.0 t/s * iq1_m: small PPL improvement via super-block scale adjustment After quantizing block scales redo the super-block scale fit. PPL(LLaMA-v2-7B ) = 9.3346 PPL(LLaMA-v2-13B) = 6.8419 PPL(LLaMA-v2-70B) = 4.8294 PPL(Mistral-7B ) = 8.1624 * iq1_m: adapt to CUDA refactoring * iq1_m: remove unused variable We have progressed to warnings being errors. * iq1_m: add to backend-ops tests * iq1_m: fix Windows ARM * iq1_m: use common definition of iq1m_scale_t * cuda: assert -> NO_DEVICE_CODE * iq1_M: PR comments --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2024-03-26 15:21:27 +01:00
[[host_name("kernel_mul_mv_iq1_m_f32")]]
kernel void kernel_mul_mv_iq1_m_f32(
device const void * src0,
device const float * src1,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
constant uint64_t & nb00,
constant uint64_t & nb01,
constant uint64_t & nb02,
constant int64_t & ne10,
constant int64_t & ne11,
constant int64_t & ne12,
constant uint64_t & nb10,
constant uint64_t & nb11,
constant uint64_t & nb12,
constant int64_t & ne0,
constant int64_t & ne1,
constant uint & r2,
constant uint & r3,
uint3 tgpig[[threadgroup_position_in_grid]],
uint tiisg[[thread_index_in_simdgroup]],
uint sgitg[[simdgroup_index_in_threadgroup]]) {
kernel_mul_mv_iq1_m_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, nullptr, tgpig, tiisg, sgitg);
IQ1_M: 1.75 bpw quantization (#6302) * iq1_m: basics * iq1_m: basics-2 * iq1_m: CUDA dequantize works Very 1st shot I get PPL = 9.76 for LLaMA-v2-7B. * iq1_m: separate shifts for each group of 8 in a block We get PPL(LLaMA-v2-7B ) = 9.2810 PPL(LLaMA-v2-13B) = 6.8105 Not bad, but slightly higher than sqrt(PPL(IQ1_S) * PPL(IQ2_XXS)) which is the expected outcome given that IQ1_M is halfway between IQ1_S and IQ2_XXS in terms of bpw. From this, we would expect PPL = 9.14 for LLaMA-v2-7B PPL = 6.63 for LLaMA-v2-13B * iq1_m: go to 3-bit scales There is slight increase in PPL, but the 0.0625 bpw reduction in size is totally worth it. We now have PPL(LLaMA-v2-7B ) = 9.4469 at 1.96 bpw PPL(LLaMA-v2-13B) = 6.8717 at 1.93 bpw PPL(LLaMA-v2-70B) = 4.8568 at 1.85 bpw * iq1_m: scalar dot product * iq1_m: AVX2 dot product * iq1_m: very slightly faster AVX2 dot product * iq1_m: ARM_NEON dot product Works, but very slow (10.5 t/s) * iq1_m: Metal - dequantize works, dot product does not * iq1_m: Metal now works About the same performance as iq1_s. * iq1_m: minor * iq1_m: checking pure iq1_m quantization It is pretty bad: PPL(LLaMA-v2-7B) = 34 if we quantize output.weight with Q4_K. * iiq1_m: slightly faster ARM_NEON dot product 10.5 t/s -> 11.65 t/s * iq1_m: faster ARM_NEON dot product 11.65 t/s -> 14.9 t/s * iq1_m: another minor ARM_NEON dot product improvement 14.9 -> 15.0 t/s * iq1_m: small PPL improvement via super-block scale adjustment After quantizing block scales redo the super-block scale fit. PPL(LLaMA-v2-7B ) = 9.3346 PPL(LLaMA-v2-13B) = 6.8419 PPL(LLaMA-v2-70B) = 4.8294 PPL(Mistral-7B ) = 8.1624 * iq1_m: adapt to CUDA refactoring * iq1_m: remove unused variable We have progressed to warnings being errors. * iq1_m: add to backend-ops tests * iq1_m: fix Windows ARM * iq1_m: use common definition of iq1m_scale_t * cuda: assert -> NO_DEVICE_CODE * iq1_M: PR comments --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2024-03-26 15:21:27 +01:00
}
[[host_name("kernel_mul_mv_iq4_nl_f32")]]
kernel void kernel_mul_mv_iq4_nl_f32(
device const void * src0,
device const float * src1,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
constant uint64_t & nb00,
constant uint64_t & nb01,
constant uint64_t & nb02,
constant int64_t & ne10,
constant int64_t & ne11,
constant int64_t & ne12,
constant uint64_t & nb10,
constant uint64_t & nb11,
constant uint64_t & nb12,
constant int64_t & ne0,
constant int64_t & ne1,
constant uint & r2,
constant uint & r3,
threadgroup int8_t * shared_values [[threadgroup(0)]],
uint3 tgpig[[threadgroup_position_in_grid]],
uint tiisg[[thread_index_in_simdgroup]],
uint sgitg[[simdgroup_index_in_threadgroup]]) {
kernel_mul_mv_iq4_nl_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, shared_values, tgpig, tiisg, sgitg);
}
[[host_name("kernel_mul_mv_iq4_xs_f32")]]
kernel void kernel_mul_mv_iq4_xs_f32(
device const void * src0,
device const float * src1,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
constant uint64_t & nb00,
constant uint64_t & nb01,
constant uint64_t & nb02,
constant int64_t & ne10,
constant int64_t & ne11,
constant int64_t & ne12,
constant uint64_t & nb10,
constant uint64_t & nb11,
constant uint64_t & nb12,
constant int64_t & ne0,
constant int64_t & ne1,
constant uint & r2,
constant uint & r3,
threadgroup int8_t * shared_values [[threadgroup(0)]],
uint3 tgpig[[threadgroup_position_in_grid]],
uint tiisg[[thread_index_in_simdgroup]],
uint sgitg[[simdgroup_index_in_threadgroup]]) {
#if QK_K == 64
kernel_mul_mv_iq4_nl_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, shared_values, tgpig, tiisg, sgitg);
#else
kernel_mul_mv_iq4_xs_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, shared_values, tgpig, tiisg, sgitg);
#endif
}
//============================= templates and their specializations =============================
// NOTE: this is not dequantizing - we are simply fitting the template
template <typename type4x4>
void dequantize_f32(device const float4x4 * src, short il, thread type4x4 & reg) {
float4x4 temp = *(((device float4x4 *)src));
for (int i = 0; i < 16; i++){
reg[i/4][i%4] = temp[i/4][i%4];
}
}
template <typename type4x4>
void dequantize_f16(device const half4x4 * src, short il, thread type4x4 & reg) {
half4x4 temp = *(((device half4x4 *)src));
for (int i = 0; i < 16; i++){
reg[i/4][i%4] = temp[i/4][i%4];
}
}
template <typename type4x4>
void dequantize_q4_0(device const block_q4_0 *xb, short il, thread type4x4 & reg) {
device const uint16_t * qs = ((device const uint16_t *)xb + 1);
const float d1 = il ? (xb->d / 16.h) : xb->d;
const float d2 = d1 / 256.f;
const float md = -8.h * xb->d;
const ushort mask0 = il ? 0x00F0 : 0x000F;
const ushort mask1 = mask0 << 8;
for (int i=0;i<8;i++) {
reg[i/2][2*(i%2)+0] = d1 * (qs[i] & mask0) + md;
reg[i/2][2*(i%2)+1] = d2 * (qs[i] & mask1) + md;
}
}
template <typename type4x4>
void dequantize_q4_1(device const block_q4_1 *xb, short il, thread type4x4 & reg) {
device const uint16_t * qs = ((device const uint16_t *)xb + 2);
const float d1 = il ? (xb->d / 16.h) : xb->d;
const float d2 = d1 / 256.f;
const float m = xb->m;
const ushort mask0 = il ? 0x00F0 : 0x000F;
const ushort mask1 = mask0 << 8;
for (int i=0;i<8;i++) {
reg[i/2][2*(i%2)+0] = ((qs[i] & mask0) * d1) + m;
reg[i/2][2*(i%2)+1] = ((qs[i] & mask1) * d2) + m;
}
}
template <typename type4x4>
void dequantize_q5_0(device const block_q5_0 *xb, short il, thread type4x4 & reg) {
device const uint16_t * qs = ((device const uint16_t *)xb + 3);
const float d = xb->d;
const float md = -16.h * xb->d;
const ushort mask = il ? 0x00F0 : 0x000F;
const uint32_t qh = *((device const uint32_t *)xb->qh);
const int x_mv = il ? 4 : 0;
const int gh_mv = il ? 12 : 0;
const int gh_bk = il ? 0 : 4;
for (int i = 0; i < 8; i++) {
// extract the 5-th bits for x0 and x1
const uint8_t xh_0 = ((qh >> (gh_mv + 2*i )) << gh_bk) & 0x10;
const uint8_t xh_1 = ((qh >> (gh_mv + 2*i+1)) << gh_bk) & 0x10;
// combine the 4-bits from qs with the 5th bit
const int32_t x0 = ((((qs[i] ) & mask) >> x_mv) | xh_0);
const int32_t x1 = ((((qs[i] >> 8) & mask) >> x_mv) | xh_1);
reg[i/2][2*(i%2)+0] = d * x0 + md;
reg[i/2][2*(i%2)+1] = d * x1 + md;
}
}
template <typename type4x4>
void dequantize_q5_1(device const block_q5_1 *xb, short il, thread type4x4 & reg) {
device const uint16_t * qs = ((device const uint16_t *)xb + 4);
const float d = xb->d;
const float m = xb->m;
const ushort mask = il ? 0x00F0 : 0x000F;
const uint32_t qh = *((device const uint32_t *)xb->qh);
const int x_mv = il ? 4 : 0;
const int gh_mv = il ? 12 : 0;
const int gh_bk = il ? 0 : 4;
for (int i = 0; i < 8; i++) {
// extract the 5-th bits for x0 and x1
const uint8_t xh_0 = ((qh >> (gh_mv + 2*i )) << gh_bk) & 0x10;
const uint8_t xh_1 = ((qh >> (gh_mv + 2*i+1)) << gh_bk) & 0x10;
// combine the 4-bits from qs with the 5th bit
const int32_t x0 = ((((qs[i] ) & mask) >> x_mv) | xh_0);
const int32_t x1 = ((((qs[i] >> 8) & mask) >> x_mv) | xh_1);
reg[i/2][2*(i%2)+0] = d * x0 + m;
reg[i/2][2*(i%2)+1] = d * x1 + m;
}
}
template <typename type4x4>
void dequantize_q8_0(device const block_q8_0 *xb, short il, thread type4x4 & reg) {
device const int8_t * qs = ((device const int8_t *)xb->qs);
const half d = xb->d;
for (int i = 0; i < 16; i++) {
reg[i/4][i%4] = (qs[i + 16*il] * d);
}
}
template <typename type4x4>
void dequantize_q2_K(device const block_q2_K *xb, short il, thread type4x4 & reg) {
const float d = xb->d;
const float min = xb->dmin;
device const uint8_t * q = (device const uint8_t *)xb->qs;
float dl, ml;
uint8_t sc = xb->scales[il];
#if QK_K == 256
q = q + 32*(il/8) + 16*(il&1);
il = (il/2)%4;
#endif
half coef = il>1 ? (il>2 ? 1/64.h : 1/16.h) : (il>0 ? 1/4.h : 1.h);
uchar mask = il>1 ? (il>2 ? 192 : 48) : (il>0 ? 12 : 3);
dl = d * (sc & 0xF) * coef, ml = min * (sc >> 4);
for (int i = 0; i < 16; ++i) {
reg[i/4][i%4] = dl * (q[i] & mask) - ml;
}
}
template <typename type4x4>
void dequantize_q3_K(device const block_q3_K *xb, short il, thread type4x4 & reg) {
const half d_all = xb->d;
device const uint8_t * q = (device const uint8_t *)xb->qs;
device const uint8_t * h = (device const uint8_t *)xb->hmask;
device const int8_t * scales = (device const int8_t *)xb->scales;
#if QK_K == 256
q = q + 32 * (il/8) + 16 * (il&1);
h = h + 16 * (il&1);
uint8_t m = 1 << (il/2);
uint16_t kmask1 = (il/4)>1 ? ((il/4)>2 ? 192 : 48) : \
((il/4)>0 ? 12 : 3);
uint16_t kmask2 = il/8 ? 0xF0 : 0x0F;
uint16_t scale_2 = scales[il%8], scale_1 = scales[8 + il%4];
int16_t dl_int = (il/4)&1 ? (scale_2&kmask2) | ((scale_1&kmask1) << 2)
: (scale_2&kmask2) | ((scale_1&kmask1) << 4);
float dl = il<8 ? d_all * (dl_int - 32.f) : d_all * (dl_int / 16.f - 32.f);
const float ml = 4.f * dl;
il = (il/2) & 3;
const half coef = il>1 ? (il>2 ? 1/64.h : 1/16.h) : (il>0 ? 1/4.h : 1.h);
const uint8_t mask = il>1 ? (il>2 ? 192 : 48) : (il>0 ? 12 : 3);
dl *= coef;
for (int i = 0; i < 16; ++i) {
reg[i/4][i%4] = dl * (q[i] & mask) - (h[i] & m ? 0 : ml);
}
#else
float kcoef = il&1 ? 1.f/16.f : 1.f;
uint16_t kmask = il&1 ? 0xF0 : 0x0F;
float dl = d_all * ((scales[il/2] & kmask) * kcoef - 8);
float coef = il>1 ? (il>2 ? 1/64.h : 1/16.h) : (il>0 ? 1/4.h : 1.h);
uint8_t mask = il>1 ? (il>2 ? 192 : 48) : (il>0 ? 12 : 3);
uint8_t m = 1<<(il*2);
for (int i = 0; i < 16; ++i) {
reg[i/4][i%4] = coef * dl * ((q[i] & mask) - ((h[i%8] & (m * (1 + i/8))) ? 0 : 4.f/coef));
}
#endif
}
static inline uchar2 get_scale_min_k4_just2(int j, int k, device const uchar * q) {
return j < 4 ? uchar2{uchar(q[j+0+k] & 63), uchar(q[j+4+k] & 63)}
: uchar2{uchar((q[j+4+k] & 0xF) | ((q[j-4+k] & 0xc0) >> 2)), uchar((q[j+4+k] >> 4) | ((q[j-0+k] & 0xc0) >> 2))};
}
template <typename type4x4>
void dequantize_q4_K(device const block_q4_K *xb, short il, thread type4x4 & reg) {
device const uchar * q = xb->qs;
#if QK_K == 256
short is = (il/4) * 2;
q = q + (il/4) * 32 + 16 * (il&1);
il = il & 3;
const uchar2 sc = get_scale_min_k4_just2(is, il/2, xb->scales);
const float d = il < 2 ? xb->d : xb->d / 16.h;
const float min = xb->dmin;
const float dl = d * sc[0];
const float ml = min * sc[1];
#else
2024-02-18 20:39:58 +01:00
(void) get_scale_min_k4_just2;
q = q + 16 * (il&1);
device const uint8_t * s = xb->scales;
device const half2 * dh = (device const half2 *)xb->d;
const float2 d = (float2)dh[0];
