llama.cpp/ggml-metal.m

3268 lines
187 KiB
Mathematica
Raw Permalink Normal View History

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
#import "ggml-metal.h"
#import "ggml-backend-impl.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
#import "ggml.h"
#import <Foundation/Foundation.h>
#import <Metal/Metal.h>
#undef MIN
#undef MAX
#define MIN(a, b) ((a) < (b) ? (a) : (b))
#define MAX(a, b) ((a) > (b) ? (a) : (b))
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
#ifdef GGML_METAL_NDEBUG
#define GGML_METAL_LOG_INFO(...)
#define GGML_METAL_LOG_WARN(...)
#define GGML_METAL_LOG_ERROR(...)
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
#else
#define GGML_METAL_LOG_INFO(...) ggml_metal_log(GGML_LOG_LEVEL_INFO, __VA_ARGS__)
#define GGML_METAL_LOG_WARN(...) ggml_metal_log(GGML_LOG_LEVEL_WARN, __VA_ARGS__)
#define GGML_METAL_LOG_ERROR(...) ggml_metal_log(GGML_LOG_LEVEL_ERROR, __VA_ARGS__)
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
#endif
#define UNUSED(x) (void)(x)
struct ggml_metal_kernel {
id<MTLComputePipelineState> pipeline;
};
enum ggml_metal_kernel_type {
GGML_METAL_KERNEL_TYPE_ADD,
GGML_METAL_KERNEL_TYPE_ADD_ROW,
GGML_METAL_KERNEL_TYPE_MUL,
GGML_METAL_KERNEL_TYPE_MUL_ROW,
GGML_METAL_KERNEL_TYPE_DIV,
GGML_METAL_KERNEL_TYPE_DIV_ROW,
GGML_METAL_KERNEL_TYPE_REPEAT_F32,
GGML_METAL_KERNEL_TYPE_REPEAT_F16,
GGML_METAL_KERNEL_TYPE_REPEAT_I32,
GGML_METAL_KERNEL_TYPE_REPEAT_I16,
GGML_METAL_KERNEL_TYPE_SCALE,
GGML_METAL_KERNEL_TYPE_SCALE_4,
GGML_METAL_KERNEL_TYPE_CLAMP,
GGML_METAL_KERNEL_TYPE_TANH,
GGML_METAL_KERNEL_TYPE_RELU,
GGML_METAL_KERNEL_TYPE_SIGMOID,
GGML_METAL_KERNEL_TYPE_GELU,
GGML_METAL_KERNEL_TYPE_GELU_4,
GGML_METAL_KERNEL_TYPE_GELU_QUICK,
GGML_METAL_KERNEL_TYPE_GELU_QUICK_4,
GGML_METAL_KERNEL_TYPE_SILU,
GGML_METAL_KERNEL_TYPE_SILU_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
GGML_METAL_KERNEL_TYPE_SOFT_MAX_F16,
GGML_METAL_KERNEL_TYPE_SOFT_MAX_F16_4,
GGML_METAL_KERNEL_TYPE_SOFT_MAX_F32,
GGML_METAL_KERNEL_TYPE_SOFT_MAX_F32_4,
GGML_METAL_KERNEL_TYPE_DIAG_MASK_INF,
GGML_METAL_KERNEL_TYPE_DIAG_MASK_INF_8,
GGML_METAL_KERNEL_TYPE_GET_ROWS_F32,
GGML_METAL_KERNEL_TYPE_GET_ROWS_F16,
GGML_METAL_KERNEL_TYPE_GET_ROWS_Q4_0,
GGML_METAL_KERNEL_TYPE_GET_ROWS_Q4_1,
GGML_METAL_KERNEL_TYPE_GET_ROWS_Q5_0,
GGML_METAL_KERNEL_TYPE_GET_ROWS_Q5_1,
GGML_METAL_KERNEL_TYPE_GET_ROWS_Q8_0,
GGML_METAL_KERNEL_TYPE_GET_ROWS_Q2_K,
GGML_METAL_KERNEL_TYPE_GET_ROWS_Q3_K,
GGML_METAL_KERNEL_TYPE_GET_ROWS_Q4_K,
GGML_METAL_KERNEL_TYPE_GET_ROWS_Q5_K,
GGML_METAL_KERNEL_TYPE_GET_ROWS_Q6_K,
GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XXS,
GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XS,
GGML_METAL_KERNEL_TYPE_GET_ROWS_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
GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ3_S,
GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_S,
GGML_METAL_KERNEL_TYPE_GET_ROWS_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
GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ1_M,
GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_NL,
GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_XS,
GGML_METAL_KERNEL_TYPE_GET_ROWS_I32,
GGML_METAL_KERNEL_TYPE_RMS_NORM,
GGML_METAL_KERNEL_TYPE_GROUP_NORM,
GGML_METAL_KERNEL_TYPE_NORM,
GGML_METAL_KERNEL_TYPE_MUL_MV_F32_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F16,
GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_1ROW,
GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_L4,
GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_0_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_1_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_0_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_1_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_Q8_0_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_Q2_K_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_Q3_K_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_K_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_K_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_Q6_K_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XXS_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XS_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_XXS_F32,
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
GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_S_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_S_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_S_F32,
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
GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_M_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_NL_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_XS_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F32_F32,
//GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F16,
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32,
//GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32_1ROW,
//GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32_L4,
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_0_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_1_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_0_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_1_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q8_0_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q2_K_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q3_K_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_K_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_K_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q6_K_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_XXS_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_XS_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_XXS_F32,
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
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_S_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_S_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ1_S_F32,
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
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ1_M_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_NL_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_XS_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_F32_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_F16_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_0_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_1_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_0_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_1_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_Q8_0_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_Q2_K_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_Q3_K_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_K_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_K_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_Q6_K_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XXS_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XS_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_XXS_F32,
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
GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_S_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_S_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_S_F32,
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
GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_M_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_NL_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_XS_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F32_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F16_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_0_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_1_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_0_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_1_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q8_0_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q2_K_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q3_K_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_K_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_K_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q6_K_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XXS_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XS_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_XXS_F32,
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
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_S_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_S_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_S_F32,
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
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_M_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_NL_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_XS_F32,
GGML_METAL_KERNEL_TYPE_ROPE_NORM_F32,
GGML_METAL_KERNEL_TYPE_ROPE_NORM_F16,
GGML_METAL_KERNEL_TYPE_ROPE_NEOX_F32,
GGML_METAL_KERNEL_TYPE_ROPE_NEOX_F16,
GGML_METAL_KERNEL_TYPE_IM2COL_F16,
GGML_METAL_KERNEL_TYPE_IM2COL_F32,
GGML_METAL_KERNEL_TYPE_UPSCALE_F32,
GGML_METAL_KERNEL_TYPE_PAD_F32,
GGML_METAL_KERNEL_TYPE_ARANGE_F32,
GGML_METAL_KERNEL_TYPE_TIMESTEP_EMBEDDING_F32,
GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_ASC,
GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_DESC,
GGML_METAL_KERNEL_TYPE_LEAKY_RELU_F32,
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
GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H64,
GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H80,
GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H96,
GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H112,
GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H128,
//GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H256, // https://github.com/ggerganov/llama.cpp/issues/7261
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
GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H128,
//GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H256, // https://github.com/ggerganov/llama.cpp/issues/7261
GGML_METAL_KERNEL_TYPE_CPY_F32_F16,
GGML_METAL_KERNEL_TYPE_CPY_F32_F32,
GGML_METAL_KERNEL_TYPE_CPY_F32_Q8_0,
GGML_METAL_KERNEL_TYPE_CPY_F32_Q4_0,
GGML_METAL_KERNEL_TYPE_CPY_F32_Q4_1,
GGML_METAL_KERNEL_TYPE_CPY_F32_Q5_0,
GGML_METAL_KERNEL_TYPE_CPY_F32_Q5_1,
GGML_METAL_KERNEL_TYPE_CPY_F32_IQ4_NL,
GGML_METAL_KERNEL_TYPE_CPY_F16_F16,
GGML_METAL_KERNEL_TYPE_CPY_F16_F32,
GGML_METAL_KERNEL_TYPE_CONCAT,
GGML_METAL_KERNEL_TYPE_SQR,
GGML_METAL_KERNEL_TYPE_SUM_ROWS,
GGML_METAL_KERNEL_TYPE_COUNT
};
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
struct ggml_metal_context {
int n_cb;
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
id<MTLDevice> device;
id<MTLCommandQueue> queue;
dispatch_queue_t d_queue;
struct ggml_metal_kernel kernels[GGML_METAL_KERNEL_TYPE_COUNT];
bool support_simdgroup_reduction;
bool support_simdgroup_mm;
bool should_capture_next_compute;
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
};
// MSL code
// TODO: move the contents here when ready
// for now it is easier to work in a separate file
// static NSString * const msl_library_source = @"see metal.metal";
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
// Here to assist with NSBundle Path Hack
@interface GGMLMetalClass : NSObject
@end
@implementation GGMLMetalClass
@end
static void ggml_metal_default_log_callback(enum ggml_log_level level, const char * msg, void * user_data) {
fprintf(stderr, "%s", msg);
UNUSED(level);
UNUSED(user_data);
}
ggml_log_callback ggml_metal_log_callback = ggml_metal_default_log_callback;
void * ggml_metal_log_user_data = NULL;
GGML_ATTRIBUTE_FORMAT(2, 3)
static void ggml_metal_log(enum ggml_log_level level, const char * format, ...){
if (ggml_metal_log_callback != NULL) {
va_list args;
va_start(args, format);
char buffer[128];
int len = vsnprintf(buffer, 128, format, args);
if (len < 128) {
ggml_metal_log_callback(level, buffer, ggml_metal_log_user_data);
} else {
char* buffer2 = malloc(len+1);
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
va_end(args);
va_start(args, format);
vsnprintf(buffer2, len+1, format, args);
buffer2[len] = 0;
ggml_metal_log_callback(level, buffer2, ggml_metal_log_user_data);
free(buffer2);
}
va_end(args);
}
}
2024-01-13 19:45:45 +01:00
static void * ggml_metal_host_malloc(size_t n) {
void * data = NULL;
#if TARGET_OS_OSX
kern_return_t err = vm_allocate((vm_map_t) mach_task_self(), (void *) &data, n, VM_FLAGS_ANYWHERE);
if (err != KERN_SUCCESS) {
GGML_METAL_LOG_ERROR("%s: error: vm_allocate failed\n", __func__);
return NULL;
}
#else
2024-01-13 19:45:45 +01:00
const int result = posix_memalign((void **) &data, sysconf(_SC_PAGESIZE), n);
if (result != 0) {
GGML_METAL_LOG_ERROR("%s: error: posix_memalign failed\n", __func__);
return NULL;
}
#endif
2024-01-13 19:45:45 +01:00
return data;
}
static struct ggml_metal_context * ggml_metal_init(int n_cb) {
GGML_METAL_LOG_INFO("%s: allocating\n", __func__);
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 TARGET_OS_OSX && !GGML_METAL_NDEBUG
// Show all the Metal device instances in the system
NSArray * devices = MTLCopyAllDevices();
for (id<MTLDevice> device in devices) {
GGML_METAL_LOG_INFO("%s: found device: %s\n", __func__, [[device name] UTF8String]);
}
[devices release]; // since it was created by a *Copy* C method
#endif
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
// Pick and show default Metal device
id<MTLDevice> device = MTLCreateSystemDefaultDevice();
GGML_METAL_LOG_INFO("%s: picking default device: %s\n", __func__, [[device name] UTF8String]);
// Configure context
struct ggml_metal_context * ctx = malloc(sizeof(struct ggml_metal_context));
ctx->device = device;
ctx->n_cb = MIN(n_cb, GGML_METAL_MAX_BUFFERS);
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
ctx->queue = [ctx->device newCommandQueue];
ctx->d_queue = dispatch_queue_create("ggml-metal", DISPATCH_QUEUE_CONCURRENT);
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
id<MTLLibrary> metal_library;
// load library
//
// - first check if the library is embedded
// - then check if the library is in the bundle
// - if not found, load the source and compile it
// - if that fails, return NULL
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
{
NSBundle * bundle = nil;
#ifdef SWIFT_PACKAGE
bundle = SWIFTPM_MODULE_BUNDLE;
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
#else
bundle = [NSBundle bundleForClass:[GGMLMetalClass class]];
#endif
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
NSError * error = nil;
#if GGML_METAL_EMBED_LIBRARY
const bool try_metallib = false;
#else
const bool try_metallib = true;
#endif
NSString * path_lib = [bundle pathForResource:@"default" ofType:@"metallib"];
if (try_metallib && path_lib != nil) {
// pre-compiled library found
NSURL * libURL = [NSURL fileURLWithPath:path_lib];
GGML_METAL_LOG_INFO("%s: loading '%s'\n", __func__, [path_lib UTF8String]);
metal_library = [ctx->device newLibraryWithURL:libURL error:&error];
2024-01-25 10:26:17 +01:00
if (error) {
GGML_METAL_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]);
return NULL;
}
} else {
#if GGML_METAL_EMBED_LIBRARY
GGML_METAL_LOG_INFO("%s: using embedded metal library\n", __func__);
extern const char ggml_metallib_start[];
extern const char ggml_metallib_end[];
NSString * src = [[NSString alloc] initWithBytes:ggml_metallib_start length:(ggml_metallib_end-ggml_metallib_start) encoding:NSUTF8StringEncoding];
#else
GGML_METAL_LOG_INFO("%s: default.metallib not found, loading from source\n", __func__);
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
NSString * path_source;
NSString * path_resource = [[NSProcessInfo processInfo].environment objectForKey:@"GGML_METAL_PATH_RESOURCES"];
GGML_METAL_LOG_INFO("%s: GGML_METAL_PATH_RESOURCES = %s\n", __func__, path_resource ? [path_resource UTF8String] : "nil");
if (path_resource) {
path_source = [path_resource stringByAppendingPathComponent:@"ggml-metal.metal"];
} else {
path_source = [bundle pathForResource:@"ggml-metal" ofType:@"metal"];
}
if (path_source == nil) {
GGML_METAL_LOG_WARN("%s: error: could not use bundle path to find ggml-metal.metal, falling back to trying cwd\n", __func__);
path_source = @"ggml-metal.metal";
}
GGML_METAL_LOG_INFO("%s: loading '%s'\n", __func__, [path_source UTF8String]);
NSString * src = [NSString stringWithContentsOfFile:path_source encoding:NSUTF8StringEncoding error:&error];
if (error) {
GGML_METAL_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]);
return NULL;
}
#endif // GGML_METAL_EMBED_LIBRARY
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
@autoreleasepool {
// dictionary of preprocessor macros
NSMutableDictionary * prep = [NSMutableDictionary dictionary];
MTLCompileOptions* options = [MTLCompileOptions new];
options.preprocessorMacros = prep;
//[options setFastMathEnabled:false];
metal_library = [ctx->device newLibraryWithSource:src options:options error:&error];
2024-01-25 10:26:17 +01:00
if (error) {
GGML_METAL_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]);
return NULL;
}
}
}
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
}
// print MTL GPU family:
GGML_METAL_LOG_INFO("%s: GPU name: %s\n", __func__, [[ctx->device name] UTF8String]);
const NSInteger MTLGPUFamilyMetal3 = 5001;
// determine max supported GPU family
// https://developer.apple.com/metal/Metal-Shading-Language-Specification.pdf
// https://developer.apple.com/metal/Metal-Feature-Set-Tables.pdf
{
for (int i = MTLGPUFamilyApple1 + 20; i >= MTLGPUFamilyApple1; --i) {
if ([ctx->device supportsFamily:i]) {
GGML_METAL_LOG_INFO("%s: GPU family: MTLGPUFamilyApple%d (%d)\n", __func__, i - (int) MTLGPUFamilyApple1 + 1, i);
break;
}
}
for (int i = MTLGPUFamilyCommon1 + 5; i >= MTLGPUFamilyCommon1; --i) {
if ([ctx->device supportsFamily:i]) {
GGML_METAL_LOG_INFO("%s: GPU family: MTLGPUFamilyCommon%d (%d)\n", __func__, i - (int) MTLGPUFamilyCommon1 + 1, i);
break;
}
}
for (int i = MTLGPUFamilyMetal3 + 5; i >= MTLGPUFamilyMetal3; --i) {
if ([ctx->device supportsFamily:i]) {
GGML_METAL_LOG_INFO("%s: GPU family: MTLGPUFamilyMetal%d (%d)\n", __func__, i - (int) MTLGPUFamilyMetal3 + 3, i);
break;
}
}
}
ctx->support_simdgroup_reduction = [ctx->device supportsFamily:MTLGPUFamilyApple7];
ctx->support_simdgroup_reduction |= [ctx->device supportsFamily:MTLGPUFamilyMetal3];
ctx->support_simdgroup_mm = [ctx->device supportsFamily:MTLGPUFamilyApple7];
GGML_METAL_LOG_INFO("%s: simdgroup reduction support = %s\n", __func__, ctx->support_simdgroup_reduction ? "true" : "false");
GGML_METAL_LOG_INFO("%s: simdgroup matrix mul. support = %s\n", __func__, ctx->support_simdgroup_mm ? "true" : "false");
GGML_METAL_LOG_INFO("%s: hasUnifiedMemory = %s\n", __func__, ctx->device.hasUnifiedMemory ? "true" : "false");
ctx->should_capture_next_compute = false;
#if TARGET_OS_OSX || (TARGET_OS_IOS && __clang_major__ >= 15)
if (@available(macOS 10.12, iOS 16.0, *)) {
GGML_METAL_LOG_INFO("%s: recommendedMaxWorkingSetSize = %8.2f MB\n", __func__, ctx->device.recommendedMaxWorkingSetSize / 1e6);
}
#elif TARGET_OS_OSX
if (ctx->device.maxTransferRate != 0) {
GGML_METAL_LOG_INFO("%s: maxTransferRate = %8.2f MB/s\n", __func__, ctx->device.maxTransferRate / 1e6);
} else {
GGML_METAL_LOG_INFO("%s: maxTransferRate = built-in GPU\n", __func__);
}
#endif
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
// load kernels
{
NSError * error = nil;
llama : refactor graph build code (#3837) * llama : factor out ggml-alloc from graph graph build functions ggml-ci * metal : disable kernel load log * llama : factor out tensor offloading outside the build call (wip) ggml-ci * llama : offload rest of the models ggml-ci * llama : update offload log messages to print node index * llama : comments * llama : support offloading result_norm + comments * llama : factor graph input into a function * llama : do tensor offload only with CUDA * llama : fix res_norm offloading * llama : try to optimize offloading code * llama : fix non-CUDA build * llama : try to fix build * llama : move refact in correct place + optimize graph input * llama : refactor tensor offloading as callback * llama : add layer index to all tensor names * llama : add functional header * llama : comment ggml-ci * llama : remove obsolete map for layer counting * llama : add llm_build helper functions (#3848) * llama : add llm_build_norm helper function ggml-ci * llama : add llm_build_ffn helper function (#3849) ggml-ci * llama : add llm_build_k_shift helper ggml-ci * llama : fix offloading after recent changes * llama : add llm_build_kv_store helper ggml-ci * llama : remove obsolete offload names * llama : fix llm_build_k_shift to use n_head_kv instead of n_head * llama : simplify falcon Q, K, V computation * llama : remove obsolete comments in build graphs * llama : add llm_build_kqv helper ggml-ci * llama : minor * llama : add LLAMA_OFFLOAD_DEBUG + fix starcoder offloading * llama : fix input allocation logic * llama : update offload functions for KQ tensors * llama : normalize tensor names ggml-ci * llama : enable warning about not offloaded tensors * llama : remove extra ; + deduplicate gate_b logic * llama : add llm_build_inp_embd helper
2023-11-01 07:04:02 +01:00
for (int i = 0; i < GGML_METAL_KERNEL_TYPE_COUNT; ++i) {
ctx->kernels[i].pipeline = nil;
}