const float dl = il<2 ? d[0] * (s[0]&0xF) : d[0] * (s[1]&0xF)/16.h;
const float ml = il<2 ? d[1] * (s[0]>>4) : d[1] * (s[1]>>4);
#endif
const ushort mask = il<2 ? 0x0F : 0xF0;
for (int i = 0; i < 16; ++i) {
reg[i/4][i%4] = dl * (q[i] & mask) - ml;
}
}
template <typename type4x4>
void dequantize_q5_K(device const block_q5_K *xb, short il, thread type4x4 & reg) {
device const uint8_t * q = xb->qs;
device const uint8_t * qh = xb->qh;
#if QK_K == 256
short is = (il/4) * 2;
q = q + 32 * (il/4) + 16 * (il&1);
qh = qh + 16 * (il&1);
uint8_t ul = 1 << (il/2);
il = il & 3;
const uchar2 sc = get_scale_min_k4_just2(is, il/2, xb->scales);
const float d = il < 2 ? xb->d : xb->d / 16.f;
const float min = xb->dmin;
const float dl = d * sc[0];
const float ml = min * sc[1];
const ushort mask = il<2 ? 0x0F : 0xF0;
const float qh_val = il<2 ? 16.f : 256.f;
for (int i = 0; i < 16; ++i) {
reg[i/4][i%4] = dl * ((q[i] & mask) + (qh[i] & ul ? qh_val : 0)) - ml;
}
#else
q = q + 16 * (il&1);
device const int8_t * s = xb->scales;
const float dl = xb->d * s[il];
uint8_t m = 1<<(il*2);
const float coef = il<2 ? 1.f : 1.f/16.f;
const ushort mask = il<2 ? 0x0F : 0xF0;
for (int i = 0; i < 16; ++i) {
reg[i/4][i%4] = coef * dl * ((q[i] & mask) - (qh[i%8] & (m*(1+i/8)) ? 0.f : 16.f/coef));
}
#endif
}
template <typename type4x4>
void dequantize_q6_K(device const block_q6_K *xb, short il, thread type4x4 & reg) {
const half d_all = xb->d;
device const uint8_t * ql = (device const uint8_t *)xb->ql;
device const uint8_t * qh = (device const uint8_t *)xb->qh;
device const int8_t * scales = (device const int8_t *)xb->scales;
#if QK_K == 256
ql = ql + 64*(il/8) + 32*((il/2)&1) + 16*(il&1);
qh = qh + 32*(il/8) + 16*(il&1);
float sc = scales[(il%2) + 2 * ((il/2))];
il = (il/2) & 3;
#else
ql = ql + 16 * (il&1);
float sc = scales[il];
#endif
const uint16_t kmask1 = il>1 ? (il>2 ? 192 : 48) : (il>0 ? 12 : 3);
const uint16_t kmask2 = il>1 ? 0xF0 : 0x0F;
const float coef = il>1 ? 1.f/16.f : 1.f;
const float ml = d_all * sc * 32.f;
const float dl = d_all * sc * coef;
for (int i = 0; i < 16; ++i) {
const half q = il&1 ? ((ql[i] & kmask2) | ((qh[i] & kmask1) << 2))
: ((ql[i] & kmask2) | ((qh[i] & kmask1) << 4));
reg[i/4][i%4] = dl * q - ml;
}
}
2024-01-08 16:02:32 +01:00
template <typename type4x4>
void dequantize_iq2_xxs(device const block_iq2_xxs * xb, short il, thread type4x4 & reg) {
// il is 0...15 for QK_K = 256 => index of block of 32 is il/2
const float d = xb->d;
const int ib32 = il/2;
il = il%2;
// il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
// each block of 32 needs 2 uint32_t's for the quants & scale, so 4 uint16_t's.
device const uint16_t * q2 = xb->qs + 4*ib32;
const uint32_t aux32_g = q2[0] | (q2[1] << 16);
const uint32_t aux32_s = q2[2] | (q2[3] << 16);
thread const uint8_t * aux8 = (thread const uint8_t *)&aux32_g;
const float dl = d * (0.5f + (aux32_s >> 28)) * 0.25f;
constant uint8_t * grid = (constant uint8_t *)(iq2xxs_grid + aux8[2*il+0]);
2024-01-08 16:02:32 +01:00
uint8_t signs = ksigns_iq2xs[(aux32_s >> 14*il) & 127];
for (int i = 0; i < 8; ++i) {
reg[i/4][i%4] = dl * grid[i] * (signs & kmask_iq2xs[i] ? -1.f : 1.f);
}
grid = (constant uint8_t *)(iq2xxs_grid + aux8[2*il+1]);
2024-01-08 16:02:32 +01:00
signs = ksigns_iq2xs[(aux32_s >> (14*il+7)) & 127];
for (int i = 0; i < 8; ++i) {
reg[2+i/4][i%4] = dl * grid[i] * (signs & kmask_iq2xs[i] ? -1.f : 1.f);
}
}
template <typename type4x4>
void dequantize_iq2_xs(device const block_iq2_xs * xb, short il, thread type4x4 & reg) {
// il is 0...15 for QK_K = 256 => index of block of 32 is il/2
const float d = xb->d;
const int ib32 = il/2;
il = il%2;
// il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
device const uint16_t * q2 = xb->qs + 4*ib32;
const float dl = d * (0.5f + ((xb->scales[ib32] >> 4*il) & 0xf)) * 0.25f;
constant uint8_t * grid = (constant uint8_t *)(iq2xs_grid + (q2[2*il+0] & 511));
uint8_t signs = ksigns_iq2xs[q2[2*il+0] >> 9];
for (int i = 0; i < 8; ++i) {
reg[i/4][i%4] = dl * grid[i] * (signs & kmask_iq2xs[i] ? -1.f : 1.f);
}
grid = (constant uint8_t *)(iq2xs_grid + (q2[2*il+1] & 511));
signs = ksigns_iq2xs[q2[2*il+1] >> 9];
for (int i = 0; i < 8; ++i) {
reg[2+i/4][i%4] = dl * grid[i] * (signs & kmask_iq2xs[i] ? -1.f : 1.f);
}
}
template <typename type4x4>
void dequantize_iq3_xxs(device const block_iq3_xxs * xb, short il, thread type4x4 & reg) {
// il is 0...15 for QK_K = 256 => index of block of 32 is il/2
const float d = xb->d;
const int ib32 = il/2;
il = il%2;
// il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
device const uint8_t * q3 = xb->qs + 8*ib32;
device const uint16_t * gas = (device const uint16_t *)(xb->qs + QK_K/4) + 2*ib32;
const uint32_t aux32 = gas[0] | (gas[1] << 16);
const float dl = d * (0.5f + (aux32 >> 28)) * 0.5f;
constant uint8_t * grid1 = (constant uint8_t *)(iq3xxs_grid + q3[4*il+0]);
constant uint8_t * grid2 = (constant uint8_t *)(iq3xxs_grid + q3[4*il+1]);
uint8_t signs = ksigns_iq2xs[(aux32 >> 14*il) & 127];
for (int i = 0; i < 4; ++i) {
reg[0][i] = dl * grid1[i] * (signs & kmask_iq2xs[i+0] ? -1.f : 1.f);
reg[1][i] = dl * grid2[i] * (signs & kmask_iq2xs[i+4] ? -1.f : 1.f);
}
grid1 = (constant uint8_t *)(iq3xxs_grid + q3[4*il+2]);
grid2 = (constant uint8_t *)(iq3xxs_grid + q3[4*il+3]);
signs = ksigns_iq2xs[(aux32 >> (14*il+7)) & 127];
for (int i = 0; i < 4; ++i) {
reg[2][i] = dl * grid1[i] * (signs & kmask_iq2xs[i+0] ? -1.f : 1.f);
reg[3][i] = dl * grid2[i] * (signs & kmask_iq2xs[i+4] ? -1.f : 1.f);
}
}
IQ3_S: a much better alternative to Q3_K (#5676) * iq4_nl: squash commits for easier rebase * Basics (quantize, dequantize) * CUDA dequantize and dot product * Slightly faster CUDA dot product (120 t/s) * Switch to 6-bit scales * Scalar dot product * AVX2 dot product * ARM_NEON dot product * Works on metal, but still slow * Slightly better Metal dot product * Another small Metal improvement * Metal dot product is getting there * Faster CUDA dot product * Add 1/8 ffn_down layers as Q5_K when no imatrix has been provided * Report the actual bpw * Add _xs mix that is 4.05 bpw for non-MoE models * Remove IQ4_XS for now, slightly adjust kvalues_iq4nl * AVX2 dot product uses Q8_0 instead of Q8_K * Add to test-backend-ops * Minor fix * Also use use Q5_K for attn_output in MoE models * Fixes after merging latest master * Switching to blocks of 32 * AVX2 for blocks of 32 * Scaler dot product for blocks of 32 * ARM_NEON dot product for blocks of 32 * Metal kernels for blocks of 32 * Slightly faster Metal kernels * Resurrecting iq3_xs After all the experimentation, nothing was better than this. * Minor PPL improvement via a block scale fudge factor * Minor improvement via 3 neighbours * iq3_xs: working scalar and AVX2 dot products * iq3_xs: ARM_NEON dot product - works but extremely slow (10 t/s) * iq3_xs: working Metal implementation * Adding IQ3_M - IQ3_XS mix with mostly Q4_K * iiq3_xs: a 3.4375 bpw variant * iq3_xs: make CUDA work for new version * iq3_xs: make scalar and AVX2 work for new version * iq3_s: make ARM_NEON work with new version * iq3_xs: make new version work on metal Performance is very similar to Q3_K_S * iq3_xs: tiny Metal speed improvement * iq3_xs: tiny Metal speed improvement * Fix stupid warning * Q3_K_XS now uses a mix of IQ3_XS and IQ3_XXS * iq3_xs: rename to iq3_s * iq3_s: make tests pass * Move Q3_K_XS mix to 3.25 bpw * Attempt to fix failing tests * Another attempt to fix the Windows builds * Attempt to fix ROCm * ROCm again * iq3_s: partial fix for QK_K = 64 * iq3_s: make it work on metal for QK_K = 64 Pleasent surprise: the coding was super-block size independent, so all it took was to delete some QK_K == 256 guards. * Will this fix ROCm? --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2024-02-24 15:23:52 +01:00
template <typename type4x4>
void dequantize_iq3_s(device const block_iq3_s * xb, short il, thread type4x4 & reg) {
// il is 0...15 for QK_K = 256 => index of block of 32 is il/2
const float d = xb->d;
const int ib32 = il/2;
il = il%2;
// il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
device const uint8_t * qs = xb->qs + 8*ib32;
device const uint8_t * signs = xb->signs + 4*ib32 + 2*il;
const uint8_t qh = xb->qh[ib32] >> 4*il;
const float dl = d * (1 + 2*((xb->scales[ib32/2] >> 4*(ib32%2)) & 0xf));
constant uint8_t * grid1 = (constant uint8_t *)(iq3s_grid + (qs[4*il+0] | ((qh << 8) & 256)));
constant uint8_t * grid2 = (constant uint8_t *)(iq3s_grid + (qs[4*il+1] | ((qh << 7) & 256)));
IQ3_S: a much better alternative to Q3_K (#5676) * iq4_nl: squash commits for easier rebase * Basics (quantize, dequantize) * CUDA dequantize and dot product * Slightly faster CUDA dot product (120 t/s) * Switch to 6-bit scales * Scalar dot product * AVX2 dot product * ARM_NEON dot product * Works on metal, but still slow * Slightly better Metal dot product * Another small Metal improvement * Metal dot product is getting there * Faster CUDA dot product * Add 1/8 ffn_down layers as Q5_K when no imatrix has been provided * Report the actual bpw * Add _xs mix that is 4.05 bpw for non-MoE models * Remove IQ4_XS for now, slightly adjust kvalues_iq4nl * AVX2 dot product uses Q8_0 instead of Q8_K * Add to test-backend-ops * Minor fix * Also use use Q5_K for attn_output in MoE models * Fixes after merging latest master * Switching to blocks of 32 * AVX2 for blocks of 32 * Scaler dot product for blocks of 32 * ARM_NEON dot product for blocks of 32 * Metal kernels for blocks of 32 * Slightly faster Metal kernels * Resurrecting iq3_xs After all the experimentation, nothing was better than this. * Minor PPL improvement via a block scale fudge factor * Minor improvement via 3 neighbours * iq3_xs: working scalar and AVX2 dot products * iq3_xs: ARM_NEON dot product - works but extremely slow (10 t/s) * iq3_xs: working Metal implementation * Adding IQ3_M - IQ3_XS mix with mostly Q4_K * iiq3_xs: a 3.4375 bpw variant * iq3_xs: make CUDA work for new version * iq3_xs: make scalar and AVX2 work for new version * iq3_s: make ARM_NEON work with new version * iq3_xs: make new version work on metal Performance is very similar to Q3_K_S * iq3_xs: tiny Metal speed improvement * iq3_xs: tiny Metal speed improvement * Fix stupid warning * Q3_K_XS now uses a mix of IQ3_XS and IQ3_XXS * iq3_xs: rename to iq3_s * iq3_s: make tests pass * Move Q3_K_XS mix to 3.25 bpw * Attempt to fix failing tests * Another attempt to fix the Windows builds * Attempt to fix ROCm * ROCm again * iq3_s: partial fix for QK_K = 64 * iq3_s: make it work on metal for QK_K = 64 Pleasent surprise: the coding was super-block size independent, so all it took was to delete some QK_K == 256 guards. * Will this fix ROCm? --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2024-02-24 15:23:52 +01:00
for (int i = 0; i < 4; ++i) {
reg[0][i] = dl * grid1[i] * select(1, -1, signs[0] & kmask_iq2xs[i+0]);
reg[1][i] = dl * grid2[i] * select(1, -1, signs[0] & kmask_iq2xs[i+4]);
}
grid1 = (constant uint8_t *)(iq3s_grid + (qs[4*il+2] | ((qh << 6) & 256)));
grid2 = (constant uint8_t *)(iq3s_grid + (qs[4*il+3] | ((qh << 5) & 256)));