llama : refactor graph build code (#3837) * llama : factor out ggml-alloc from graph graph build functions ggml-ci * metal : disable kernel load log * llama : factor out tensor offloading outside the build call (wip) ggml-ci * llama : offload rest of the models ggml-ci * llama : update offload log messages to print node index * llama : comments * llama : support offloading result_norm + comments * llama : factor graph input into a function * llama : do tensor offload only with CUDA * llama : fix res_norm offloading * llama : try to optimize offloading code * llama : fix non-CUDA build * llama : try to fix build * llama : move refact in correct place + optimize graph input * llama : refactor tensor offloading as callback * llama : add layer index to all tensor names * llama : add functional header * llama : comment ggml-ci * llama : remove obsolete map for layer counting * llama : add llm_build helper functions (#3848) * llama : add llm_build_norm helper function ggml-ci * llama : add llm_build_ffn helper function (#3849) ggml-ci * llama : add llm_build_k_shift helper ggml-ci * llama : fix offloading after recent changes * llama : add llm_build_kv_store helper ggml-ci * llama : remove obsolete offload names * llama : fix llm_build_k_shift to use n_head_kv instead of n_head * llama : simplify falcon Q, K, V computation * llama : remove obsolete comments in build graphs * llama : add llm_build_kqv helper ggml-ci * llama : minor * llama : add LLAMA_OFFLOAD_DEBUG + fix starcoder offloading * llama : fix input allocation logic * llama : update offload functions for KQ tensors * llama : normalize tensor names ggml-ci * llama : enable warning about not offloaded tensors * llama : remove extra ; + deduplicate gate_b logic * llama : add llm_build_inp_embd helper
2023-11-01 07:04:02 +01:00
/*
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
GGML_METAL_LOG_INFO("%s: loaded %-40s %16p | th_max = %4d | th_width = %4d\n", __func__, "kernel_"#name, (void *) kernel->pipeline, \
(int) kernel->pipeline.maxTotalThreadsPerThreadgroup, \
(int) kernel->pipeline.threadExecutionWidth); \
llama : refactor graph build code (#3837) * llama : factor out ggml-alloc from graph graph build functions ggml-ci * metal : disable kernel load log * llama : factor out tensor offloading outside the build call (wip) ggml-ci * llama : offload rest of the models ggml-ci * llama : update offload log messages to print node index * llama : comments * llama : support offloading result_norm + comments * llama : factor graph input into a function * llama : do tensor offload only with CUDA * llama : fix res_norm offloading * llama : try to optimize offloading code * llama : fix non-CUDA build * llama : try to fix build * llama : move refact in correct place + optimize graph input * llama : refactor tensor offloading as callback * llama : add layer index to all tensor names * llama : add functional header * llama : comment ggml-ci * llama : remove obsolete map for layer counting * llama : add llm_build helper functions (#3848) * llama : add llm_build_norm helper function ggml-ci * llama : add llm_build_ffn helper function (#3849) ggml-ci * llama : add llm_build_k_shift helper ggml-ci * llama : fix offloading after recent changes * llama : add llm_build_kv_store helper ggml-ci * llama : remove obsolete offload names * llama : fix llm_build_k_shift to use n_head_kv instead of n_head * llama : simplify falcon Q, K, V computation * llama : remove obsolete comments in build graphs * llama : add llm_build_kqv helper ggml-ci * llama : minor * llama : add LLAMA_OFFLOAD_DEBUG + fix starcoder offloading * llama : fix input allocation logic * llama : update offload functions for KQ tensors * llama : normalize tensor names ggml-ci * llama : enable warning about not offloaded tensors * llama : remove extra ; + deduplicate gate_b logic * llama : add llm_build_inp_embd helper
2023-11-01 07:04:02 +01:00
*/
#define GGML_METAL_ADD_KERNEL(e, name, supported) \
if (supported) { \
struct ggml_metal_kernel * kernel = &ctx->kernels[e]; \
id<MTLFunction> metal_function = [metal_library newFunctionWithName:@"kernel_"#name]; \
kernel->pipeline = [ctx->device newComputePipelineStateWithFunction:metal_function error:&error]; \
[metal_function release]; \
if (error) { \
GGML_METAL_LOG_ERROR("%s: error: load pipeline error: %s\n", __func__, [[error description] UTF8String]); \
[metal_library release]; \
return NULL; \
} \
} else { \
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
GGML_METAL_LOG_WARN("%s: skipping %-40s (not supported)\n", __func__, "kernel_"#name); \
}
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
// simd_sum and simd_max requires MTLGPUFamilyApple7
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
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ADD, add, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ADD_ROW, add_row, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL, mul, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_ROW, mul_row, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIV, div, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIV_ROW, div_row, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_REPEAT_F32, repeat_f32, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_REPEAT_F16, repeat_f16, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_REPEAT_I32, repeat_i32, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_REPEAT_I16, repeat_i16, true);
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
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SCALE, scale, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SCALE_4, scale_4, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CLAMP, clamp, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_TANH, tanh, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_RELU, relu, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SIGMOID, sigmoid, true);
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
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU, gelu, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU_4, gelu_4, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU_QUICK, gelu_quick, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU_QUICK_4, gelu_quick_4, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SILU, silu, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SILU_4, silu_4, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX_F16, soft_max_f16, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX_F16_4, soft_max_f16_4, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX_F32, soft_max_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX_F32_4, soft_max_f32_4, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIAG_MASK_INF, diag_mask_inf, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIAG_MASK_INF_8, diag_mask_inf_8, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_F32, get_rows_f32, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_F16, get_rows_f16, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q4_0, get_rows_q4_0, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q4_1, get_rows_q4_1, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q5_0, get_rows_q5_0, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q5_1, get_rows_q5_1, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q8_0, get_rows_q8_0, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q2_K, get_rows_q2_K, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q3_K, get_rows_q3_K, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q4_K, get_rows_q4_K, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q5_K, get_rows_q5_K, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q6_K, get_rows_q6_K, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XXS, get_rows_iq2_xxs, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XS, get_rows_iq2_xs, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ3_XXS, get_rows_iq3_xxs, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ3_S, get_rows_iq3_s, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_S, get_rows_iq2_s, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ1_S, get_rows_iq1_s, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ1_M, get_rows_iq1_m, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_NL, get_rows_iq4_nl, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_XS, get_rows_iq4_xs, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_I32, get_rows_i32, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_RMS_NORM, rms_norm, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GROUP_NORM, group_norm, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_NORM, norm, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F32_F32, mul_mv_f32_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F16, mul_mv_f16_f16, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32, mul_mv_f16_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_1ROW, mul_mv_f16_f32_1row, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_L4, mul_mv_f16_f32_l4, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_0_F32, mul_mv_q4_0_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_1_F32, mul_mv_q4_1_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_0_F32, mul_mv_q5_0_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_1_F32, mul_mv_q5_1_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q8_0_F32, mul_mv_q8_0_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q2_K_F32, mul_mv_q2_K_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q3_K_F32, mul_mv_q3_K_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_K_F32, mul_mv_q4_K_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_K_F32, mul_mv_q5_K_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q6_K_F32, mul_mv_q6_K_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XXS_F32, mul_mv_iq2_xxs_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XS_F32, mul_mv_iq2_xs_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_XXS_F32, mul_mv_iq3_xxs_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_S_F32, mul_mv_iq3_s_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_S_F32, mul_mv_iq2_s_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_S_F32, mul_mv_iq1_s_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_M_F32, mul_mv_iq1_m_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_NL_F32, mul_mv_iq4_nl_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_XS_F32, mul_mv_iq4_xs_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F32_F32, mul_mv_id_f32_f32, ctx->support_simdgroup_reduction);
//GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F16, mul_mv_id_f16_f16, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32, mul_mv_id_f16_f32, ctx->support_simdgroup_reduction);
//GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32_1ROW, mul_mv_id_f16_f32_1row, ctx->support_simdgroup_reduction);
//GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32_L4, mul_mv_id_f16_f32_l4, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_0_F32, mul_mv_id_q4_0_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_1_F32, mul_mv_id_q4_1_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_0_F32, mul_mv_id_q5_0_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_1_F32, mul_mv_id_q5_1_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q8_0_F32, mul_mv_id_q8_0_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q2_K_F32, mul_mv_id_q2_K_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q3_K_F32, mul_mv_id_q3_K_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_K_F32, mul_mv_id_q4_K_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_K_F32, mul_mv_id_q5_K_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q6_K_F32, mul_mv_id_q6_K_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_XXS_F32, mul_mv_id_iq2_xxs_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_XS_F32, mul_mv_id_iq2_xs_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_XXS_F32, mul_mv_id_iq3_xxs_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_S_F32, mul_mv_id_iq3_s_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_S_F32, mul_mv_id_iq2_s_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ1_S_F32, mul_mv_id_iq1_s_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ1_M_F32, mul_mv_id_iq1_m_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_NL_F32, mul_mv_id_iq4_nl_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_XS_F32, mul_mv_id_iq4_xs_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_F32_F32, mul_mm_f32_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_F16_F32, mul_mm_f16_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_0_F32, mul_mm_q4_0_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_1_F32, mul_mm_q4_1_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_0_F32, mul_mm_q5_0_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_1_F32, mul_mm_q5_1_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q8_0_F32, mul_mm_q8_0_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q2_K_F32, mul_mm_q2_K_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q3_K_F32, mul_mm_q3_K_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_K_F32, mul_mm_q4_K_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_K_F32, mul_mm_q5_K_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q6_K_F32, mul_mm_q6_K_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XXS_F32, mul_mm_iq2_xxs_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XS_F32, mul_mm_iq2_xs_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_XXS_F32, mul_mm_iq3_xxs_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_S_F32, mul_mm_iq3_s_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_S_F32, mul_mm_iq2_s_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_S_F32, mul_mm_iq1_s_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_M_F32, mul_mm_iq1_m_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_NL_F32, mul_mm_iq4_nl_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_XS_F32, mul_mm_iq4_xs_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F32_F32, mul_mm_id_f32_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F16_F32, mul_mm_id_f16_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_0_F32, mul_mm_id_q4_0_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_1_F32, mul_mm_id_q4_1_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_0_F32, mul_mm_id_q5_0_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_1_F32, mul_mm_id_q5_1_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q8_0_F32, mul_mm_id_q8_0_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q2_K_F32, mul_mm_id_q2_K_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q3_K_F32, mul_mm_id_q3_K_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_K_F32, mul_mm_id_q4_K_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_K_F32, mul_mm_id_q5_K_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q6_K_F32, mul_mm_id_q6_K_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XXS_F32, mul_mm_id_iq2_xxs_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XS_F32, mul_mm_id_iq2_xs_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_XXS_F32, mul_mm_id_iq3_xxs_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_S_F32, mul_mm_id_iq3_s_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_S_F32, mul_mm_id_iq2_s_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_S_F32, mul_mm_id_iq1_s_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_M_F32, mul_mm_id_iq1_m_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_NL_F32, mul_mm_id_iq4_nl_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_XS_F32, mul_mm_id_iq4_xs_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ROPE_NORM_F32, rope_norm_f32, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ROPE_NORM_F16, rope_norm_f16, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ROPE_NEOX_F32, rope_neox_f32, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ROPE_NEOX_F16, rope_neox_f16, true);
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
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_IM2COL_F16, im2col_f16, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_IM2COL_F32, im2col_f32, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_UPSCALE_F32, upscale_f32, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_PAD_F32, pad_f32, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_TIMESTEP_EMBEDDING_F32, timestep_embedding_f32, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARANGE_F32, arange_f32, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_ASC, argsort_f32_i32_asc, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_DESC, argsort_f32_i32_desc, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_LEAKY_RELU_F32, leaky_relu_f32, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H64, flash_attn_ext_f16_h64, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H80, flash_attn_ext_f16_h80, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H96, flash_attn_ext_f16_h96, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H112, flash_attn_ext_f16_h112, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H128, flash_attn_ext_f16_h128, ctx->support_simdgroup_mm);
//GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H256, flash_attn_ext_f16_h256, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H128, flash_attn_ext_vec_f16_h128, ctx->support_simdgroup_reduction);
//GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H256, flash_attn_ext_vec_f16_h256, ctx->support_simdgroup_reduction);
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
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_F16, cpy_f32_f16, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_F32, cpy_f32_f32, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q8_0, cpy_f32_q8_0, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q4_0, cpy_f32_q4_0, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q4_1, cpy_f32_q4_1, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q5_0, cpy_f32_q5_0, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q5_1, cpy_f32_q5_1, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_IQ4_NL, cpy_f32_iq4_nl, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F16_F16, cpy_f16_f16, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F16_F32, cpy_f16_f32, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CONCAT, concat, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SQR, sqr, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SUM_ROWS, sum_rows, true);
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
}
[metal_library release];
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
return ctx;
}
2024-01-13 19:45:45 +01:00
static void ggml_metal_free(struct ggml_metal_context * ctx) {
GGML_METAL_LOG_INFO("%s: deallocating\n", __func__);
for (int i = 0; i < GGML_METAL_KERNEL_TYPE_COUNT; ++i) {
[ctx->kernels[i].pipeline release];
}
[ctx->queue release];
[ctx->device release];
dispatch_release(ctx->d_queue);
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
free(ctx);
}
// temporarily defined here for compatibility between ggml-backend and the old API
struct ggml_backend_metal_buffer {
void * data;
size_t size;
id<MTLBuffer> metal;
};
struct ggml_backend_metal_buffer_context {
void * all_data;
size_t all_size;
bool owned;
// multiple buffers are used only to avoid the maximum buffer size limitation when using mmap
int n_buffers;
struct ggml_backend_metal_buffer buffers[GGML_METAL_MAX_BUFFERS];
};
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
// finds the Metal buffer that contains the tensor data on the GPU device
// the assumption is that there is 1-to-1 mapping between the host and device memory buffers, so we can find the
// Metal buffer based on the host memory pointer
//
static id<MTLBuffer> ggml_metal_get_buffer(struct ggml_tensor * t, size_t * offs) {
//GGML_METAL_LOG_INFO("%s: data tensor '%16s', offs_data = %8ld, offs_eval = %8ld, offs_cach = %8ld\n", __func__, t->name, offs_data, offs_eval, offs_cach);
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 tsize = ggml_nbytes(t);
ggml_backend_buffer_t buffer = t->view_src ? t->view_src->buffer : t->buffer;
struct ggml_backend_metal_buffer_context * buf_ctx = (struct ggml_backend_metal_buffer_context *) buffer->context;
// find the view that contains the tensor fully
for (int i = 0; i < buf_ctx->n_buffers; ++i) {
const int64_t ioffs = (int64_t) t->data - (int64_t) buf_ctx->buffers[i].data;
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_METAL_LOG_INFO("ioffs = %10ld, tsize = %10ld, sum = %10ld, buf_ctx->buffers[%d].size = %10ld\n", ioffs, tsize, ioffs + tsize, i, buf_ctx->buffers[i].size);
if (ioffs >= 0 && ioffs + tsize <= (int64_t) buf_ctx->buffers[i].size) {
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
*offs = (size_t) ioffs;
//GGML_METAL_LOG_INFO("%s: tensor '%16s', offs = %8ld\n", __func__, t->name, *offs);
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
return buf_ctx->buffers[i].metal;
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_METAL_LOG_ERROR("%s: error: tensor '%s' buffer is nil\n", __func__, t->name);
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
return nil;
}
static bool ggml_metal_supports_op(const struct ggml_metal_context * ctx, const struct ggml_tensor * op) {
switch (op->op) {
case GGML_OP_UNARY:
switch (ggml_get_unary_op(op)) {
case GGML_UNARY_OP_TANH:
case GGML_UNARY_OP_RELU:
case GGML_UNARY_OP_SIGMOID:
case GGML_UNARY_OP_GELU:
case GGML_UNARY_OP_GELU_QUICK:
case GGML_UNARY_OP_SILU:
return ggml_is_contiguous(op->src[0]);
default:
return false;
}
case GGML_OP_NONE:
case GGML_OP_RESHAPE:
case GGML_OP_VIEW:
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
case GGML_OP_TRANSPOSE:
case GGML_OP_PERMUTE:
case GGML_OP_CONCAT:
case GGML_OP_ADD:
case GGML_OP_ACC:
case GGML_OP_MUL:
case GGML_OP_DIV:
case GGML_OP_REPEAT:
case GGML_OP_SCALE:
case GGML_OP_CLAMP:
case GGML_OP_SQR:
case GGML_OP_SUM_ROWS:
return true;
case GGML_OP_SOFT_MAX:
case GGML_OP_RMS_NORM:
case GGML_OP_GROUP_NORM:
return ctx->support_simdgroup_reduction;
case GGML_OP_NORM:
case GGML_OP_ROPE:
case GGML_OP_IM2COL:
return true;
case GGML_OP_POOL_1D:
case GGML_OP_POOL_2D:
return false;
case GGML_OP_UPSCALE:
case GGML_OP_PAD:
case GGML_OP_ARANGE:
case GGML_OP_TIMESTEP_EMBEDDING:
case GGML_OP_ARGSORT:
case GGML_OP_LEAKY_RELU:
return true;
case GGML_OP_FLASH_ATTN_EXT:
if (op->src[1]->type != GGML_TYPE_F16) {
return false;
}
if (op->src[2]->type != GGML_TYPE_F16) {
return false;
}
if (op->src[0]->ne[0] == 256) {
return false;
}
return ctx->support_simdgroup_mm; // TODO: over-restricted for vec-kernels
case GGML_OP_MUL_MAT:
case GGML_OP_MUL_MAT_ID:
return ctx->support_simdgroup_reduction &&
(op->src[0]->type != GGML_TYPE_F32 || op->src[1]->type == GGML_TYPE_F32);
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
case GGML_OP_CPY:
case GGML_OP_DUP:
case GGML_OP_CONT:
{
switch (op->src[0]->type) {
case GGML_TYPE_F32:
switch (op->type) {
case GGML_TYPE_F16:
case GGML_TYPE_F32:
case GGML_TYPE_Q8_0:
case GGML_TYPE_Q4_0:
case GGML_TYPE_Q4_1:
case GGML_TYPE_Q5_0:
case GGML_TYPE_Q5_1:
case GGML_TYPE_IQ4_NL:
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
return true;
default:
return false;
}
case GGML_TYPE_F16:
switch (op->type) {
case GGML_TYPE_F16:
case GGML_TYPE_F32:
return true;
default:
return false;
}
default:
return false;
};
}
case GGML_OP_DIAG_MASK_INF:
case GGML_OP_GET_ROWS:
{
ggml : introduce bfloat16 support (#6412) * Introduce bfloat16 support Many models on Hugging Face (e.g. Mistral, TinyLLaMA) use bfloat16 as their canonical floating point format. ┌sign │ │ ┌exponent │ │ │ │ ┌mantissa │ │ │ │┌──┴───┐┌─┴───┐ 0b0000000000000000 brain16 This encoding has the same number of exponent bits as float32. That makes conversion relatively straightforward, even in the absence of hardware support. For example, converting brain16 to binary32 means simply shifting 16 bits to the left. ┌sign │ │ ┌exponent │ │ │ │ ┌mantissa │ │ │ │┌──┴───┐┌─┴───────────────────┐ 0b00000000000000000000000000000000 IEEE binary32 The issue is that converting bf16 to fp16 can result in information loss. Only 13% of bf16 numbers can be precisely represented in fp16 which in practice ends up being 99.71% of Mistral 7b v0.2's weights however there is currently no way other than fp32 to get the others ┌sign │ │ ┌exponent │ │ │ │ ┌mantissa │ │ │ │┌─┴─┐┌─┴──────┐ 0b0000000000000000 IEEE binary16 This change fixes that, by adding a bf16 data type to GGML. Support for CPU inference has been implemented along with optimizations for the AVX2, AVX512, and AVX512BF16 ISAs. Perplexity on Mistral 7b 0.2 improves somewhere around -0.0024 to -0.0046 compared to using fp16 * Remove GGML code that's not needed * Minimize the GGML API surface area for BF16 * Remove bf16 luts * Make the GGML header look nicer * Fix documentation * Apply ggerganov's fixes for test-backend-ops * Add BF16 code for new ggml_validate_row_data() function
2024-05-08 08:30:09 +02:00
return op->src[0]->type != GGML_TYPE_BF16 && op->ne[3] == 1;
}
default:
return false;
}
}
static enum ggml_status ggml_metal_graph_compute(
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
struct ggml_metal_context * ctx,
struct ggml_cgraph * gf) {
@autoreleasepool {
MTLComputePassDescriptor * edesc = MTLComputePassDescriptor.computePassDescriptor;
2024-01-13 19:45:45 +01:00
edesc.dispatchType = MTLDispatchTypeSerial;
// create multiple command buffers and enqueue them
// then, we encode the graph into the command buffers in parallel
const int n_nodes = gf->n_nodes;
const int n_cb = ctx->n_cb;
const int n_nodes_per_cb = (n_nodes + n_cb - 1) / n_cb;
const bool should_capture = ctx->should_capture_next_compute;
if (should_capture) {
ctx->should_capture_next_compute = false;
MTLCaptureDescriptor * descriptor = [MTLCaptureDescriptor new];
descriptor.captureObject = ctx->queue;
NSError * error = nil;
if (![[MTLCaptureManager sharedCaptureManager] startCaptureWithDescriptor:descriptor error:&error]) {
GGML_METAL_LOG_ERROR("%s: error: unable to start capture '%s'\n", __func__, [[error localizedDescription] UTF8String]);
GGML_ASSERT(!"capture failed");
}
}
id<MTLCommandBuffer> command_buffer_builder[n_cb];
for (int cb_idx = 0; cb_idx < n_cb; ++cb_idx) {
id<MTLCommandBuffer> command_buffer = [ctx->queue commandBufferWithUnretainedReferences];
command_buffer_builder[cb_idx] = command_buffer;
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
// enqueue the command buffers in order to specify their execution order
[command_buffer enqueue];
}
const id<MTLCommandBuffer> *command_buffers = command_buffer_builder;
dispatch_apply(n_cb, ctx->d_queue, ^(size_t iter) {
const int cb_idx = iter;
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
size_t offs_src0 = 0;
size_t offs_src1 = 0;
size_t offs_src2 = 0;
size_t offs_dst = 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
id<MTLCommandBuffer> command_buffer = command_buffers[cb_idx];
id<MTLComputeCommandEncoder> encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
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 int node_start = (cb_idx + 0) * n_nodes_per_cb;
const int node_end = MIN((cb_idx == n_cb - 1) ? n_nodes : (cb_idx + 1) * n_nodes_per_cb, n_nodes);
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 i = node_start; i < node_end; ++i) {
if (i == -1) {
[encoder memoryBarrierWithScope:MTLBarrierScopeBuffers];
continue;
}
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
//GGML_METAL_LOG_INFO("%s: encoding node %3d, op = %8s\n", __func__, i, ggml_op_name(gf->nodes[i]->op));
struct ggml_tensor * src0 = gf->nodes[i]->src[0];
struct ggml_tensor * src1 = gf->nodes[i]->src[1];
struct ggml_tensor * src2 = gf->nodes[i]->src[2];
struct ggml_tensor * dst = gf->nodes[i];
llama : greatly reduce output buffer memory usage (#6122) * llama : greatly reduce logits memory usage * llama : more compact state saving and reloading * llama : fix lctx.n_outputs not being set before building graph * perplexity : adapt to the logits API changes * perplexity : fix Winogrande, use correct logits for second choice start The first logits used to evaluate the second choice were not from the end of the common prefix; instead, they were the logits from the end of the first choice. This has been corrected. The previous implementation sometimes had outliers in the scores of choices for some tasks, and the logic to skip choices words in the log-likelihood evaluation probably was an attempt to reduce those, but it was complex and didn't quite seem to be the right thing. This is simpler now, and the outlier scores aren't there anymore. * perplexity : normalize spaces and punctuation in Winogrande sentences * llama : fix embedding conditions * llama : fix llama_get_embeddings_ith when the resulting id is 0 * llama : fix wrong n_outputs in llama_set_inputs A mismatch happened when using a smaller n_ubatch than n_batch and then using llama_batch_get_one(). The decision of what n_outputs should be now almost fully depends on how lctx.n_outputs is set in llama_decode_internal. The conditions are simpler this way. * llama : when saving the state, recalculate n_outputs This ensures the correct number of outputs for the entire previous batch is stored in the session file, even when n_ubatch is smaller than n_batch. * llama : fix not-skipping outputs of non-causal models * llama : fix running a batch with n_outputs == 0 It previously worked because lctx.inp_out_ids was not initialized, so it pointed to some garbage address which was somehow still valid when I ran my tests. * llama : keep same graph topology even when n_outputs == 0 * ggml : saner ggml_can_repeat with empty tensors * ggml : future-proof ggml_is_empty by using GGML_MAX_DIMS - 1 * ggml : do not multi-thread ops returning empty tensors * ggml : make ggml_is_empty public and work with views * llama : use a vector for ctx->output_ids * llama : rework reallocation logic for llama_output_reserve Now comparing the actual size with the new total size of the output buffer to allow more efficient enabling and disabling of the embeddings and/or logits output in the future. * ggml : skip empty tensors in all backends * llama : fix llama_output_reserve nullptr deref when new_size is 0 * perplexity : make Winogrande work as it does on master The problems with the Winogrande implementation will need to be fixed in a separate PR to ease review. * llama : clearer error messages for invalid logits or embeddings ids * llama : assert all models that can have inp_out_ids Since the graph topology is now constant, this presence check can be done even when there are no outputs. * llama : assert logits and embd buffers exist before writing to them * llama : handle errors from llama_output_reserve at call sites * perplexity : make hellaswag and multiple-choice outputs identical to master Due to how the KV cache is updated, the logprobs for tokens in a batch are very slightly affected by the other tokens present in the batch, so to make hellaswag and multiple-choice return exactly the same results as on master, the last token of each sequence needs to be evaluated even though its output is not used at all. This will probably be changed back in the future to make these benchmarks a tiny bit faster. * perplexity : fix division by zero when using less than 100 multiple-choice tasks * llama : allow loading state saved with a different ctx size When loading a session file, the context size is now only required to be at least enough to load the KV cells contained in that session file, instead of requiring to use exactly the same context size as when saving. Doing this enables the use-case of extending or shrinking the context size of a saved session. This breaks existing session files because the meaning of kv_buf_size is slightly changed (previously it was the size of the whole KV cache, now it's only the size of the saved part of it). This allows for finer-grained sanity checks when loading in an effort to keep kv_buf_size useful even when the kv_size is changed. * llama : minor ggml-ci * readme : update recent API changes, and warn about Vulkan --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-03-26 15:46:41 +01:00
if (ggml_is_empty(dst)) {
continue;
}
switch (dst->op) {
case GGML_OP_NONE:
case GGML_OP_RESHAPE:
case GGML_OP_VIEW:
case GGML_OP_TRANSPOSE:
case GGML_OP_PERMUTE:
{
// noop -> next node
} continue;
default:
{
} break;
}
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
if (!ggml_metal_supports_op(ctx, dst)) {
GGML_METAL_LOG_ERROR("%s: error: unsupported op '%s'\n", __func__, ggml_op_desc(dst));
GGML_ASSERT(!"unsupported op");
}
if (should_capture) {
[encoder pushDebugGroup:[NSString stringWithCString:ggml_op_desc(dst) encoding:NSUTF8StringEncoding]];
}
const int64_t ne00 = src0 ? src0->ne[0] : 0;
const int64_t ne01 = src0 ? src0->ne[1] : 0;
const int64_t ne02 = src0 ? src0->ne[2] : 0;
const int64_t ne03 = src0 ? src0->ne[3] : 0;
const uint64_t nb00 = src0 ? src0->nb[0] : 0;