IQ3_S: a much better alternative to Q3_K (#5676) * iq4_nl: squash commits for easier rebase * Basics (quantize, dequantize) * CUDA dequantize and dot product * Slightly faster CUDA dot product (120 t/s) * Switch to 6-bit scales * Scalar dot product * AVX2 dot product * ARM_NEON dot product * Works on metal, but still slow * Slightly better Metal dot product * Another small Metal improvement * Metal dot product is getting there * Faster CUDA dot product * Add 1/8 ffn_down layers as Q5_K when no imatrix has been provided * Report the actual bpw * Add _xs mix that is 4.05 bpw for non-MoE models * Remove IQ4_XS for now, slightly adjust kvalues_iq4nl * AVX2 dot product uses Q8_0 instead of Q8_K * Add to test-backend-ops * Minor fix * Also use use Q5_K for attn_output in MoE models * Fixes after merging latest master * Switching to blocks of 32 * AVX2 for blocks of 32 * Scaler dot product for blocks of 32 * ARM_NEON dot product for blocks of 32 * Metal kernels for blocks of 32 * Slightly faster Metal kernels * Resurrecting iq3_xs After all the experimentation, nothing was better than this. * Minor PPL improvement via a block scale fudge factor * Minor improvement via 3 neighbours * iq3_xs: working scalar and AVX2 dot products * iq3_xs: ARM_NEON dot product - works but extremely slow (10 t/s) * iq3_xs: working Metal implementation * Adding IQ3_M - IQ3_XS mix with mostly Q4_K * iiq3_xs: a 3.4375 bpw variant * iq3_xs: make CUDA work for new version * iq3_xs: make scalar and AVX2 work for new version * iq3_s: make ARM_NEON work with new version * iq3_xs: make new version work on metal Performance is very similar to Q3_K_S * iq3_xs: tiny Metal speed improvement * iq3_xs: tiny Metal speed improvement * Fix stupid warning * Q3_K_XS now uses a mix of IQ3_XS and IQ3_XXS * iq3_xs: rename to iq3_s * iq3_s: make tests pass * Move Q3_K_XS mix to 3.25 bpw * Attempt to fix failing tests * Another attempt to fix the Windows builds * Attempt to fix ROCm * ROCm again * iq3_s: partial fix for QK_K = 64 * iq3_s: make it work on metal for QK_K = 64 Pleasent surprise: the coding was super-block size independent, so all it took was to delete some QK_K == 256 guards. * Will this fix ROCm? --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2024-02-24 15:23:52 +01:00
for (int i = 0; i < 4; ++i) {
reg[2][i] = dl * grid1[i] * select(1, -1, signs[1] & kmask_iq2xs[i+0]);
reg[3][i] = dl * grid2[i] * select(1, -1, signs[1] & kmask_iq2xs[i+4]);
}
}
template <typename type4x4>
void dequantize_iq2_s(device const block_iq2_s * xb, short il, thread type4x4 & reg) {
// il is 0...15 for QK_K = 256 => index of block of 32 is il/2
const float d = xb->d;
const int ib32 = il/2;
il = il%2;
// il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
device const uint8_t * qs = xb->qs + 4*ib32 + 2*il;
device const uint8_t * signs = qs + QK_K/8;
const uint8_t qh = xb->qh[ib32] >> 4*il;
const float dl = d * (0.5f + ((xb->scales[ib32] >> 4*il) & 0xf)) * 0.25f;
constant uint8_t * grid1 = (constant uint8_t *)(iq2s_grid + (qs[0] | ((qh << 8) & 0x300)));
constant uint8_t * grid2 = (constant uint8_t *)(iq2s_grid + (qs[1] | ((qh << 6) & 0x300)));
for (int i = 0; i < 8; ++i) {
reg[i/4+0][i%4] = dl * grid1[i] * select(1, -1, signs[0] & kmask_iq2xs[i]);
reg[i/4+2][i%4] = dl * grid2[i] * select(1, -1, signs[1] & kmask_iq2xs[i]);
}
}
template <typename type4x4>
void dequantize_iq1_s(device const block_iq1_s * xb, short il, thread type4x4 & reg) {
// il is 0...15 for QK_K = 256 => index of block of 32 is il/2
const int ib32 = il/2;
il = il%2;
const float d = xb->d;
device const uint8_t * qs = xb->qs + 4*ib32 + 2*il;
device const uint16_t * qh = xb->qh;
const float dl = d * (2*((qh[ib32] >> 12) & 7) + 1);
const float ml = dl * (qh[ib32] & 0x8000 ? -1 - IQ1S_DELTA : -1 + IQ1S_DELTA);
const uint16_t h = qh[ib32] >> 6*il;
constant uint8_t * grid1 = (constant uint8_t *)(iq1s_grid_gpu + (qs[0] | ((h << 8) & 0x700)));
constant uint8_t * grid2 = (constant uint8_t *)(iq1s_grid_gpu + (qs[1] | ((h << 5) & 0x700)));
for (int i = 0; i < 4; ++i) {
reg[0][i] = dl * (grid1[i] & 0xf) + ml;
reg[1][i] = dl * (grid1[i] >> 4) + ml;
reg[2][i] = dl * (grid2[i] & 0xf) + ml;
reg[3][i] = dl * (grid2[i] >> 4) + ml;
}
}
IQ1_M: 1.75 bpw quantization (#6302) * iq1_m: basics * iq1_m: basics-2 * iq1_m: CUDA dequantize works Very 1st shot I get PPL = 9.76 for LLaMA-v2-7B. * iq1_m: separate shifts for each group of 8 in a block We get PPL(LLaMA-v2-7B ) = 9.2810 PPL(LLaMA-v2-13B) = 6.8105 Not bad, but slightly higher than sqrt(PPL(IQ1_S) * PPL(IQ2_XXS)) which is the expected outcome given that IQ1_M is halfway between IQ1_S and IQ2_XXS in terms of bpw. From this, we would expect PPL = 9.14 for LLaMA-v2-7B PPL = 6.63 for LLaMA-v2-13B * iq1_m: go to 3-bit scales There is slight increase in PPL, but the 0.0625 bpw reduction in size is totally worth it. We now have PPL(LLaMA-v2-7B ) = 9.4469 at 1.96 bpw PPL(LLaMA-v2-13B) = 6.8717 at 1.93 bpw PPL(LLaMA-v2-70B) = 4.8568 at 1.85 bpw * iq1_m: scalar dot product * iq1_m: AVX2 dot product * iq1_m: very slightly faster AVX2 dot product * iq1_m: ARM_NEON dot product Works, but very slow (10.5 t/s) * iq1_m: Metal - dequantize works, dot product does not * iq1_m: Metal now works About the same performance as iq1_s. * iq1_m: minor * iq1_m: checking pure iq1_m quantization It is pretty bad: PPL(LLaMA-v2-7B) = 34 if we quantize output.weight with Q4_K. * iiq1_m: slightly faster ARM_NEON dot product 10.5 t/s -> 11.65 t/s * iq1_m: faster ARM_NEON dot product 11.65 t/s -> 14.9 t/s * iq1_m: another minor ARM_NEON dot product improvement 14.9 -> 15.0 t/s * iq1_m: small PPL improvement via super-block scale adjustment After quantizing block scales redo the super-block scale fit. PPL(LLaMA-v2-7B ) = 9.3346 PPL(LLaMA-v2-13B) = 6.8419 PPL(LLaMA-v2-70B) = 4.8294 PPL(Mistral-7B ) = 8.1624 * iq1_m: adapt to CUDA refactoring * iq1_m: remove unused variable We have progressed to warnings being errors. * iq1_m: add to backend-ops tests * iq1_m: fix Windows ARM * iq1_m: use common definition of iq1m_scale_t * cuda: assert -> NO_DEVICE_CODE * iq1_M: PR comments --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2024-03-26 15:21:27 +01:00
template <typename type4x4>
void dequantize_iq1_m(device const block_iq1_m * xb, short il, thread type4x4 & reg) {
// il is 0...15 for QK_K = 256 => index of block of 32 is il/2
const int ib32 = il/2;
il = il%2;
device const uint16_t * sc = (device const uint16_t *)xb->scales;
#if QK_K == 64
const float d = xb->d;
#else
iq1m_scale_t scale;
IQ1_M: 1.75 bpw quantization (#6302) * iq1_m: basics * iq1_m: basics-2 * iq1_m: CUDA dequantize works Very 1st shot I get PPL = 9.76 for LLaMA-v2-7B. * iq1_m: separate shifts for each group of 8 in a block We get PPL(LLaMA-v2-7B ) = 9.2810 PPL(LLaMA-v2-13B) = 6.8105 Not bad, but slightly higher than sqrt(PPL(IQ1_S) * PPL(IQ2_XXS)) which is the expected outcome given that IQ1_M is halfway between IQ1_S and IQ2_XXS in terms of bpw. From this, we would expect PPL = 9.14 for LLaMA-v2-7B PPL = 6.63 for LLaMA-v2-13B * iq1_m: go to 3-bit scales There is slight increase in PPL, but the 0.0625 bpw reduction in size is totally worth it. We now have PPL(LLaMA-v2-7B ) = 9.4469 at 1.96 bpw PPL(LLaMA-v2-13B) = 6.8717 at 1.93 bpw PPL(LLaMA-v2-70B) = 4.8568 at 1.85 bpw * iq1_m: scalar dot product * iq1_m: AVX2 dot product * iq1_m: very slightly faster AVX2 dot product * iq1_m: ARM_NEON dot product Works, but very slow (10.5 t/s) * iq1_m: Metal - dequantize works, dot product does not * iq1_m: Metal now works About the same performance as iq1_s. * iq1_m: minor * iq1_m: checking pure iq1_m quantization It is pretty bad: PPL(LLaMA-v2-7B) = 34 if we quantize output.weight with Q4_K. * iiq1_m: slightly faster ARM_NEON dot product 10.5 t/s -> 11.65 t/s * iq1_m: faster ARM_NEON dot product 11.65 t/s -> 14.9 t/s * iq1_m: another minor ARM_NEON dot product improvement 14.9 -> 15.0 t/s * iq1_m: small PPL improvement via super-block scale adjustment After quantizing block scales redo the super-block scale fit. PPL(LLaMA-v2-7B ) = 9.3346 PPL(LLaMA-v2-13B) = 6.8419 PPL(LLaMA-v2-70B) = 4.8294 PPL(Mistral-7B ) = 8.1624 * iq1_m: adapt to CUDA refactoring * iq1_m: remove unused variable We have progressed to warnings being errors. * iq1_m: add to backend-ops tests * iq1_m: fix Windows ARM * iq1_m: use common definition of iq1m_scale_t * cuda: assert -> NO_DEVICE_CODE * iq1_M: PR comments --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2024-03-26 15:21:27 +01:00
scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000);
const float d = scale.f16;
#endif
IQ1_M: 1.75 bpw quantization (#6302) * iq1_m: basics * iq1_m: basics-2 * iq1_m: CUDA dequantize works Very 1st shot I get PPL = 9.76 for LLaMA-v2-7B. * iq1_m: separate shifts for each group of 8 in a block We get PPL(LLaMA-v2-7B ) = 9.2810 PPL(LLaMA-v2-13B) = 6.8105 Not bad, but slightly higher than sqrt(PPL(IQ1_S) * PPL(IQ2_XXS)) which is the expected outcome given that IQ1_M is halfway between IQ1_S and IQ2_XXS in terms of bpw. From this, we would expect PPL = 9.14 for LLaMA-v2-7B PPL = 6.63 for LLaMA-v2-13B * iq1_m: go to 3-bit scales There is slight increase in PPL, but the 0.0625 bpw reduction in size is totally worth it. We now have PPL(LLaMA-v2-7B ) = 9.4469 at 1.96 bpw PPL(LLaMA-v2-13B) = 6.8717 at 1.93 bpw PPL(LLaMA-v2-70B) = 4.8568 at 1.85 bpw * iq1_m: scalar dot product * iq1_m: AVX2 dot product * iq1_m: very slightly faster AVX2 dot product * iq1_m: ARM_NEON dot product Works, but very slow (10.5 t/s) * iq1_m: Metal - dequantize works, dot product does not * iq1_m: Metal now works About the same performance as iq1_s. * iq1_m: minor * iq1_m: checking pure iq1_m quantization It is pretty bad: PPL(LLaMA-v2-7B) = 34 if we quantize output.weight with Q4_K. * iiq1_m: slightly faster ARM_NEON dot product 10.5 t/s -> 11.65 t/s * iq1_m: faster ARM_NEON dot product 11.65 t/s -> 14.9 t/s * iq1_m: another minor ARM_NEON dot product improvement 14.9 -> 15.0 t/s * iq1_m: small PPL improvement via super-block scale adjustment After quantizing block scales redo the super-block scale fit. PPL(LLaMA-v2-7B ) = 9.3346 PPL(LLaMA-v2-13B) = 6.8419 PPL(LLaMA-v2-70B) = 4.8294 PPL(Mistral-7B ) = 8.1624 * iq1_m: adapt to CUDA refactoring * iq1_m: remove unused variable We have progressed to warnings being errors. * iq1_m: add to backend-ops tests * iq1_m: fix Windows ARM * iq1_m: use common definition of iq1m_scale_t * cuda: assert -> NO_DEVICE_CODE * iq1_M: PR comments --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2024-03-26 15:21:27 +01:00
device const uint8_t * qs = xb->qs + 4*ib32 + 2*il;
device const uint8_t * qh = xb->qh + 2*ib32 + il;
#if QK_K == 64
const float dl = d * (2*((sc[ib32/2] >> (8*(ib32%2)+4*il)) & 0xf) + 1);
#else
IQ1_M: 1.75 bpw quantization (#6302) * iq1_m: basics * iq1_m: basics-2 * iq1_m: CUDA dequantize works Very 1st shot I get PPL = 9.76 for LLaMA-v2-7B. * iq1_m: separate shifts for each group of 8 in a block We get PPL(LLaMA-v2-7B ) = 9.2810 PPL(LLaMA-v2-13B) = 6.8105 Not bad, but slightly higher than sqrt(PPL(IQ1_S) * PPL(IQ2_XXS)) which is the expected outcome given that IQ1_M is halfway between IQ1_S and IQ2_XXS in terms of bpw. From this, we would expect PPL = 9.14 for LLaMA-v2-7B PPL = 6.63 for LLaMA-v2-13B * iq1_m: go to 3-bit scales There is slight increase in PPL, but the 0.0625 bpw reduction in size is totally worth it. We now have PPL(LLaMA-v2-7B ) = 9.4469 at 1.96 bpw PPL(LLaMA-v2-13B) = 6.8717 at 1.93 bpw PPL(LLaMA-v2-70B) = 4.8568 at 1.85 bpw * iq1_m: scalar dot product * iq1_m: AVX2 dot product * iq1_m: very slightly faster AVX2 dot product * iq1_m: ARM_NEON dot product Works, but very slow (10.5 t/s) * iq1_m: Metal - dequantize works, dot product does not * iq1_m: Metal now works About the same performance as iq1_s. * iq1_m: minor * iq1_m: checking pure iq1_m quantization It is pretty bad: PPL(LLaMA-v2-7B) = 34 if we quantize output.weight with Q4_K. * iiq1_m: slightly faster ARM_NEON dot product 10.5 t/s -> 11.65 t/s * iq1_m: faster ARM_NEON dot product 11.65 t/s -> 14.9 t/s * iq1_m: another minor ARM_NEON dot product improvement 14.9 -> 15.0 t/s * iq1_m: small PPL improvement via super-block scale adjustment After quantizing block scales redo the super-block scale fit. PPL(LLaMA-v2-7B ) = 9.3346 PPL(LLaMA-v2-13B) = 6.8419 PPL(LLaMA-v2-70B) = 4.8294 PPL(Mistral-7B ) = 8.1624 * iq1_m: adapt to CUDA refactoring * iq1_m: remove unused variable We have progressed to warnings being errors. * iq1_m: add to backend-ops tests * iq1_m: fix Windows ARM * iq1_m: use common definition of iq1m_scale_t * cuda: assert -> NO_DEVICE_CODE * iq1_M: PR comments --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2024-03-26 15:21:27 +01:00
const float dl = d * (2*((sc[ib32/2] >> (6*(ib32%2)+3*il)) & 7) + 1);
#endif
IQ1_M: 1.75 bpw quantization (#6302) * iq1_m: basics * iq1_m: basics-2 * iq1_m: CUDA dequantize works Very 1st shot I get PPL = 9.76 for LLaMA-v2-7B. * iq1_m: separate shifts for each group of 8 in a block We get PPL(LLaMA-v2-7B ) = 9.2810 PPL(LLaMA-v2-13B) = 6.8105 Not bad, but slightly higher than sqrt(PPL(IQ1_S) * PPL(IQ2_XXS)) which is the expected outcome given that IQ1_M is halfway between IQ1_S and IQ2_XXS in terms of bpw. From this, we would expect PPL = 9.14 for LLaMA-v2-7B PPL = 6.63 for LLaMA-v2-13B * iq1_m: go to 3-bit scales There is slight increase in PPL, but the 0.0625 bpw reduction in size is totally worth it. We now have PPL(LLaMA-v2-7B ) = 9.4469 at 1.96 bpw PPL(LLaMA-v2-13B) = 6.8717 at 1.93 bpw PPL(LLaMA-v2-70B) = 4.8568 at 1.85 bpw * iq1_m: scalar dot product * iq1_m: AVX2 dot product * iq1_m: very slightly faster AVX2 dot product * iq1_m: ARM_NEON dot product Works, but very slow (10.5 t/s) * iq1_m: Metal - dequantize works, dot product does not * iq1_m: Metal now works About the same performance as iq1_s. * iq1_m: minor * iq1_m: checking pure iq1_m quantization It is pretty bad: PPL(LLaMA-v2-7B) = 34 if we quantize output.weight with Q4_K. * iiq1_m: slightly faster ARM_NEON dot product 10.5 t/s -> 11.65 t/s * iq1_m: faster ARM_NEON dot product 11.65 t/s -> 14.9 t/s * iq1_m: another minor ARM_NEON dot product improvement 14.9 -> 15.0 t/s * iq1_m: small PPL improvement via super-block scale adjustment After quantizing block scales redo the super-block scale fit. PPL(LLaMA-v2-7B ) = 9.3346 PPL(LLaMA-v2-13B) = 6.8419 PPL(LLaMA-v2-70B) = 4.8294 PPL(Mistral-7B ) = 8.1624 * iq1_m: adapt to CUDA refactoring * iq1_m: remove unused variable We have progressed to warnings being errors. * iq1_m: add to backend-ops tests * iq1_m: fix Windows ARM * iq1_m: use common definition of iq1m_scale_t * cuda: assert -> NO_DEVICE_CODE * iq1_M: PR comments --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2024-03-26 15:21:27 +01:00
const float ml1 = dl * (qh[0] & 0x08 ? -1 - IQ1M_DELTA : -1 + IQ1M_DELTA);
const float ml2 = dl * (qh[0] & 0x80 ? -1 - IQ1M_DELTA : -1 + IQ1M_DELTA);
constant uint8_t * grid1 = (constant uint8_t *)(iq1s_grid_gpu + (qs[0] | ((qh[0] << 8) & 0x700)));
constant uint8_t * grid2 = (constant uint8_t *)(iq1s_grid_gpu + (qs[1] | ((qh[0] << 4) & 0x700)));
for (int i = 0; i < 4; ++i) {
reg[0][i] = dl * (grid1[i] & 0xf) + ml1;
reg[1][i] = dl * (grid1[i] >> 4) + ml1;
reg[2][i] = dl * (grid2[i] & 0xf) + ml2;
reg[3][i] = dl * (grid2[i] >> 4) + ml2;
}
}
template <typename type4x4>
void dequantize_iq4_nl(device const block_iq4_nl * xb, short il, thread type4x4 & reg) {
device const uint16_t * q4 = (device const uint16_t *)xb->qs;
const float d = xb->d;
uint32_t aux32;
thread const uint8_t * q8 = (thread const uint8_t *)&aux32;
for (int i = 0; i < 4; ++i) {
aux32 = ((q4[2*i] | (q4[2*i+1] << 16)) >> 4*il) & 0x0f0f0f0f;
reg[i][0] = d * kvalues_iq4nl_f[q8[0]];
reg[i][1] = d * kvalues_iq4nl_f[q8[1]];
reg[i][2] = d * kvalues_iq4nl_f[q8[2]];
reg[i][3] = d * kvalues_iq4nl_f[q8[3]];
}
}
template <typename type4x4>
void dequantize_iq4_xs(device const block_iq4_xs * xb, short il, thread type4x4 & reg) {
#if QK_K == 64
dequantize_iq4_nl(xb, il, reg);
#else
// il is 0...15 for QK_K = 256 => index of block of 32 is il/2
const int ib32 = il/2;
il = il%2;
// il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
device const uint32_t * q4 = (device const uint32_t *)xb->qs + 4*ib32;
const int ls = ((xb->scales_l[ib32/2] >> 4*(ib32%2)) & 0xf) | (((xb->scales_h >> 2*ib32) & 3) << 4);
const float d = (float)xb->d * (ls - 32);
uint32_t aux32;
thread const uint8_t * q8 = (thread const uint8_t *)&aux32;
for (int i = 0; i < 4; ++i) {
aux32 = (q4[i] >> 4*il) & 0x0f0f0f0f;
reg[i][0] = d * kvalues_iq4nl_f[q8[0]];
reg[i][1] = d * kvalues_iq4nl_f[q8[1]];
reg[i][2] = d * kvalues_iq4nl_f[q8[2]];
reg[i][3] = d * kvalues_iq4nl_f[q8[3]];
}
#endif
}
template<typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread float4x4 &)>
kernel void kernel_get_rows(
device const void * src0,
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 13:04:25 +01:00
device const char * src1,
device float * dst,
constant int64_t & ne00,
constant uint64_t & nb01,
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 13:04:25 +01:00
constant uint64_t & nb02,
constant int64_t & ne10,
constant uint64_t & nb10,
constant uint64_t & nb11,
constant uint64_t & nb1,
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 13:04:25 +01:00
constant uint64_t & nb2,
uint3 tgpig[[threadgroup_position_in_grid]],
uint tiitg[[thread_index_in_threadgroup]],
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 13:04:25 +01:00
uint3 tptg [[threads_per_threadgroup]]) {
//const int64_t i = tgpig;
//const int64_t r = ((device int32_t *) src1)[i];
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 13:04:25 +01:00
const int64_t i10 = tgpig.x;
const int64_t i11 = tgpig.y;
const int64_t r = ((device int32_t *) ((device char *) src1 + i11*nb11 + i10*nb10))[0];
const int64_t i02 = i11;
for (int64_t ind = tiitg; ind < ne00/16; ind += tptg.x) {
float4x4 temp;
dequantize_func(
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 13:04:25 +01:00
((device const block_q *) ((device char *) src0 + r*nb01 + i02*nb02)) + ind/nl, ind%nl, temp);
*(((device float4x4 *) ((device char *) dst + i11*nb2 + i10*nb1)) + ind) = temp;
}
}
kernel void kernel_get_rows_f32(
device const void * src0,
device const char * src1,
device float * dst,
constant int64_t & ne00,
constant uint64_t & nb01,
constant uint64_t & nb02,
constant int64_t & ne10,
constant uint64_t & nb10,
constant uint64_t & nb11,
constant uint64_t & nb1,
constant uint64_t & nb2,
uint3 tgpig[[threadgroup_position_in_grid]],
uint tiitg[[thread_index_in_threadgroup]],
uint3 tptg [[threads_per_threadgroup]]) {
const int64_t i10 = tgpig.x;
const int64_t i11 = tgpig.y;
const int64_t r = ((device int32_t *) ((device char *) src1 + i11*nb11 + i10*nb10))[0];
const int64_t i02 = i11;
for (int ind = tiitg; ind < ne00; ind += tptg.x) {
((device float *) ((device char *) dst + i11*nb2 + i10*nb1))[ind] =
((device float *) ((device char *) src0 + r*nb01 + i02*nb02))[ind];
}
}
kernel void kernel_get_rows_f16(
device const void * src0,
device const char * src1,
device float * dst,
constant int64_t & ne00,
constant uint64_t & nb01,
constant uint64_t & nb02,
constant int64_t & ne10,
constant uint64_t & nb10,
constant uint64_t & nb11,
constant uint64_t & nb1,
constant uint64_t & nb2,
uint3 tgpig[[threadgroup_position_in_grid]],
uint tiitg[[thread_index_in_threadgroup]],
uint3 tptg [[threads_per_threadgroup]]) {
const int64_t i10 = tgpig.x;
const int64_t i11 = tgpig.y;
const int64_t r = ((device int32_t *) ((device char *) src1 + i11*nb11 + i10*nb10))[0];
const int64_t i02 = i11;
for (int ind = tiitg; ind < ne00; ind += tptg.x) {
((device float *) ((device char *) dst + i11*nb2 + i10*nb1))[ind] =
((device half *) ((device char *) src0 + r*nb01 + i02*nb02))[ind];
}
}
kernel void kernel_get_rows_i32(
device const void * src0,
device const char * src1,
device int32_t * dst,
constant int64_t & ne00,
constant uint64_t & nb01,
constant uint64_t & nb02,
constant int64_t & ne10,
constant uint64_t & nb10,
constant uint64_t & nb11,
constant uint64_t & nb1,
constant uint64_t & nb2,
uint3 tgpig[[threadgroup_position_in_grid]],
uint tiitg[[thread_index_in_threadgroup]],
uint3 tptg [[threads_per_threadgroup]]) {
const int64_t i10 = tgpig.x;
const int64_t i11 = tgpig.y;
const int64_t r = ((device int32_t *) ((device char *) src1 + i11*nb11 + i10*nb10))[0];
const int64_t i02 = i11;
for (int ind = tiitg; ind < ne00; ind += tptg.x) {
((device int32_t *) ((device char *) dst + i11*nb2 + i10*nb1))[ind] =
((device int32_t *) ((device char *) src0 + r*nb01 + i02*nb02))[ind];
}
}
#define BLOCK_SIZE_M 64 // 8 simdgroup matrices from matrix A
#define BLOCK_SIZE_N 32 // 4 simdgroup matrices from matrix B
#define BLOCK_SIZE_K 32
#define THREAD_MAT_M 4 // each thread take 4 simdgroup matrices from matrix A
#define THREAD_MAT_N 2 // each thread take 2 simdgroup matrices from matrix B
#define THREAD_PER_BLOCK 128
#define THREAD_PER_ROW 2 // 2 thread for each row in matrix A to load numbers
#define THREAD_PER_COL 4 // 4 thread for each row in matrix B to load numbers
#define SG_MAT_SIZE 64 // simdgroup matrix is of shape 8x8
#define SG_MAT_ROW 8
// each block_q contains 16*nl weights
template<typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread half4x4 &)>
void kernel_mul_mm_impl(device const uchar * src0,
device const uchar * src1,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne02,
constant uint64_t & nb01,
constant uint64_t & nb02,
constant int64_t & ne12,
constant uint64_t & nb10,
constant uint64_t & nb11,
constant uint64_t & nb12,
constant int64_t & ne0,
constant int64_t & ne1,
constant uint & r2,
constant uint & r3,
threadgroup uchar * shared_memory [[threadgroup(0)]],
uint3 tgpig[[threadgroup_position_in_grid]],
uint tiitg[[thread_index_in_threadgroup]],
uint sgitg[[simdgroup_index_in_threadgroup]]) {
threadgroup half * sa = (threadgroup half *)(shared_memory);
threadgroup float * sb = (threadgroup float *)(shared_memory + 4096);
const uint r0 = tgpig.y;
const uint r1 = tgpig.x;
const uint im = tgpig.z;
// if this block is of 64x32 shape or smaller
short n_rows = (ne0 - r0 * BLOCK_SIZE_M < BLOCK_SIZE_M) ? (ne0 - r0 * BLOCK_SIZE_M) : BLOCK_SIZE_M;
short n_cols = (ne1 - r1 * BLOCK_SIZE_N < BLOCK_SIZE_N) ? (ne1 - r1 * BLOCK_SIZE_N) : BLOCK_SIZE_N;
// a thread shouldn't load data outside of the matrix
short thread_row = ((short)tiitg/THREAD_PER_ROW) < n_rows ? ((short)tiitg/THREAD_PER_ROW) : n_rows - 1;
short thread_col = ((short)tiitg/THREAD_PER_COL) < n_cols ? ((short)tiitg/THREAD_PER_COL) : n_cols - 1;
simdgroup_half8x8 ma[4];
simdgroup_float8x8 mb[2];
simdgroup_float8x8 c_res[8];
for (int i = 0; i < 8; i++){
c_res[i] = make_filled_simdgroup_matrix<float, 8>(0.f);
}
short il = (tiitg % THREAD_PER_ROW);
const uint i12 = im%ne12;
const uint i13 = im/ne12;
uint offset0 = (i12/r2)*nb02 + (i13/r3)*(nb02*ne02);
ushort offset1 = il/nl;
device const block_q * x = (device const block_q *)(src0 + (r0 * BLOCK_SIZE_M + thread_row) * nb01 + offset0) + offset1;
device const float * y = (device const float *)(src1
+ nb12 * im
+ nb11 * (r1 * BLOCK_SIZE_N + thread_col)
+ nb10 * (BLOCK_SIZE_K / THREAD_PER_COL * (tiitg % THREAD_PER_COL)));
for (int loop_k = 0; loop_k < ne00; loop_k += BLOCK_SIZE_K) {
// load data and store to threadgroup memory
half4x4 temp_a;
dequantize_func(x, il, temp_a);
threadgroup_barrier(mem_flags::mem_threadgroup);
#pragma unroll(16)
for (int i = 0; i < 16; i++) {
*(sa + SG_MAT_SIZE * ((tiitg / THREAD_PER_ROW / 8) \
+ (tiitg % THREAD_PER_ROW) * 16 + (i / 8) * 8) \
+ (tiitg / THREAD_PER_ROW) % 8 + (i & 7) * 8) = temp_a[i/4][i%4];
}