const uint64_t nb01 = src0 ? src0->nb[1] : 0;
const uint64_t nb02 = src0 ? src0->nb[2] : 0;
const uint64_t nb03 = src0 ? src0->nb[3] : 0;
const int64_t ne10 = src1 ? src1->ne[0] : 0;
const int64_t ne11 = src1 ? src1->ne[1] : 0;
const int64_t ne12 = src1 ? src1->ne[2] : 0;
const int64_t ne13 = src1 ? src1->ne[3] : 0;
const uint64_t nb10 = src1 ? src1->nb[0] : 0;
const uint64_t nb11 = src1 ? src1->nb[1] : 0;
const uint64_t nb12 = src1 ? src1->nb[2] : 0;
const uint64_t nb13 = src1 ? src1->nb[3] : 0;
const int64_t ne20 = src2 ? src2->ne[0] : 0;
const int64_t ne21 = src2 ? src2->ne[1] : 0;
const int64_t ne22 = src2 ? src2->ne[2] : 0; GGML_UNUSED(ne22);
const int64_t ne23 = src2 ? src2->ne[3] : 0; GGML_UNUSED(ne23);
const uint64_t nb20 = src2 ? src2->nb[0] : 0; GGML_UNUSED(nb20);
const uint64_t nb21 = src2 ? src2->nb[1] : 0;
const uint64_t nb22 = src2 ? src2->nb[2] : 0;
const uint64_t nb23 = src2 ? src2->nb[3] : 0;
const int64_t ne0 = dst ? dst->ne[0] : 0;
const int64_t ne1 = dst ? dst->ne[1] : 0;
const int64_t ne2 = dst ? dst->ne[2] : 0;
const int64_t ne3 = dst ? dst->ne[3] : 0;
const uint64_t nb0 = dst ? dst->nb[0] : 0;
const uint64_t nb1 = dst ? dst->nb[1] : 0;
const uint64_t nb2 = dst ? dst->nb[2] : 0;
const uint64_t nb3 = dst ? dst->nb[3] : 0;
const enum ggml_type src0t = src0 ? src0->type : GGML_TYPE_COUNT;
const enum ggml_type src1t = src1 ? src1->type : GGML_TYPE_COUNT;
const enum ggml_type dstt = dst ? dst->type : GGML_TYPE_COUNT;
id<MTLBuffer> id_src0 = src0 ? ggml_metal_get_buffer(src0, &offs_src0) : nil;
id<MTLBuffer> id_src1 = src1 ? ggml_metal_get_buffer(src1, &offs_src1) : nil;
id<MTLBuffer> id_src2 = src2 ? ggml_metal_get_buffer(src2, &offs_src2) : nil;
id<MTLBuffer> id_dst = dst ? ggml_metal_get_buffer(dst, &offs_dst) : nil;
//GGML_METAL_LOG_INFO("%s: op - %s\n", __func__, ggml_op_name(dst->op));
//if (src0) {
// GGML_METAL_LOG_INFO("%s: src0 - %4s [%5lld, %5lld, %5lld], %d, %s\n", __func__, ggml_type_name(src0t), ne00, ne01, ne02,
// ggml_is_contiguous(src0), src0->name);
//}
//if (src1) {
// GGML_METAL_LOG_INFO("%s: src1 - %4s [%5lld, %5lld, %5lld], %d, %s\n", __func__, ggml_type_name(src1t), ne10, ne11, ne12,
// ggml_is_contiguous(src1), src1->name);
//}
//if (dst) {
// GGML_METAL_LOG_INFO("%s: dst - %4s [%5lld, %5lld, %5lld], 1, %s\n", __func__, ggml_type_name(dstt), ne0, ne1, ne2,
// dst->name);
//}
switch (dst->op) {
case GGML_OP_CONCAT:
{
id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CONCAT].pipeline;
const int32_t dim = ((int32_t *) dst->op_params)[0];
[encoder setComputePipelineState:pipeline];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
[encoder setBuffer:id_dst offset:offs_dst atIndex:2];
[encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3];
[encoder setBytes:&ne01 length:sizeof(ne01) atIndex:4];
[encoder setBytes:&ne02 length:sizeof(ne02) atIndex:5];
[encoder setBytes:&ne03 length:sizeof(ne03) atIndex:6];
[encoder setBytes:&nb00 length:sizeof(nb00) atIndex:7];
[encoder setBytes:&nb01 length:sizeof(nb01) atIndex:8];
[encoder setBytes:&nb02 length:sizeof(nb02) atIndex:9];
[encoder setBytes:&nb03 length:sizeof(nb03) atIndex:10];
[encoder setBytes:&ne10 length:sizeof(ne10) atIndex:11];
[encoder setBytes:&ne11 length:sizeof(ne11) atIndex:12];
[encoder setBytes:&ne12 length:sizeof(ne12) atIndex:13];
[encoder setBytes:&ne13 length:sizeof(ne13) atIndex:14];
[encoder setBytes:&nb10 length:sizeof(nb10) atIndex:15];
[encoder setBytes:&nb11 length:sizeof(nb11) atIndex:16];
[encoder setBytes:&nb12 length:sizeof(nb12) atIndex:17];
[encoder setBytes:&nb13 length:sizeof(nb13) atIndex:18];
[encoder setBytes:&ne0 length:sizeof(ne0) atIndex:19];
[encoder setBytes:&ne1 length:sizeof(ne1) atIndex:20];
[encoder setBytes:&ne2 length:sizeof(ne2) atIndex:21];
[encoder setBytes:&ne3 length:sizeof(ne3) atIndex:22];
[encoder setBytes:&nb0 length:sizeof(nb0) atIndex:23];
[encoder setBytes:&nb1 length:sizeof(nb1) atIndex:24];
[encoder setBytes:&nb2 length:sizeof(nb2) atIndex:25];
[encoder setBytes:&nb3 length:sizeof(nb3) atIndex:26];
[encoder setBytes:&dim length:sizeof(dim) atIndex:27];
const int nth = MIN(1024, ne0);
[encoder dispatchThreadgroups:MTLSizeMake(ne1, ne2, ne3) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
} break;
case GGML_OP_ADD:
case GGML_OP_MUL:
case GGML_OP_DIV:
{
GGML_ASSERT(src0t == GGML_TYPE_F32);
GGML_ASSERT(src1t == GGML_TYPE_F32);
const size_t offs = 0;
bool bcast_row = false;
int64_t nb = ne00; // used by the "row" kernels
id<MTLComputePipelineState> pipeline = nil;
if (ggml_nelements(src1) == ne10 && ggml_is_contiguous(src1) && ne00 % 4 == 0 && ne10 % 4 == 0) {
GGML_ASSERT(ggml_is_contiguous(src0));
// src1 is a row
GGML_ASSERT(ne11 == 1);
nb = ne00 / 4;
switch (dst->op) {
case GGML_OP_ADD: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ADD_ROW].pipeline; break;
case GGML_OP_MUL: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_ROW].pipeline; break;
case GGML_OP_DIV: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_DIV_ROW].pipeline; break;
default: GGML_ASSERT(false);
}
bcast_row = true;
} else {
switch (dst->op) {
case GGML_OP_ADD: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ADD].pipeline; break;
case GGML_OP_MUL: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL].pipeline; break;
case GGML_OP_DIV: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_DIV].pipeline; break;
default: GGML_ASSERT(false);
}
}
[encoder setComputePipelineState:pipeline];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
[encoder setBuffer:id_dst offset:offs_dst atIndex:2];
[encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3];
[encoder setBytes:&ne01 length:sizeof(ne01) atIndex:4];
[encoder setBytes:&ne02 length:sizeof(ne02) atIndex:5];
[encoder setBytes:&ne03 length:sizeof(ne03) atIndex:6];
[encoder setBytes:&nb00 length:sizeof(nb00) atIndex:7];
[encoder setBytes:&nb01 length:sizeof(nb01) atIndex:8];
[encoder setBytes:&nb02 length:sizeof(nb02) atIndex:9];
[encoder setBytes:&nb03 length:sizeof(nb03) atIndex:10];
[encoder setBytes:&ne10 length:sizeof(ne10) atIndex:11];
[encoder setBytes:&ne11 length:sizeof(ne11) atIndex:12];
[encoder setBytes:&ne12 length:sizeof(ne12) atIndex:13];
[encoder setBytes:&ne13 length:sizeof(ne13) atIndex:14];
[encoder setBytes:&nb10 length:sizeof(nb10) atIndex:15];
[encoder setBytes:&nb11 length:sizeof(nb11) atIndex:16];
[encoder setBytes:&nb12 length:sizeof(nb12) atIndex:17];
[encoder setBytes:&nb13 length:sizeof(nb13) atIndex:18];
[encoder setBytes:&ne0 length:sizeof(ne0) atIndex:19];
[encoder setBytes:&ne1 length:sizeof(ne1) atIndex:20];
[encoder setBytes:&ne2 length:sizeof(ne2) atIndex:21];
[encoder setBytes:&ne3 length:sizeof(ne3) atIndex:22];
[encoder setBytes:&nb0 length:sizeof(nb0) atIndex:23];
[encoder setBytes:&nb1 length:sizeof(nb1) atIndex:24];
[encoder setBytes:&nb2 length:sizeof(nb2) atIndex:25];
[encoder setBytes:&nb3 length:sizeof(nb3) atIndex:26];
[encoder setBytes:&offs length:sizeof(offs) atIndex:27];
[encoder setBytes:&nb length:sizeof(nb) atIndex:28];
if (bcast_row) {
const int64_t n = ggml_nelements(dst)/4;
[encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
} else {
const int nth = MIN((int) pipeline.maxTotalThreadsPerThreadgroup, ne0);
[encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
}
} break;
case GGML_OP_REPEAT:
{
id<MTLComputePipelineState> pipeline;
switch (src0t) {
case GGML_TYPE_F32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_REPEAT_F32].pipeline; break;
case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_REPEAT_F16].pipeline; break;
case GGML_TYPE_I32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_REPEAT_I32].pipeline; break;
case GGML_TYPE_I16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_REPEAT_I16].pipeline; break;
default: GGML_ASSERT(false);
}
[encoder setComputePipelineState:pipeline];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_dst offset:offs_dst atIndex:1];
[encoder setBytes:&ne00 length:sizeof(ne00) atIndex:2];
[encoder setBytes:&ne01 length:sizeof(ne01) atIndex:3];
[encoder setBytes:&ne02 length:sizeof(ne02) atIndex:4];
[encoder setBytes:&ne03 length:sizeof(ne03) atIndex:5];
[encoder setBytes:&nb00 length:sizeof(nb00) atIndex:6];
[encoder setBytes:&nb01 length:sizeof(nb01) atIndex:7];
[encoder setBytes:&nb02 length:sizeof(nb02) atIndex:8];
[encoder setBytes:&nb03 length:sizeof(nb03) atIndex:9];
[encoder setBytes:&ne0 length:sizeof(ne0) atIndex:10];
[encoder setBytes:&ne1 length:sizeof(ne1) atIndex:11];
[encoder setBytes:&ne2 length:sizeof(ne2) atIndex:12];
[encoder setBytes:&ne3 length:sizeof(ne3) atIndex:13];
[encoder setBytes:&nb0 length:sizeof(nb0) atIndex:14];
[encoder setBytes:&nb1 length:sizeof(nb1) atIndex:15];
[encoder setBytes:&nb2 length:sizeof(nb2) atIndex:16];
[encoder setBytes:&nb3 length:sizeof(nb3) atIndex:17];
const int nth = MIN((int) pipeline.maxTotalThreadsPerThreadgroup, ne0);
[encoder dispatchThreadgroups:MTLSizeMake(ne1, ne2, ne3) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
} break;
case GGML_OP_ACC:
{
GGML_ASSERT(src0t == GGML_TYPE_F32);
GGML_ASSERT(src1t == GGML_TYPE_F32);
GGML_ASSERT(dstt == GGML_TYPE_F32);
GGML_ASSERT(ggml_is_contiguous(src0));
GGML_ASSERT(ggml_is_contiguous(src1));
const size_t pnb1 = ((int32_t *) dst->op_params)[0];
const size_t pnb2 = ((int32_t *) dst->op_params)[1];
const size_t pnb3 = ((int32_t *) dst->op_params)[2];
const size_t offs = ((int32_t *) dst->op_params)[3];
const bool inplace = (bool) ((int32_t *) dst->op_params)[4];
if (!inplace) {
// run a separete kernel to cpy src->dst
// not sure how to avoid this
// TODO: make a simpler cpy_bytes kernel
const id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_F32].pipeline;
[encoder setComputePipelineState:pipeline];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_dst offset:offs_dst atIndex:1];
[encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:2];
[encoder setBytes:&ne01 length:sizeof( int64_t) atIndex:3];
[encoder setBytes:&ne02 length:sizeof( int64_t) atIndex:4];
[encoder setBytes:&ne03 length:sizeof( int64_t) atIndex:5];
[encoder setBytes:&nb00 length:sizeof(uint64_t) atIndex:6];
[encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:7];
[encoder setBytes:&nb02 length:sizeof(uint64_t) atIndex:8];
[encoder setBytes:&nb03 length:sizeof(uint64_t) atIndex:9];
[encoder setBytes:&ne0 length:sizeof( int64_t) atIndex:10];
[encoder setBytes:&ne1 length:sizeof( int64_t) atIndex:11];
[encoder setBytes:&ne2 length:sizeof( int64_t) atIndex:12];
[encoder setBytes:&ne3 length:sizeof( int64_t) atIndex:13];
[encoder setBytes:&nb0 length:sizeof(uint64_t) atIndex:14];
[encoder setBytes:&nb1 length:sizeof(uint64_t) atIndex:15];
[encoder setBytes:&nb2 length:sizeof(uint64_t) atIndex:16];
[encoder setBytes:&nb3 length:sizeof(uint64_t) atIndex:17];
const int nth = MIN((int) pipeline.maxTotalThreadsPerThreadgroup, ne00);
[encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