*(threadgroup float2x4 *)(sb + (tiitg % THREAD_PER_COL) * 8 * 32 + 8 * (tiitg / THREAD_PER_COL)) = *((device float2x4 *)y);
il = (il + 2 < nl) ? il + 2 : il % 2;
x = (il < 2) ? x + (2+nl-1)/nl : x;
y += BLOCK_SIZE_K;
threadgroup_barrier(mem_flags::mem_threadgroup);
// load matrices from threadgroup memory and conduct outer products
threadgroup half * lsma = (sa + THREAD_MAT_M * SG_MAT_SIZE * (sgitg % 2));
threadgroup float * lsmb = (sb + THREAD_MAT_N * SG_MAT_SIZE * (sgitg / 2));
#pragma unroll(4)
for (int ik = 0; ik < BLOCK_SIZE_K / 8; ik++) {
#pragma unroll(4)
for (int i = 0; i < 4; i++) {
simdgroup_load(ma[i],lsma + SG_MAT_SIZE * i);
}
simdgroup_barrier(mem_flags::mem_none);
#pragma unroll(2)
for (int i = 0; i < 2; i++) {
simdgroup_load(mb[i],lsmb + SG_MAT_SIZE * i);
}
lsma += BLOCK_SIZE_M / SG_MAT_ROW * SG_MAT_SIZE;
lsmb += BLOCK_SIZE_N / SG_MAT_ROW * SG_MAT_SIZE;
#pragma unroll(8)
for (int i = 0; i < 8; i++){
simdgroup_multiply_accumulate(c_res[i], mb[i/4], ma[i%4], c_res[i]);
}
}
}
if ((r0 + 1) * BLOCK_SIZE_M <= ne0 && (r1 + 1) * BLOCK_SIZE_N <= ne1) {
device float * C = dst + (BLOCK_SIZE_M * r0 + 32 * (sgitg & 1)) \
+ (BLOCK_SIZE_N * r1 + 16 * (sgitg >> 1)) * ne0 + im*ne1*ne0;
for (int i = 0; i < 8; i++) {
simdgroup_store(c_res[i], C + 8 * (i%4) + 8 * ne0 * (i/4), ne0);
}
} else {
// block is smaller than 64x32, we should avoid writing data outside of the matrix
threadgroup_barrier(mem_flags::mem_threadgroup);
threadgroup float * temp_str = ((threadgroup float *)shared_memory) \
+ 32 * (sgitg&1) + (16 * (sgitg>>1)) * BLOCK_SIZE_M;
for (int i = 0; i < 8; i++) {
simdgroup_store(c_res[i], temp_str + 8 * (i%4) + 8 * BLOCK_SIZE_M * (i/4), BLOCK_SIZE_M);
}
threadgroup_barrier(mem_flags::mem_threadgroup);
device float * C = dst + (BLOCK_SIZE_M * r0) + (BLOCK_SIZE_N * r1) * ne0 + im*ne1*ne0;
if (sgitg == 0) {
for (int i = 0; i < n_rows; i++) {
for (int j = tiitg; j < n_cols; j += BLOCK_SIZE_N) {
*(C + i + j * ne0) = *(temp_str + i + j * BLOCK_SIZE_M);
}
}
}
}
}
// same as kernel_mul_mm_impl, but src1 and dst are accessed via indices stored in rowids
template<typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread half4x4 &)>
void kernel_mul_mm_id_impl(
device const uchar * src0,
device const uchar * src1,
threadgroup ushort2 * rowids,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne02,
constant uint64_t & nb01,
constant uint64_t & nb02,
constant int64_t & ne11,
constant int64_t & ne12,
constant uint64_t & nb10,
constant uint64_t & nb11,
constant uint64_t & nb12,
constant int64_t & ne0,
int64_t ne1,
int64_t ne0ne1,
threadgroup uchar * shared_memory,
uint3 tgpig[[threadgroup_position_in_grid]],
uint tiitg[[thread_index_in_threadgroup]],
uint sgitg[[simdgroup_index_in_threadgroup]]) {
threadgroup half * sa = (threadgroup half *)(shared_memory);
threadgroup float * sb = (threadgroup float *)(shared_memory + 4096);
const uint r0 = tgpig.y;
const uint r1 = tgpig.x;
if (r1 * BLOCK_SIZE_N >= ne1) return;
// if this block is of 64x32 shape or smaller
short n_rows = (ne0 - r0 * BLOCK_SIZE_M < BLOCK_SIZE_M) ? (ne0 - r0 * BLOCK_SIZE_M) : BLOCK_SIZE_M;
short n_cols = (ne1 - r1 * BLOCK_SIZE_N < BLOCK_SIZE_N) ? (ne1 - r1 * BLOCK_SIZE_N) : BLOCK_SIZE_N;
// a thread shouldn't load data outside of the matrix
short thread_row = ((short)tiitg/THREAD_PER_ROW) < n_rows ? ((short)tiitg/THREAD_PER_ROW) : n_rows - 1;
short thread_col = ((short)tiitg/THREAD_PER_COL) < n_cols ? ((short)tiitg/THREAD_PER_COL) : n_cols - 1;
simdgroup_half8x8 ma[4];
simdgroup_float8x8 mb[2];
simdgroup_float8x8 c_res[8];
for (int i = 0; i < 8; i++){
c_res[i] = make_filled_simdgroup_matrix<float, 8>(0.f);
}
short il = (tiitg % THREAD_PER_ROW);
ushort offset1 = il/nl;
threadgroup const auto & id = rowids[r1 * BLOCK_SIZE_N + thread_col];
device const block_q * x = (device const block_q *)(src0 + (r0 * BLOCK_SIZE_M + thread_row) * nb01) + offset1;
device const float * y = (device const float *)(src1
+ nb12 * id[1]
+ nb11 * (id[0] % ne11)
+ nb10 * (BLOCK_SIZE_K / THREAD_PER_COL * (tiitg % THREAD_PER_COL)));
for (int loop_k = 0; loop_k < ne00; loop_k += BLOCK_SIZE_K) {
// load data and store to threadgroup memory
half4x4 temp_a;
dequantize_func(x, il, temp_a);
threadgroup_barrier(mem_flags::mem_threadgroup);
for (int i = 0; i < 16; i++) {
*(sa + SG_MAT_SIZE * ((tiitg / THREAD_PER_ROW / 8) \
+ (tiitg % THREAD_PER_ROW) * 16 + (i / 8) * 8) \
+ (tiitg / THREAD_PER_ROW) % 8 + (i & 7) * 8) = temp_a[i/4][i%4];
}
*(threadgroup float2x4 *)(sb + (tiitg % THREAD_PER_COL) * 8 * 32 + 8 * (tiitg / THREAD_PER_COL)) = *((device float2x4 *)y);
il = (il + 2 < nl) ? il + 2 : il % 2;
x = (il < 2) ? x + (2+nl-1)/nl : x;
y += BLOCK_SIZE_K;
threadgroup_barrier(mem_flags::mem_threadgroup);
// load matrices from threadgroup memory and conduct outer products
threadgroup half * lsma = (sa + THREAD_MAT_M * SG_MAT_SIZE * (sgitg % 2));
threadgroup float * lsmb = (sb + THREAD_MAT_N * SG_MAT_SIZE * (sgitg / 2));
for (int ik = 0; ik < BLOCK_SIZE_K / 8; ik++) {
for (int i = 0; i < 4; i++) {
simdgroup_load(ma[i], lsma + SG_MAT_SIZE * i);
}
simdgroup_barrier(mem_flags::mem_none);
for (int i = 0; i < 2; i++) {
simdgroup_load(mb[i], lsmb + SG_MAT_SIZE * i);
}
lsma += BLOCK_SIZE_M / SG_MAT_ROW * SG_MAT_SIZE;
lsmb += BLOCK_SIZE_N / SG_MAT_ROW * SG_MAT_SIZE;
for (int i = 0; i < 8; i++){
simdgroup_multiply_accumulate(c_res[i], mb[i/4], ma[i%4], c_res[i]);
}
}
}
{
threadgroup_barrier(mem_flags::mem_threadgroup);
threadgroup float * temp_str = ((threadgroup float *)shared_memory) \
+ 32 * (sgitg&1) + (16 * (sgitg>>1)) * BLOCK_SIZE_M;
for (int i = 0; i < 8; i++) {
simdgroup_store(c_res[i], temp_str + 8 * (i%4) + 8 * BLOCK_SIZE_M * (i/4), BLOCK_SIZE_M);
}
threadgroup_barrier(mem_flags::mem_threadgroup);
device float * C = dst + (BLOCK_SIZE_M * r0);
if (sgitg == 0) {
for (int j = tiitg; j < n_cols; j += BLOCK_SIZE_N) {
threadgroup const auto & jid = rowids[r1 * BLOCK_SIZE_N + j];
int joff = jid[0] * ne0 + jid[1] * ne0ne1;
for (int i = 0; i < n_rows; i++) {
*(C + i + joff) = *(temp_str + i + j * BLOCK_SIZE_M);
}
}
}
}
}
template<typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread half4x4 &)>
kernel void kernel_mul_mm(device const uchar * src0,
device const uchar * src1,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne02,
constant uint64_t & nb01,
constant uint64_t & nb02,
constant int64_t & ne12,
constant uint64_t & nb10,
constant uint64_t & nb11,
constant uint64_t & nb12,
constant int64_t & ne0,
constant int64_t & ne1,
constant uint & r2,
constant uint & r3,
threadgroup uchar * shared_memory [[threadgroup(0)]],
uint3 tgpig[[threadgroup_position_in_grid]],
uint tiitg[[thread_index_in_threadgroup]],
uint sgitg[[simdgroup_index_in_threadgroup]]) {
kernel_mul_mm_impl<block_q, nl, dequantize_func>(
src0,
src1,
dst,
ne00,
ne02,
nb01,
nb02,
ne12,
nb10,
nb11,
nb12,
ne0,
ne1,
r2,
r3,
shared_memory,
tgpig,
tiitg,
sgitg);
}
template<typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread half4x4 &)>
kernel void kernel_mul_mm_id(
ggml : mul_mat_id use the same tensor for all the experts (#6387) * ggml : update mul_mat_id to use the same tensor for all the experts * update cuda * minor * update metal * update test-backend-ops * fix cuda * Update ggml-metal.m Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * update convert.py * update convert-hf-to-gguf.py * update convert.py for mixtral hf models * Update convert-hf-to-gguf.py Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * cuda : support non-pow-2 number of experts * allow quantize to work for split and merged experts models in the same way * cleanup + disable mmap automatically with split tensors models * update imatrix * test-backend-ops : test qwen argsort * update grok model loading * llama : add merged experts tensors to the grok tensor map * minor * gguf : bump version * fix quantizing of merged experts * convert-hf-to-gguf.py : update grok (untested) * make linter happy * cuda/argsort : use shared memory instead of pool memory * convert : fix grok tensor names * metal : add support for non-pow-2 argsort * llama : more loader cleanup, better error checking * cuda : fix warning * llama : still use mmap for loading old models, but copy the data to a host buffer * add review note * llama : remove ffn tensor counting + add sanity check ggml-ci * convert : fix handling of n_experts == None ggml-ci * imatrix : fix ncall counters * llama : produce error if imatrix size does not match * quantize : terminate on errors + trace logs ggml-ci * metal : pad shared memory to 16 bytes --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-04-03 15:07:05 +02:00
device const uchar * src0s,
device const uchar * src1,
device float * dst,
ggml : mul_mat_id use the same tensor for all the experts (#6387) * ggml : update mul_mat_id to use the same tensor for all the experts * update cuda * minor * update metal * update test-backend-ops * fix cuda * Update ggml-metal.m Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * update convert.py * update convert-hf-to-gguf.py * update convert.py for mixtral hf models * Update convert-hf-to-gguf.py Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * cuda : support non-pow-2 number of experts * allow quantize to work for split and merged experts models in the same way * cleanup + disable mmap automatically with split tensors models * update imatrix * test-backend-ops : test qwen argsort * update grok model loading * llama : add merged experts tensors to the grok tensor map * minor * gguf : bump version * fix quantizing of merged experts * convert-hf-to-gguf.py : update grok (untested) * make linter happy * cuda/argsort : use shared memory instead of pool memory * convert : fix grok tensor names * metal : add support for non-pow-2 argsort * llama : more loader cleanup, better error checking * cuda : fix warning * llama : still use mmap for loading old models, but copy the data to a host buffer * add review note * llama : remove ffn tensor counting + add sanity check ggml-ci * convert : fix handling of n_experts == None ggml-ci * imatrix : fix ncall counters * llama : produce error if imatrix size does not match * quantize : terminate on errors + trace logs ggml-ci * metal : pad shared memory to 16 bytes --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-04-03 15:07:05 +02:00
device const uchar * ids,
constant int64_t & nei0,
constant int64_t & nei1,
constant uint64_t & nbi1,
constant int64_t & ne00,
constant int64_t & ne02,
constant uint64_t & nb01,
constant uint64_t & nb02,
constant int64_t & ne11,
constant int64_t & ne12,
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 13:04:25 +01:00
constant int64_t & ne13,
constant uint64_t & nb10,
constant uint64_t & nb11,
constant uint64_t & nb12,
constant int64_t & ne0,
constant int64_t & ne1,
constant uint64_t & nb1,
threadgroup uchar * shared_memory [[threadgroup(0)]],
uint3 tgpig[[threadgroup_position_in_grid]],
uint tiitg[[thread_index_in_threadgroup]],
uint sgitg[[simdgroup_index_in_threadgroup]]) {
const int32_t i02 = tgpig.z;
tgpig.z = 0;
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 13:04:25 +01:00
device const uchar * src0 = src0s + i02*nb02;
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 13:04:25 +01:00
// row indices
threadgroup ushort2 * rowids = (threadgroup ushort2 *)(shared_memory + 8192);
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 13:04:25 +01:00
// TODO: parallelize this loop
int64_t _ne1 = 0;
for (ushort ii1 = 0; ii1 < nei1; ii1++) {
for (ushort ii0 = 0; ii0 < nei0; ii0++) {
int32_t id = ((device int32_t *) (ids + ii1*nbi1))[ii0];
if (id == i02) {
//if (tiitg == 0) {
rowids[_ne1] = ushort2(ii0, ii1);
//}
_ne1++;
}
}
}
threadgroup_barrier(mem_flags::mem_threadgroup);
kernel_mul_mm_id_impl<block_q, nl, dequantize_func>(
ggml : mul_mat_id use the same tensor for all the experts (#6387) * ggml : update mul_mat_id to use the same tensor for all the experts * update cuda * minor * update metal * update test-backend-ops * fix cuda * Update ggml-metal.m Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * update convert.py * update convert-hf-to-gguf.py * update convert.py for mixtral hf models * Update convert-hf-to-gguf.py Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * cuda : support non-pow-2 number of experts * allow quantize to work for split and merged experts models in the same way * cleanup + disable mmap automatically with split tensors models * update imatrix * test-backend-ops : test qwen argsort * update grok model loading * llama : add merged experts tensors to the grok tensor map * minor * gguf : bump version * fix quantizing of merged experts * convert-hf-to-gguf.py : update grok (untested) * make linter happy * cuda/argsort : use shared memory instead of pool memory * convert : fix grok tensor names * metal : add support for non-pow-2 argsort * llama : more loader cleanup, better error checking * cuda : fix warning * llama : still use mmap for loading old models, but copy the data to a host buffer * add review note * llama : remove ffn tensor counting + add sanity check ggml-ci * convert : fix handling of n_experts == None ggml-ci * imatrix : fix ncall counters * llama : produce error if imatrix size does not match * quantize : terminate on errors + trace logs ggml-ci * metal : pad shared memory to 16 bytes --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-04-03 15:07:05 +02:00
src0,
src1,
rowids,
dst,
ne00,
ne02,
nb01,
nb02,
ne11,
ne12,
nb10,
nb11,
nb12,
ne0,
_ne1,
ne0*ne1,
shared_memory,
tgpig,
tiitg,
sgitg);
}
#if QK_K == 256
#define QK_NL 16
#else
#define QK_NL 4
#endif
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 13:04:25 +01:00
//
// get rows
//
typedef void (get_rows_t)(
device const void * src0,
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 13:04:25 +01:00
device const char * src1,
device float * dst,
constant int64_t & ne00,
constant uint64_t & nb01,
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 13:04:25 +01:00
constant uint64_t & nb02,
constant int64_t & ne10,
constant uint64_t & nb10,
constant uint64_t & nb11,
constant uint64_t & nb1,
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 13:04:25 +01:00
constant uint64_t & nb2,
uint3, uint, uint3);
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 13:04:25 +01:00
//template [[host_name("kernel_get_rows_f32")]] kernel get_rows_t kernel_get_rows<float4x4, 1, dequantize_f32>;
//template [[host_name("kernel_get_rows_f16")]] kernel get_rows_t kernel_get_rows<half4x4, 1, dequantize_f16>;
template [[host_name("kernel_get_rows_q4_0")]] kernel get_rows_t kernel_get_rows<block_q4_0, 2, dequantize_q4_0>;
template [[host_name("kernel_get_rows_q4_1")]] kernel get_rows_t kernel_get_rows<block_q4_1, 2, dequantize_q4_1>;
template [[host_name("kernel_get_rows_q5_0")]] kernel get_rows_t kernel_get_rows<block_q5_0, 2, dequantize_q5_0>;
template [[host_name("kernel_get_rows_q5_1")]] kernel get_rows_t kernel_get_rows<block_q5_1, 2, dequantize_q5_1>;
template [[host_name("kernel_get_rows_q8_0")]] kernel get_rows_t kernel_get_rows<block_q8_0, 2, dequantize_q8_0>;
template [[host_name("kernel_get_rows_q2_K")]] kernel get_rows_t kernel_get_rows<block_q2_K, QK_NL, dequantize_q2_K>;
template [[host_name("kernel_get_rows_q3_K")]] kernel get_rows_t kernel_get_rows<block_q3_K, QK_NL, dequantize_q3_K>;
template [[host_name("kernel_get_rows_q4_K")]] kernel get_rows_t kernel_get_rows<block_q4_K, QK_NL, dequantize_q4_K>;
template [[host_name("kernel_get_rows_q5_K")]] kernel get_rows_t kernel_get_rows<block_q5_K, QK_NL, dequantize_q5_K>;
template [[host_name("kernel_get_rows_q6_K")]] kernel get_rows_t kernel_get_rows<block_q6_K, QK_NL, dequantize_q6_K>;
2024-01-08 16:02:32 +01:00
template [[host_name("kernel_get_rows_iq2_xxs")]] kernel get_rows_t kernel_get_rows<block_iq2_xxs, QK_NL, dequantize_iq2_xxs>;
template [[host_name("kernel_get_rows_iq2_xs")]] kernel get_rows_t kernel_get_rows<block_iq2_xs, QK_NL, dequantize_iq2_xs>;
template [[host_name("kernel_get_rows_iq3_xxs")]] kernel get_rows_t kernel_get_rows<block_iq3_xxs, QK_NL, dequantize_iq3_xxs>;
IQ3_S: a much better alternative to Q3_K (#5676) * iq4_nl: squash commits for easier rebase * Basics (quantize, dequantize) * CUDA dequantize and dot product * Slightly faster CUDA dot product (120 t/s) * Switch to 6-bit scales * Scalar dot product * AVX2 dot product * ARM_NEON dot product * Works on metal, but still slow * Slightly better Metal dot product * Another small Metal improvement * Metal dot product is getting there * Faster CUDA dot product * Add 1/8 ffn_down layers as Q5_K when no imatrix has been provided * Report the actual bpw * Add _xs mix that is 4.05 bpw for non-MoE models * Remove IQ4_XS for now, slightly adjust kvalues_iq4nl * AVX2 dot product uses Q8_0 instead of Q8_K * Add to test-backend-ops * Minor fix * Also use use Q5_K for attn_output in MoE models * Fixes after merging latest master * Switching to blocks of 32 * AVX2 for blocks of 32 * Scaler dot product for blocks of 32 * ARM_NEON dot product for blocks of 32 * Metal kernels for blocks of 32 * Slightly faster Metal kernels * Resurrecting iq3_xs After all the experimentation, nothing was better than this. * Minor PPL improvement via a block scale fudge factor * Minor improvement via 3 neighbours * iq3_xs: working scalar and AVX2 dot products * iq3_xs: ARM_NEON dot product - works but extremely slow (10 t/s) * iq3_xs: working Metal implementation * Adding IQ3_M - IQ3_XS mix with mostly Q4_K * iiq3_xs: a 3.4375 bpw variant * iq3_xs: make CUDA work for new version * iq3_xs: make scalar and AVX2 work for new version * iq3_s: make ARM_NEON work with new version * iq3_xs: make new version work on metal Performance is very similar to Q3_K_S * iq3_xs: tiny Metal speed improvement * iq3_xs: tiny Metal speed improvement * Fix stupid warning * Q3_K_XS now uses a mix of IQ3_XS and IQ3_XXS * iq3_xs: rename to iq3_s * iq3_s: make tests pass * Move Q3_K_XS mix to 3.25 bpw * Attempt to fix failing tests * Another attempt to fix the Windows builds * Attempt to fix ROCm * ROCm again * iq3_s: partial fix for QK_K = 64 * iq3_s: make it work on metal for QK_K = 64 Pleasent surprise: the coding was super-block size independent, so all it took was to delete some QK_K == 256 guards. * Will this fix ROCm? --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2024-02-24 15:23:52 +01:00
template [[host_name("kernel_get_rows_iq3_s")]] kernel get_rows_t kernel_get_rows<block_iq3_s, QK_NL, dequantize_iq3_s>;
template [[host_name("kernel_get_rows_iq2_s")]] kernel get_rows_t kernel_get_rows<block_iq2_s, QK_NL, dequantize_iq2_s>;
template [[host_name("kernel_get_rows_iq1_s")]] kernel get_rows_t kernel_get_rows<block_iq1_s, QK_NL, dequantize_iq1_s>;
IQ1_M: 1.75 bpw quantization (#6302) * iq1_m: basics * iq1_m: basics-2 * iq1_m: CUDA dequantize works Very 1st shot I get PPL = 9.76 for LLaMA-v2-7B. * iq1_m: separate shifts for each group of 8 in a block We get PPL(LLaMA-v2-7B ) = 9.2810 PPL(LLaMA-v2-13B) = 6.8105 Not bad, but slightly higher than sqrt(PPL(IQ1_S) * PPL(IQ2_XXS)) which is the expected outcome given that IQ1_M is halfway between IQ1_S and IQ2_XXS in terms of bpw. From this, we would expect PPL = 9.14 for LLaMA-v2-7B PPL = 6.63 for LLaMA-v2-13B * iq1_m: go to 3-bit scales There is slight increase in PPL, but the 0.0625 bpw reduction in size is totally worth it. We now have PPL(LLaMA-v2-7B ) = 9.4469 at 1.96 bpw PPL(LLaMA-v2-13B) = 6.8717 at 1.93 bpw PPL(LLaMA-v2-70B) = 4.8568 at 1.85 bpw * iq1_m: scalar dot product * iq1_m: AVX2 dot product * iq1_m: very slightly faster AVX2 dot product * iq1_m: ARM_NEON dot product Works, but very slow (10.5 t/s) * iq1_m: Metal - dequantize works, dot product does not * iq1_m: Metal now works About the same performance as iq1_s. * iq1_m: minor * iq1_m: checking pure iq1_m quantization It is pretty bad: PPL(LLaMA-v2-7B) = 34 if we quantize output.weight with Q4_K. * iiq1_m: slightly faster ARM_NEON dot product 10.5 t/s -> 11.65 t/s * iq1_m: faster ARM_NEON dot product 11.65 t/s -> 14.9 t/s * iq1_m: another minor ARM_NEON dot product improvement 14.9 -> 15.0 t/s * iq1_m: small PPL improvement via super-block scale adjustment After quantizing block scales redo the super-block scale fit. PPL(LLaMA-v2-7B ) = 9.3346 PPL(LLaMA-v2-13B) = 6.8419 PPL(LLaMA-v2-70B) = 4.8294 PPL(Mistral-7B ) = 8.1624 * iq1_m: adapt to CUDA refactoring * iq1_m: remove unused variable We have progressed to warnings being errors. * iq1_m: add to backend-ops tests * iq1_m: fix Windows ARM * iq1_m: use common definition of iq1m_scale_t * cuda: assert -> NO_DEVICE_CODE * iq1_M: PR comments --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2024-03-26 15:21:27 +01:00
template [[host_name("kernel_get_rows_iq1_m")]] kernel get_rows_t kernel_get_rows<block_iq1_m, QK_NL, dequantize_iq1_m>;
template [[host_name("kernel_get_rows_iq4_nl")]] kernel get_rows_t kernel_get_rows<block_iq4_nl, 2, dequantize_iq4_nl>;
#if QK_K == 64
template [[host_name("kernel_get_rows_iq4_xs")]] kernel get_rows_t kernel_get_rows<block_iq4_xs, 2, dequantize_iq4_xs>;
#else
template [[host_name("kernel_get_rows_iq4_xs")]] kernel get_rows_t kernel_get_rows<block_iq4_xs, QK_NL, dequantize_iq4_xs>;
#endif
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 13:04:25 +01:00
//
// matrix-matrix multiplication
//
typedef decltype(kernel_mul_mm<float4x4, 1, dequantize_f32>) mat_mm_t;
template [[host_name("kernel_mul_mm_f32_f32")]] kernel mat_mm_t kernel_mul_mm<float4x4, 1, dequantize_f32>;
template [[host_name("kernel_mul_mm_f16_f32")]] kernel mat_mm_t kernel_mul_mm<half4x4, 1, dequantize_f16>;
template [[host_name("kernel_mul_mm_q4_0_f32")]] kernel mat_mm_t kernel_mul_mm<block_q4_0, 2, dequantize_q4_0>;
template [[host_name("kernel_mul_mm_q4_1_f32")]] kernel mat_mm_t kernel_mul_mm<block_q4_1, 2, dequantize_q4_1>;
template [[host_name("kernel_mul_mm_q5_0_f32")]] kernel mat_mm_t kernel_mul_mm<block_q5_0, 2, dequantize_q5_0>;
template [[host_name("kernel_mul_mm_q5_1_f32")]] kernel mat_mm_t kernel_mul_mm<block_q5_1, 2, dequantize_q5_1>;
template [[host_name("kernel_mul_mm_q8_0_f32")]] kernel mat_mm_t kernel_mul_mm<block_q8_0, 2, dequantize_q8_0>;
template [[host_name("kernel_mul_mm_q2_K_f32")]] kernel mat_mm_t kernel_mul_mm<block_q2_K, QK_NL, dequantize_q2_K>;
template [[host_name("kernel_mul_mm_q3_K_f32")]] kernel mat_mm_t kernel_mul_mm<block_q3_K, QK_NL, dequantize_q3_K>;
template [[host_name("kernel_mul_mm_q4_K_f32")]] kernel mat_mm_t kernel_mul_mm<block_q4_K, QK_NL, dequantize_q4_K>;
template [[host_name("kernel_mul_mm_q5_K_f32")]] kernel mat_mm_t kernel_mul_mm<block_q5_K, QK_NL, dequantize_q5_K>;
template [[host_name("kernel_mul_mm_q6_K_f32")]] kernel mat_mm_t kernel_mul_mm<block_q6_K, QK_NL, dequantize_q6_K>;
2024-01-08 16:02:32 +01:00
template [[host_name("kernel_mul_mm_iq2_xxs_f32")]] kernel mat_mm_t kernel_mul_mm<block_iq2_xxs, QK_NL, dequantize_iq2_xxs>;
template [[host_name("kernel_mul_mm_iq2_xs_f32")]] kernel mat_mm_t kernel_mul_mm<block_iq2_xs, QK_NL, dequantize_iq2_xs>;
template [[host_name("kernel_mul_mm_iq3_xxs_f32")]] kernel mat_mm_t kernel_mul_mm<block_iq3_xxs, QK_NL, dequantize_iq3_xxs>;