}
const id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ADD].pipeline;
[encoder setComputePipelineState:pipeline];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
[encoder setBuffer:id_dst offset:offs_dst atIndex:2];
[encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3];
[encoder setBytes:&ne01 length:sizeof(ne01) atIndex:4];
[encoder setBytes:&ne02 length:sizeof(ne02) atIndex:5];
[encoder setBytes:&ne03 length:sizeof(ne03) atIndex:6];
[encoder setBytes:&nb00 length:sizeof(nb00) atIndex:7];
[encoder setBytes:&pnb1 length:sizeof(pnb1) atIndex:8];
[encoder setBytes:&pnb2 length:sizeof(pnb2) atIndex:9];
[encoder setBytes:&pnb3 length:sizeof(pnb3) atIndex:10];
[encoder setBytes:&ne10 length:sizeof(ne10) atIndex:11];
[encoder setBytes:&ne11 length:sizeof(ne11) atIndex:12];
[encoder setBytes:&ne12 length:sizeof(ne12) atIndex:13];
[encoder setBytes:&ne13 length:sizeof(ne13) atIndex:14];
[encoder setBytes:&nb10 length:sizeof(nb10) atIndex:15];
[encoder setBytes:&nb11 length:sizeof(nb11) atIndex:16];
[encoder setBytes:&nb12 length:sizeof(nb12) atIndex:17];
[encoder setBytes:&nb13 length:sizeof(nb13) atIndex:18];
[encoder setBytes:&ne0 length:sizeof(ne0) atIndex:19];
[encoder setBytes:&ne1 length:sizeof(ne1) atIndex:20];
[encoder setBytes:&ne2 length:sizeof(ne2) atIndex:21];
[encoder setBytes:&ne3 length:sizeof(ne3) atIndex:22];
[encoder setBytes:&nb0 length:sizeof(nb0) atIndex:23];
[encoder setBytes:&pnb1 length:sizeof(pnb1) atIndex:24];
[encoder setBytes:&pnb2 length:sizeof(pnb2) atIndex:25];
[encoder setBytes:&pnb3 length:sizeof(pnb3) atIndex:26];
[encoder setBytes:&offs length:sizeof(offs) atIndex:27];
const int nth = MIN((int) pipeline.maxTotalThreadsPerThreadgroup, ne00);
[encoder dispatchThreadgroups:MTLSizeMake(ne11, ne12, ne13) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
} break;
case GGML_OP_SCALE:
{
GGML_ASSERT(ggml_is_contiguous(src0));
float scale;
memcpy(&scale, dst->op_params, sizeof(scale));
int64_t n = ggml_nelements(dst);
id<MTLComputePipelineState> pipeline = nil;
if (n % 4 == 0) {
n /= 4;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SCALE_4].pipeline;
} else {
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SCALE].pipeline;
}
[encoder setComputePipelineState:pipeline];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_dst offset:offs_dst atIndex:1];
[encoder setBytes:&scale length:sizeof(scale) atIndex:2];
[encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
} break;
case GGML_OP_CLAMP:
2024-05-11 15:57:53 +02:00
{
id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CLAMP].pipeline;
2024-05-11 15:57:53 +02:00
float min;
float max;
memcpy(&min, ((int32_t *) dst->op_params) + 0, sizeof(float));
memcpy(&max, ((int32_t *) dst->op_params) + 1, sizeof(float));
2024-05-11 15:57:53 +02:00
[encoder setComputePipelineState:pipeline];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_dst offset:offs_dst atIndex:1];
[encoder setBytes:&min length:sizeof(min) atIndex:2];
[encoder setBytes:&max length:sizeof(max) atIndex:3];
2024-05-11 15:57:53 +02:00
const int64_t n = ggml_nelements(dst);
2024-05-11 15:57:53 +02:00
[encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
} break;
case GGML_OP_UNARY:
switch (ggml_get_unary_op(gf->nodes[i])) {
// we are not taking into account the strides, so for now require contiguous tensors
GGML_ASSERT(ggml_is_contiguous(src0));
case GGML_UNARY_OP_TANH:
{
id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_TANH].pipeline;
[encoder setComputePipelineState:pipeline];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_dst offset:offs_dst atIndex:1];
const int64_t n = ggml_nelements(dst);
[encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
} break;
case GGML_UNARY_OP_RELU:
{
id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_RELU].pipeline;
[encoder setComputePipelineState:pipeline];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_dst offset:offs_dst atIndex:1];
const int64_t n = ggml_nelements(dst);
[encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
} break;
case GGML_UNARY_OP_SIGMOID:
{
id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SIGMOID].pipeline;
[encoder setComputePipelineState:pipeline];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_dst offset:offs_dst atIndex:1];
const int64_t n = ggml_nelements(dst);
[encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
} break;
case GGML_UNARY_OP_GELU:
{
int64_t n = ggml_nelements(dst);
id<MTLComputePipelineState> pipeline = nil;
if (n % 4 == 0) {
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GELU_4].pipeline;
n /= 4;
} else {
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GELU].pipeline;
}
[encoder setComputePipelineState:pipeline];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_dst offset:offs_dst atIndex:1];
[encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
} break;
case GGML_UNARY_OP_GELU_QUICK:
{
int64_t n = ggml_nelements(dst);
id<MTLComputePipelineState> pipeline = nil;
if (n % 4 == 0) {
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GELU_QUICK_4].pipeline;
n /= 4;
} else {
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GELU_QUICK].pipeline;
}
[encoder setComputePipelineState:pipeline];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_dst offset:offs_dst atIndex:1];
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
[encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
} break;
case GGML_UNARY_OP_SILU:
{
int64_t n = ggml_nelements(dst);
id<MTLComputePipelineState> pipeline = nil;
if (n % 4 == 0) {
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SILU_4].pipeline;
n /= 4;
} else {
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SILU].pipeline;
}
[encoder setComputePipelineState:pipeline];
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
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_dst offset:offs_dst atIndex:1];
[encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
} break;
default:
{
GGML_METAL_LOG_WARN("%s: node %3d, op = %8s not implemented\n", __func__, i, ggml_op_name(dst->op));
GGML_ASSERT(false);
}
} break;
case GGML_OP_SQR:
{
GGML_ASSERT(ggml_is_contiguous(src0));
id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SQR].pipeline;
[encoder setComputePipelineState:pipeline];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_dst offset:offs_dst atIndex:1];
const int64_t n = ggml_nelements(dst);
[encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
} break;
case GGML_OP_SUM_ROWS:
{
GGML_ASSERT(src0->nb[0] == ggml_type_size(src0->type));
id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SUM_ROWS].pipeline;
[encoder setComputePipelineState:pipeline];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_dst offset:offs_dst atIndex:1];
[encoder setBytes:&ne00 length:sizeof(ne00) atIndex:2];
[encoder setBytes:&ne01 length:sizeof(ne01) atIndex:3];
[encoder setBytes:&ne02 length:sizeof(ne02) atIndex:4];
[encoder setBytes:&ne03 length:sizeof(ne03) atIndex:5];
[encoder setBytes:&nb00 length:sizeof(nb00) atIndex:6];
[encoder setBytes:&nb01 length:sizeof(nb01) atIndex:7];
[encoder setBytes:&nb02 length:sizeof(nb02) atIndex:8];
[encoder setBytes:&nb03 length:sizeof(nb03) atIndex:9];
[encoder setBytes:&ne10 length:sizeof(ne10) atIndex:10];
[encoder setBytes:&ne11 length:sizeof(ne11) atIndex:11];
[encoder setBytes:&ne12 length:sizeof(ne12) atIndex:12];
[encoder setBytes:&ne13 length:sizeof(ne13) atIndex:13];
[encoder setBytes:&nb10 length:sizeof(nb10) atIndex:14];
[encoder setBytes:&nb11 length:sizeof(nb11) atIndex:15];
[encoder setBytes:&nb12 length:sizeof(nb12) atIndex:16];
[encoder setBytes:&nb13 length:sizeof(nb13) atIndex:17];
[encoder setBytes:&ne0 length:sizeof(ne0) atIndex:18];
[encoder setBytes:&ne1 length:sizeof(ne1) atIndex:19];
[encoder setBytes:&ne2 length:sizeof(ne2) atIndex:20];
[encoder setBytes:&ne3 length:sizeof(ne3) atIndex:21];
[encoder setBytes:&nb0 length:sizeof(nb0) atIndex:22];
[encoder setBytes:&nb1 length:sizeof(nb1) atIndex:23];
[encoder setBytes:&nb2 length:sizeof(nb2) atIndex:24];
[encoder setBytes:&nb3 length:sizeof(nb3) atIndex:25];
[encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
} break;
case GGML_OP_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
GGML_ASSERT(!src1 || src1->type == GGML_TYPE_F16 || src1->type == GGML_TYPE_F32);
int nth = 32; // SIMD width
id<MTLComputePipelineState> pipeline = nil;
const bool use_f16 = (src1 && src1->type == GGML_TYPE_F16);
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
if (ne00%4 == 0) {
while (nth < ne00/4 && nth*ne01*ne02*ne03 < 256) {
nth *= 2;
}
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
if (use_f16) {
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SOFT_MAX_F16_4].pipeline;
} else {
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SOFT_MAX_F32_4].pipeline;
}
} else {
while (nth < ne00 && nth*ne01*ne02*ne03 < 256) {
nth *= 2;
}
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
if (use_f16) {
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SOFT_MAX_F16].pipeline;
} else {
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SOFT_MAX_F32].pipeline;
}
}
float scale;
float max_bias;
memcpy(&scale, ((int32_t *) dst->op_params) + 0, sizeof(scale));
memcpy(&max_bias, ((int32_t *) dst->op_params) + 1, sizeof(max_bias));
const int64_t nrows_x = ggml_nrows(src0);
const int64_t nrows_y = src0->ne[1];
const uint32_t n_head = nrows_x/nrows_y;
const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head));
const float m0 = powf(2.0f, -(max_bias ) / n_head_log2);
const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
[encoder setComputePipelineState:pipeline];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
if (id_src1) {
[encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
} else {
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:1];
}
[encoder setBuffer:id_dst offset:offs_dst atIndex:2];
[encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3];
[encoder setBytes:&ne01 length:sizeof(ne01) atIndex:4];
[encoder setBytes:&ne02 length:sizeof(ne02) atIndex:5];
[encoder setBytes:&scale length:sizeof(scale) atIndex:6];
[encoder setBytes:&max_bias length:sizeof(max_bias) atIndex:7];
[encoder setBytes:&m0 length:sizeof(m0) atIndex:8];
[encoder setBytes:&m1 length:sizeof(m1) atIndex:9];
[encoder setBytes:&n_head_log2 length:sizeof(n_head_log2) atIndex:10];
[encoder setThreadgroupMemoryLength:32*sizeof(float) atIndex:0];
[encoder dispatchThreadgroups:MTLSizeMake(ne01*ne02*ne03, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
} break;
case GGML_OP_DIAG_MASK_INF:
{
const int n_past = ((int32_t *)(dst->op_params))[0];
id<MTLComputePipelineState> pipeline = nil;
if (ne00%8 == 0) {
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_DIAG_MASK_INF_8].pipeline;
} else {
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_DIAG_MASK_INF].pipeline;