IQ3_S: a much better alternative to Q3_K (#5676) * iq4_nl: squash commits for easier rebase * Basics (quantize, dequantize) * CUDA dequantize and dot product * Slightly faster CUDA dot product (120 t/s) * Switch to 6-bit scales * Scalar dot product * AVX2 dot product * ARM_NEON dot product * Works on metal, but still slow * Slightly better Metal dot product * Another small Metal improvement * Metal dot product is getting there * Faster CUDA dot product * Add 1/8 ffn_down layers as Q5_K when no imatrix has been provided * Report the actual bpw * Add _xs mix that is 4.05 bpw for non-MoE models * Remove IQ4_XS for now, slightly adjust kvalues_iq4nl * AVX2 dot product uses Q8_0 instead of Q8_K * Add to test-backend-ops * Minor fix * Also use use Q5_K for attn_output in MoE models * Fixes after merging latest master * Switching to blocks of 32 * AVX2 for blocks of 32 * Scaler dot product for blocks of 32 * ARM_NEON dot product for blocks of 32 * Metal kernels for blocks of 32 * Slightly faster Metal kernels * Resurrecting iq3_xs After all the experimentation, nothing was better than this. * Minor PPL improvement via a block scale fudge factor * Minor improvement via 3 neighbours * iq3_xs: working scalar and AVX2 dot products * iq3_xs: ARM_NEON dot product - works but extremely slow (10 t/s) * iq3_xs: working Metal implementation * Adding IQ3_M - IQ3_XS mix with mostly Q4_K * iiq3_xs: a 3.4375 bpw variant * iq3_xs: make CUDA work for new version * iq3_xs: make scalar and AVX2 work for new version * iq3_s: make ARM_NEON work with new version * iq3_xs: make new version work on metal Performance is very similar to Q3_K_S * iq3_xs: tiny Metal speed improvement * iq3_xs: tiny Metal speed improvement * Fix stupid warning * Q3_K_XS now uses a mix of IQ3_XS and IQ3_XXS * iq3_xs: rename to iq3_s * iq3_s: make tests pass * Move Q3_K_XS mix to 3.25 bpw * Attempt to fix failing tests * Another attempt to fix the Windows builds * Attempt to fix ROCm * ROCm again * iq3_s: partial fix for QK_K = 64 * iq3_s: make it work on metal for QK_K = 64 Pleasent surprise: the coding was super-block size independent, so all it took was to delete some QK_K == 256 guards. * Will this fix ROCm? --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2024-02-24 15:23:52 +01:00
template [[host_name("kernel_mul_mm_iq3_s_f32")]] kernel mat_mm_t kernel_mul_mm<block_iq3_s, QK_NL, dequantize_iq3_s>;
template [[host_name("kernel_mul_mm_iq2_s_f32")]] kernel mat_mm_t kernel_mul_mm<block_iq2_s, QK_NL, dequantize_iq2_s>;
template [[host_name("kernel_mul_mm_iq1_s_f32")]] kernel mat_mm_t kernel_mul_mm<block_iq1_s, QK_NL, dequantize_iq1_s>;
IQ1_M: 1.75 bpw quantization (#6302) * iq1_m: basics * iq1_m: basics-2 * iq1_m: CUDA dequantize works Very 1st shot I get PPL = 9.76 for LLaMA-v2-7B. * iq1_m: separate shifts for each group of 8 in a block We get PPL(LLaMA-v2-7B ) = 9.2810 PPL(LLaMA-v2-13B) = 6.8105 Not bad, but slightly higher than sqrt(PPL(IQ1_S) * PPL(IQ2_XXS)) which is the expected outcome given that IQ1_M is halfway between IQ1_S and IQ2_XXS in terms of bpw. From this, we would expect PPL = 9.14 for LLaMA-v2-7B PPL = 6.63 for LLaMA-v2-13B * iq1_m: go to 3-bit scales There is slight increase in PPL, but the 0.0625 bpw reduction in size is totally worth it. We now have PPL(LLaMA-v2-7B ) = 9.4469 at 1.96 bpw PPL(LLaMA-v2-13B) = 6.8717 at 1.93 bpw PPL(LLaMA-v2-70B) = 4.8568 at 1.85 bpw * iq1_m: scalar dot product * iq1_m: AVX2 dot product * iq1_m: very slightly faster AVX2 dot product * iq1_m: ARM_NEON dot product Works, but very slow (10.5 t/s) * iq1_m: Metal - dequantize works, dot product does not * iq1_m: Metal now works About the same performance as iq1_s. * iq1_m: minor * iq1_m: checking pure iq1_m quantization It is pretty bad: PPL(LLaMA-v2-7B) = 34 if we quantize output.weight with Q4_K. * iiq1_m: slightly faster ARM_NEON dot product 10.5 t/s -> 11.65 t/s * iq1_m: faster ARM_NEON dot product 11.65 t/s -> 14.9 t/s * iq1_m: another minor ARM_NEON dot product improvement 14.9 -> 15.0 t/s * iq1_m: small PPL improvement via super-block scale adjustment After quantizing block scales redo the super-block scale fit. PPL(LLaMA-v2-7B ) = 9.3346 PPL(LLaMA-v2-13B) = 6.8419 PPL(LLaMA-v2-70B) = 4.8294 PPL(Mistral-7B ) = 8.1624 * iq1_m: adapt to CUDA refactoring * iq1_m: remove unused variable We have progressed to warnings being errors. * iq1_m: add to backend-ops tests * iq1_m: fix Windows ARM * iq1_m: use common definition of iq1m_scale_t * cuda: assert -> NO_DEVICE_CODE * iq1_M: PR comments --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2024-03-26 15:21:27 +01:00
template [[host_name("kernel_mul_mm_iq1_m_f32")]] kernel mat_mm_t kernel_mul_mm<block_iq1_m, QK_NL, dequantize_iq1_m>;
template [[host_name("kernel_mul_mm_iq4_nl_f32")]] kernel mat_mm_t kernel_mul_mm<block_iq4_nl, 2, dequantize_iq4_nl>;
#if QK_K == 64
template [[host_name("kernel_mul_mm_iq4_xs_f32")]] kernel mat_mm_t kernel_mul_mm<block_iq4_nl, 2, dequantize_iq4_xs>;
#else
template [[host_name("kernel_mul_mm_iq4_xs_f32")]] kernel mat_mm_t kernel_mul_mm<block_iq4_xs, QK_NL, dequantize_iq4_xs>;
#endif
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 13:04:25 +01:00
//
// indirect matrix-matrix multiplication
//
typedef decltype(kernel_mul_mm_id<float4x4, 1, dequantize_f32>) mat_mm_id_t;
template [[host_name("kernel_mul_mm_id_f32_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<float4x4, 1, dequantize_f32>;
template [[host_name("kernel_mul_mm_id_f16_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<half4x4, 1, dequantize_f16>;
template [[host_name("kernel_mul_mm_id_q4_0_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_q4_0, 2, dequantize_q4_0>;
template [[host_name("kernel_mul_mm_id_q4_1_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_q4_1, 2, dequantize_q4_1>;
template [[host_name("kernel_mul_mm_id_q5_0_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_q5_0, 2, dequantize_q5_0>;
template [[host_name("kernel_mul_mm_id_q5_1_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_q5_1, 2, dequantize_q5_1>;
template [[host_name("kernel_mul_mm_id_q8_0_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_q8_0, 2, dequantize_q8_0>;
template [[host_name("kernel_mul_mm_id_q2_K_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_q2_K, QK_NL, dequantize_q2_K>;
template [[host_name("kernel_mul_mm_id_q3_K_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_q3_K, QK_NL, dequantize_q3_K>;
template [[host_name("kernel_mul_mm_id_q4_K_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_q4_K, QK_NL, dequantize_q4_K>;
template [[host_name("kernel_mul_mm_id_q5_K_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_q5_K, QK_NL, dequantize_q5_K>;
template [[host_name("kernel_mul_mm_id_q6_K_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_q6_K, QK_NL, dequantize_q6_K>;
2024-01-08 16:02:32 +01:00
template [[host_name("kernel_mul_mm_id_iq2_xxs_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_iq2_xxs, QK_NL, dequantize_iq2_xxs>;
template [[host_name("kernel_mul_mm_id_iq2_xs_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_iq2_xs, QK_NL, dequantize_iq2_xs>;
template [[host_name("kernel_mul_mm_id_iq3_xxs_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_iq3_xxs, QK_NL, dequantize_iq3_xxs>;
IQ3_S: a much better alternative to Q3_K (#5676) * iq4_nl: squash commits for easier rebase * Basics (quantize, dequantize) * CUDA dequantize and dot product * Slightly faster CUDA dot product (120 t/s) * Switch to 6-bit scales * Scalar dot product * AVX2 dot product * ARM_NEON dot product * Works on metal, but still slow * Slightly better Metal dot product * Another small Metal improvement * Metal dot product is getting there * Faster CUDA dot product * Add 1/8 ffn_down layers as Q5_K when no imatrix has been provided * Report the actual bpw * Add _xs mix that is 4.05 bpw for non-MoE models * Remove IQ4_XS for now, slightly adjust kvalues_iq4nl * AVX2 dot product uses Q8_0 instead of Q8_K * Add to test-backend-ops * Minor fix * Also use use Q5_K for attn_output in MoE models * Fixes after merging latest master * Switching to blocks of 32 * AVX2 for blocks of 32 * Scaler dot product for blocks of 32 * ARM_NEON dot product for blocks of 32 * Metal kernels for blocks of 32 * Slightly faster Metal kernels * Resurrecting iq3_xs After all the experimentation, nothing was better than this. * Minor PPL improvement via a block scale fudge factor * Minor improvement via 3 neighbours * iq3_xs: working scalar and AVX2 dot products * iq3_xs: ARM_NEON dot product - works but extremely slow (10 t/s) * iq3_xs: working Metal implementation * Adding IQ3_M - IQ3_XS mix with mostly Q4_K * iiq3_xs: a 3.4375 bpw variant * iq3_xs: make CUDA work for new version * iq3_xs: make scalar and AVX2 work for new version * iq3_s: make ARM_NEON work with new version * iq3_xs: make new version work on metal Performance is very similar to Q3_K_S * iq3_xs: tiny Metal speed improvement * iq3_xs: tiny Metal speed improvement * Fix stupid warning * Q3_K_XS now uses a mix of IQ3_XS and IQ3_XXS * iq3_xs: rename to iq3_s * iq3_s: make tests pass * Move Q3_K_XS mix to 3.25 bpw * Attempt to fix failing tests * Another attempt to fix the Windows builds * Attempt to fix ROCm * ROCm again * iq3_s: partial fix for QK_K = 64 * iq3_s: make it work on metal for QK_K = 64 Pleasent surprise: the coding was super-block size independent, so all it took was to delete some QK_K == 256 guards. * Will this fix ROCm? --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2024-02-24 15:23:52 +01:00
template [[host_name("kernel_mul_mm_id_iq3_s_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_iq3_s, QK_NL, dequantize_iq3_s>;
template [[host_name("kernel_mul_mm_id_iq2_s_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_iq2_s, QK_NL, dequantize_iq2_s>;
template [[host_name("kernel_mul_mm_id_iq1_s_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_iq1_s, QK_NL, dequantize_iq1_s>;
IQ1_M: 1.75 bpw quantization (#6302) * iq1_m: basics * iq1_m: basics-2 * iq1_m: CUDA dequantize works Very 1st shot I get PPL = 9.76 for LLaMA-v2-7B. * iq1_m: separate shifts for each group of 8 in a block We get PPL(LLaMA-v2-7B ) = 9.2810 PPL(LLaMA-v2-13B) = 6.8105 Not bad, but slightly higher than sqrt(PPL(IQ1_S) * PPL(IQ2_XXS)) which is the expected outcome given that IQ1_M is halfway between IQ1_S and IQ2_XXS in terms of bpw. From this, we would expect PPL = 9.14 for LLaMA-v2-7B PPL = 6.63 for LLaMA-v2-13B * iq1_m: go to 3-bit scales There is slight increase in PPL, but the 0.0625 bpw reduction in size is totally worth it. We now have PPL(LLaMA-v2-7B ) = 9.4469 at 1.96 bpw PPL(LLaMA-v2-13B) = 6.8717 at 1.93 bpw PPL(LLaMA-v2-70B) = 4.8568 at 1.85 bpw * iq1_m: scalar dot product * iq1_m: AVX2 dot product * iq1_m: very slightly faster AVX2 dot product * iq1_m: ARM_NEON dot product Works, but very slow (10.5 t/s) * iq1_m: Metal - dequantize works, dot product does not * iq1_m: Metal now works About the same performance as iq1_s. * iq1_m: minor * iq1_m: checking pure iq1_m quantization It is pretty bad: PPL(LLaMA-v2-7B) = 34 if we quantize output.weight with Q4_K. * iiq1_m: slightly faster ARM_NEON dot product 10.5 t/s -> 11.65 t/s * iq1_m: faster ARM_NEON dot product 11.65 t/s -> 14.9 t/s * iq1_m: another minor ARM_NEON dot product improvement 14.9 -> 15.0 t/s * iq1_m: small PPL improvement via super-block scale adjustment After quantizing block scales redo the super-block scale fit. PPL(LLaMA-v2-7B ) = 9.3346 PPL(LLaMA-v2-13B) = 6.8419 PPL(LLaMA-v2-70B) = 4.8294 PPL(Mistral-7B ) = 8.1624 * iq1_m: adapt to CUDA refactoring * iq1_m: remove unused variable We have progressed to warnings being errors. * iq1_m: add to backend-ops tests * iq1_m: fix Windows ARM * iq1_m: use common definition of iq1m_scale_t * cuda: assert -> NO_DEVICE_CODE * iq1_M: PR comments --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2024-03-26 15:21:27 +01:00
template [[host_name("kernel_mul_mm_id_iq1_m_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_iq1_m, QK_NL, dequantize_iq1_m>;
template [[host_name("kernel_mul_mm_id_iq4_nl_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_iq4_nl, 2, dequantize_iq4_nl>;
#if QK_K == 64
template [[host_name("kernel_mul_mm_id_iq4_xs_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_iq4_xs, 2, dequantize_iq4_xs>;
#else
template [[host_name("kernel_mul_mm_id_iq4_xs_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_iq4_xs, QK_NL, dequantize_iq4_xs>;
#endif
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 13:04:25 +01:00
//
// matrix-vector multiplication
//
typedef void (kernel_mul_mv_impl_t)(
device const char * src0,
device const char * src1,
device float * dst,
int64_t ne00,
int64_t ne01,
int64_t ne02,
uint64_t nb00,
uint64_t nb01,
uint64_t nb02,
int64_t ne10,
int64_t ne11,
int64_t ne12,
uint64_t nb10,
uint64_t nb11,
uint64_t nb12,
int64_t ne0,
int64_t ne1,
uint r2,
uint r3,
uint3 tgpig,
uint tiisg);
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 13:04:25 +01:00
typedef void (kernel_mul_mv2_impl_t)(
device const void * src0,
device const float * src1,
device float * dst,
int64_t ne00,
int64_t ne01,
int64_t ne02,
int64_t ne10,
int64_t ne12,
int64_t ne0,
int64_t ne1,
uint r2,
uint r3,
threadgroup int8_t * shared_values,
uint3 tgpig,
uint tiisg,
uint sgitg);
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 13:04:25 +01:00
template<kernel_mul_mv_impl_t impl_fn>
void mmv_fn(
device const char * src0,
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 13:04:25 +01:00
device const char * src1,
device float * dst,
int64_t ne00,
int64_t ne01,
int64_t ne02,
uint64_t nb00,
uint64_t nb01,
uint64_t nb02,
int64_t ne10,
int64_t ne11,
int64_t ne12,
int64_t ne13,
uint64_t nb10,
uint64_t nb11,
uint64_t nb12,
int64_t ne0,
int64_t ne1,
uint64_t nb1,
uint r2,
uint r3,
threadgroup int8_t * shared_values,
uint3 tgpig,
uint tiitg,
uint tiisg,
uint sgitg) {
impl_fn(src0,src1,dst,ne00,ne01,ne02,nb00,nb01,nb02,ne10,ne11,ne12,nb10,nb11,nb12,ne0,ne1,r2,r3,tgpig,tiisg);
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 13:04:25 +01:00
}
template<kernel_mul_mv2_impl_t impl_fn>
void mmv_fn(
device const char * src0,
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 13:04:25 +01:00
device const char * src1,
device float * dst,
int64_t ne00,
int64_t ne01,
int64_t ne02,
uint64_t nb00,
uint64_t nb01,
uint64_t nb02,
int64_t ne10,
int64_t ne11,
int64_t ne12,
int64_t ne13,
uint64_t nb10,
uint64_t nb11,
uint64_t nb12,
int64_t ne0,
int64_t ne1,
uint64_t nb1,
uint r2,
uint r3,
threadgroup int8_t * shared_values,
uint3 tgpig,
uint tiitg,
uint tiisg,
uint sgitg) {
impl_fn(src0,(const device float *)src1,dst,ne00,ne01,ne02,ne10,ne12,ne0,ne1,r2,r3,shared_values,tgpig,tiisg,sgitg);
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 13:04:25 +01:00
}
typedef decltype(mmv_fn<kernel_mul_mv_f32_f32_impl>) mul_mv_impl_fn_t;
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 13:04:25 +01:00
template<mul_mv_impl_fn_t impl_fn>
kernel void kernel_mul_mv_id(
ggml : mul_mat_id use the same tensor for all the experts (#6387) * ggml : update mul_mat_id to use the same tensor for all the experts * update cuda * minor * update metal * update test-backend-ops * fix cuda * Update ggml-metal.m Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * update convert.py * update convert-hf-to-gguf.py * update convert.py for mixtral hf models * Update convert-hf-to-gguf.py Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * cuda : support non-pow-2 number of experts * allow quantize to work for split and merged experts models in the same way * cleanup + disable mmap automatically with split tensors models * update imatrix * test-backend-ops : test qwen argsort * update grok model loading * llama : add merged experts tensors to the grok tensor map * minor * gguf : bump version * fix quantizing of merged experts * convert-hf-to-gguf.py : update grok (untested) * make linter happy * cuda/argsort : use shared memory instead of pool memory * convert : fix grok tensor names * metal : add support for non-pow-2 argsort * llama : more loader cleanup, better error checking * cuda : fix warning * llama : still use mmap for loading old models, but copy the data to a host buffer * add review note * llama : remove ffn tensor counting + add sanity check ggml-ci * convert : fix handling of n_experts == None ggml-ci * imatrix : fix ncall counters * llama : produce error if imatrix size does not match * quantize : terminate on errors + trace logs ggml-ci * metal : pad shared memory to 16 bytes --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-04-03 15:07:05 +02:00
device const char * src0s,
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 13:04:25 +01:00
device const char * src1,
device float * dst,
ggml : mul_mat_id use the same tensor for all the experts (#6387) * ggml : update mul_mat_id to use the same tensor for all the experts * update cuda * minor * update metal * update test-backend-ops * fix cuda * Update ggml-metal.m Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * update convert.py * update convert-hf-to-gguf.py * update convert.py for mixtral hf models * Update convert-hf-to-gguf.py Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * cuda : support non-pow-2 number of experts * allow quantize to work for split and merged experts models in the same way * cleanup + disable mmap automatically with split tensors models * update imatrix * test-backend-ops : test qwen argsort * update grok model loading * llama : add merged experts tensors to the grok tensor map * minor * gguf : bump version * fix quantizing of merged experts * convert-hf-to-gguf.py : update grok (untested) * make linter happy * cuda/argsort : use shared memory instead of pool memory * convert : fix grok tensor names * metal : add support for non-pow-2 argsort * llama : more loader cleanup, better error checking * cuda : fix warning * llama : still use mmap for loading old models, but copy the data to a host buffer * add review note * llama : remove ffn tensor counting + add sanity check ggml-ci * convert : fix handling of n_experts == None ggml-ci * imatrix : fix ncall counters * llama : produce error if imatrix size does not match * quantize : terminate on errors + trace logs ggml-ci * metal : pad shared memory to 16 bytes --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-04-03 15:07:05 +02:00
device const char * ids,
constant int64_t & nei0,
constant int64_t & nei1,
constant uint64_t & nbi1,
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 13:04:25 +01:00
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
constant uint64_t & nb00,
constant uint64_t & nb01,
constant uint64_t & nb02,
constant int64_t & ne10,
constant int64_t & ne11,
constant int64_t & ne12,
constant int64_t & ne13,
constant uint64_t & nb10,
constant uint64_t & nb11,
constant uint64_t & nb12,
constant int64_t & ne0,
constant int64_t & ne1,
constant uint64_t & nb1,
threadgroup int8_t * shared_values [[threadgroup(0)]],
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 13:04:25 +01:00
uint3 tgpig[[threadgroup_position_in_grid]],
uint tiitg[[thread_index_in_threadgroup]],
uint tiisg[[thread_index_in_simdgroup]],
uint sgitg[[simdgroup_index_in_threadgroup]]) {
const int iid1 = tgpig.z/nei0;
const int idx = tgpig.z%nei0;
tgpig.z = 0;
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 13:04:25 +01:00
const int32_t i02 = ((device const int32_t *) (ids + iid1*nbi1))[idx];
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 13:04:25 +01:00
const int64_t i11 = idx % ne11;
const int64_t i12 = iid1;
const int64_t i1 = idx;
const int64_t i2 = i12;
device const char * src0_cur = src0s + i02*nb02;
device const char * src1_cur = src1 + i11*nb11 + i12*nb12;
device float * dst_cur = dst + i1*ne0 + i2*ne1*ne0;
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 13:04:25 +01:00
impl_fn(
/* src0 */ src0_cur,
/* src1 */ src1_cur,
/* dst */ dst_cur,
/* ne00 */ ne00,
/* ne01 */ ne01,
/* ne02 */ 1,//ne02,
/* nb00 */ nb00,
/* nb01 */ nb01,
/* nb02 */ nb02,
/* ne10 */ ne10,
/* ne11 */ 1,//ne11,
/* ne12 */ 1,//ne12,
/* ne13 */ 1,//ne13,
/* nb10 */ nb10,
/* nb11 */ nb11,
/* nb12 */ nb12,
/* ne0 */ ne0,
/* ne1 */ 1,//ne1,
/* nb1 */ nb1,
/* r2 */ 1,
/* r3 */ 1,
shared_values,
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 13:04:25 +01:00
tgpig,
tiitg,
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 13:04:25 +01:00
tiisg,
sgitg);
}
typedef decltype(kernel_mul_mv_id<mmv_fn<kernel_mul_mv_f32_f32_impl>>) kernel_mul_mv_id_t;
template [[host_name("kernel_mul_mv_id_f32_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_f32_f32_impl>>;
template [[host_name("kernel_mul_mv_id_f16_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_f16_f32_impl>>;
template [[host_name("kernel_mul_mv_id_q8_0_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_q8_0_f32_impl>>;
template [[host_name("kernel_mul_mv_id_q4_0_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<mul_vec_q_n_f32_impl<block_q4_0, N_DST, N_SIMDGROUP, N_SIMDWIDTH>>>;
template [[host_name("kernel_mul_mv_id_q4_1_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<mul_vec_q_n_f32_impl<block_q4_1, N_DST, N_SIMDGROUP, N_SIMDWIDTH>>>;
template [[host_name("kernel_mul_mv_id_q5_0_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<mul_vec_q_n_f32_impl<block_q5_0, N_DST, N_SIMDGROUP, N_SIMDWIDTH>>>;
template [[host_name("kernel_mul_mv_id_q5_1_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<mul_vec_q_n_f32_impl<block_q5_1, N_DST, N_SIMDGROUP, N_SIMDWIDTH>>>;
template [[host_name("kernel_mul_mv_id_q2_K_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_q2_K_f32_impl>>;
template [[host_name("kernel_mul_mv_id_q3_K_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_q3_K_f32_impl>>;
template [[host_name("kernel_mul_mv_id_q4_K_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_q4_K_f32_impl>>;
template [[host_name("kernel_mul_mv_id_q5_K_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_q5_K_f32_impl>>;
template [[host_name("kernel_mul_mv_id_q6_K_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_q6_K_f32_impl>>;
template [[host_name("kernel_mul_mv_id_iq1_s_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_iq1_s_f32_impl>>;
template [[host_name("kernel_mul_mv_id_iq1_m_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_iq1_m_f32_impl>>;
template [[host_name("kernel_mul_mv_id_iq2_xxs_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_iq2_xxs_f32_impl>>;
template [[host_name("kernel_mul_mv_id_iq2_xs_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_iq2_xs_f32_impl>>;
template [[host_name("kernel_mul_mv_id_iq3_xxs_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_iq3_xxs_f32_impl>>;
template [[host_name("kernel_mul_mv_id_iq3_s_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_iq3_s_f32_impl>>;
template [[host_name("kernel_mul_mv_id_iq2_s_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_iq2_s_f32_impl>>;
template [[host_name("kernel_mul_mv_id_iq4_nl_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_iq4_nl_f32_impl>>;
#if QK_K != 64
template [[host_name("kernel_mul_mv_id_iq4_xs_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_iq4_xs_f32_impl>>;
#endif
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 13:04:25 +01:00