}
[encoder setComputePipelineState:pipeline];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_dst offset:offs_dst atIndex:1];
[encoder setBytes:&ne00 length:sizeof(ne00) atIndex:2];
[encoder setBytes:&ne01 length:sizeof(ne01) atIndex:3];
[encoder setBytes:&n_past length:sizeof(int) atIndex:4];
2023-06-17 19:24:11 +02:00
if (ne00%8 == 0) {
[encoder dispatchThreadgroups:MTLSizeMake(ne00*ne01*ne02/8, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
}
else {
[encoder dispatchThreadgroups:MTLSizeMake(ne00, ne01, ne02) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
}
} break;
case GGML_OP_MUL_MAT:
{
GGML_ASSERT(ne00 == ne10);
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
GGML_ASSERT(ne12 % ne02 == 0);
GGML_ASSERT(ne13 % ne03 == 0);
2023-06-17 19:24:11 +02:00
const uint r2 = ne12/ne02;
const uint r3 = ne13/ne03;
2023-06-17 19:24:11 +02:00
// find the break-even point where the matrix-matrix kernel becomes more efficient compared
// to the matrix-vector kernel
int ne11_mm_min = 1;
#if 0
// the numbers below are measured on M2 Ultra for 7B and 13B models
// these numbers do not translate to other devices or model sizes
// TODO: need to find a better approach
if ([ctx->device.name isEqualToString:@"Apple M2 Ultra"]) {
switch (src0t) {
case GGML_TYPE_F16: ne11_mm_min = 2; break;
case GGML_TYPE_Q8_0: ne11_mm_min = 7; break;
case GGML_TYPE_Q2_K: ne11_mm_min = 15; break;
case GGML_TYPE_Q3_K: ne11_mm_min = 7; break;
case GGML_TYPE_Q4_0:
case GGML_TYPE_Q4_1: ne11_mm_min = 15; break;
case GGML_TYPE_Q4_K: ne11_mm_min = 11; break;
case GGML_TYPE_Q5_0: // not tested yet
case GGML_TYPE_Q5_1: ne11_mm_min = 13; break; // not tested yet
case GGML_TYPE_Q5_K: ne11_mm_min = 7; break;
case GGML_TYPE_Q6_K: ne11_mm_min = 7; break;
default: ne11_mm_min = 1; break;
}
}
#endif
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
// for now the matrix-matrix multiplication kernel only works on A14+/M1+ SoCs
// AMD GPU and older A-chips will reuse matrix-vector multiplication kernel
if ([ctx->device supportsFamily:MTLGPUFamilyApple7] &&
!ggml_is_transposed(src0) &&
!ggml_is_transposed(src1) &&
src1t == GGML_TYPE_F32 &&
ne00 % 32 == 0 && ne00 >= 64 &&
(ne11 > ne11_mm_min || (ggml_is_quantized(src0t) && ne12 > 1))) {
//printf("matrix: ne00 = %6d, ne01 = %6d, ne02 = %6d, ne11 = %6d, ne12 = %6d\n", ne00, ne01, ne02, ne11, ne12);
llama : add custom RoPE (#2054) * Implement customizable RoPE The original RoPE has pre-defined parameters theta_i = 10000^(−2(i−1)/d), for i in [1, 2, ..., d/2] Our customizable RoPE, ggml_rope_custom_inplace, uses theta_i = scale * base^(−2(i−1)/d), for i in [1, 2, ..., d/2] with the default matches the original scale = 1.0 base = 10000 The new command line arguments --rope-freq-base --rope-freq-scale set the two new RoPE parameter. Recent researches show changing these two parameters extends the context limit with minimal loss. 1. Extending Context to 8K kaiokendev https://kaiokendev.github.io/til#extending-context-to-8k 2. Extending Context Window of Large Language Models via Positional Interpolation Shouyuan Chen, Sherman Wong, Liangjian Chen, Yuandong Tian https://arxiv.org/abs/2306.15595 3. NTK-Aware Scaled RoPE allows LLaMA models to have extended (8k+) context size without any fine-tuning and minimal perplexity degradation. https://www.reddit.com/user/bloc97 https://www.reddit.com/r/LocalLLaMA/comments/14lz7j5/ntkaware_scaled_rope_allows_llama_models_to_have/ For the bold, try adding the following command line parameters to your favorite model: -c 16384 --rope-freq-base 80000 --rope-freq-scale 0.5 * ggml-metal: fix custom rope * common: fix argument names in help * llama: increase MEM_REQ_EVAL for MODEL_3B It avoids crashing for quantized weights on CPU. Better ways to calculate the required buffer size would be better. * llama: make MEM_REQ_EVAL depend on n_ctx * server: use proper Content-Type in curl examples Without the header Content-Type: application/json, curl will POST with Content-Type: application/x-www-form-urlencoded Though our simple server doesn't care, the httplib.h used has a limit with CPPHTTPLIB_FORM_URL_ENCODED_PAYLOAD_MAX_LENGTH 8192 With Content-Type: application/json, we can send large json data. * style : minor fixes, mostly indentations * ggml : fix asserts --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-07-15 12:34:16 +02:00
// some Metal matrix data types require aligned pointers
// ref: https://developer.apple.com/metal/Metal-Shading-Language-Specification.pdf (Table 2.5)
switch (src0->type) {
case GGML_TYPE_F32: GGML_ASSERT(nb01 % 16 == 0); break;
case GGML_TYPE_F16: GGML_ASSERT(nb01 % 8 == 0); break;
default: break;
}
id<MTLComputePipelineState> pipeline = nil;
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
switch (src0->type) {
case GGML_TYPE_F32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_F32_F32 ].pipeline; break;
case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_F16_F32 ].pipeline; break;
case GGML_TYPE_Q4_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_0_F32 ].pipeline; break;
case GGML_TYPE_Q4_1: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_1_F32 ].pipeline; break;
case GGML_TYPE_Q5_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_0_F32 ].pipeline; break;
case GGML_TYPE_Q5_1: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_1_F32 ].pipeline; break;
case GGML_TYPE_Q8_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q8_0_F32 ].pipeline; break;
case GGML_TYPE_Q2_K: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q2_K_F32 ].pipeline; break;
case GGML_TYPE_Q3_K: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q3_K_F32 ].pipeline; break;
case GGML_TYPE_Q4_K: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_K_F32 ].pipeline; break;
case GGML_TYPE_Q5_K: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_K_F32 ].pipeline; break;
case GGML_TYPE_Q6_K: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q6_K_F32 ].pipeline; break;
case GGML_TYPE_IQ2_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XXS_F32].pipeline; break;
case GGML_TYPE_IQ2_XS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XS_F32 ].pipeline; break;
case GGML_TYPE_IQ3_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_XXS_F32].pipeline; break;
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
case GGML_TYPE_IQ3_S: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_S_F32 ].pipeline; break;
case GGML_TYPE_IQ2_S: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_S_F32 ].pipeline; break;
case GGML_TYPE_IQ1_S: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_S_F32 ].pipeline; break;
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
case GGML_TYPE_IQ1_M: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_M_F32 ].pipeline; break;
case GGML_TYPE_IQ4_NL: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_NL_F32 ].pipeline; break;
case GGML_TYPE_IQ4_XS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_XS_F32 ].pipeline; break;
default: GGML_ASSERT(false && "MUL MAT-MAT not implemented");
}
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
[encoder setComputePipelineState:pipeline];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
[encoder setBuffer:id_dst offset:offs_dst atIndex:2];
[encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3];
[encoder setBytes:&ne02 length:sizeof(ne02) atIndex:4];
[encoder setBytes:&nb01 length:sizeof(nb01) atIndex:5];
[encoder setBytes:&nb02 length:sizeof(nb02) atIndex:6];
[encoder setBytes:&ne12 length:sizeof(ne12) atIndex:7];
[encoder setBytes:&nb10 length:sizeof(nb10) atIndex:8];
[encoder setBytes:&nb11 length:sizeof(nb11) atIndex:9];
[encoder setBytes:&nb12 length:sizeof(nb12) atIndex:10];
[encoder setBytes:&ne0 length:sizeof(ne0) atIndex:11];
[encoder setBytes:&ne1 length:sizeof(ne1) atIndex:12];
[encoder setBytes:&r2 length:sizeof(r2) atIndex:13];
[encoder setBytes:&r3 length:sizeof(r3) atIndex:14];
[encoder setThreadgroupMemoryLength:8192 atIndex:0];
[encoder dispatchThreadgroups:MTLSizeMake( (ne11 + 31)/32, (ne01 + 63)/64, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(128, 1, 1)];
} else {
int nth0 = 32;
int nth1 = 1;
int nrows = 1;
//printf("vector: ne00 = %6d, ne01 = %6d, ne02 = %6d, ne11 = %6d, ne12 = %6d\n", ne00, ne01, ne02, ne11, ne12);
id<MTLComputePipelineState> pipeline = nil;
// use custom matrix x vector kernel
switch (src0t) {
case GGML_TYPE_F32:
{
GGML_ASSERT(src1t == GGML_TYPE_F32);
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F32_F32].pipeline;
nrows = 4;
} break;
case GGML_TYPE_F16:
{
nth0 = 32;
nth1 = 1;
if (src1t == GGML_TYPE_F32) {
if (ne11 * ne12 < 4) {
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_1ROW].pipeline;
} else if (ne00 >= 128 && ne01 >= 8 && ne00%4 == 0) {
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_L4].pipeline;
nrows = ne11;
} else {
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32].pipeline;
nrows = 4;
}
} else {
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F16].pipeline;
nrows = 4;
}
} break;
case GGML_TYPE_Q4_0:
{
nth0 = 8;
nth1 = 8;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_0_F32].pipeline;
} break;
case GGML_TYPE_Q4_1:
{
nth0 = 8;
nth1 = 8;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_1_F32].pipeline;
} break;
case GGML_TYPE_Q5_0:
{
nth0 = 8;
nth1 = 8;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_0_F32].pipeline;
} break;
case GGML_TYPE_Q5_1:
{
nth0 = 8;
nth1 = 8;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_1_F32].pipeline;
} break;
case GGML_TYPE_Q8_0:
{
nth0 = 8;
nth1 = 8;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q8_0_F32].pipeline;
} break;
case GGML_TYPE_Q2_K:
{
nth0 = 2;
nth1 = 32;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q2_K_F32].pipeline;
} break;
case GGML_TYPE_Q3_K:
{
nth0 = 2;
nth1 = 32;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q3_K_F32].pipeline;
} break;
case GGML_TYPE_Q4_K:
{
nth0 = 4; //1;
nth1 = 8; //32;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_K_F32].pipeline;
} break;
case GGML_TYPE_Q5_K:
{
nth0 = 2;
nth1 = 32;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_K_F32].pipeline;
} break;
case GGML_TYPE_Q6_K:
{
nth0 = 2;
nth1 = 32;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q6_K_F32].pipeline;
} break;
case GGML_TYPE_IQ2_XXS:
{
nth0 = 4;
nth1 = 16;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XXS_F32].pipeline;
} break;
case GGML_TYPE_IQ2_XS:
{
nth0 = 4;
nth1 = 16;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XS_F32].pipeline;
} break;
case GGML_TYPE_IQ3_XXS:
{
nth0 = 4;
nth1 = 16;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_XXS_F32].pipeline;
} break;
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
case GGML_TYPE_IQ3_S:
{
nth0 = 4;
nth1 = 16;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_S_F32].pipeline;
} break;
case GGML_TYPE_IQ2_S:
{
nth0 = 4;
nth1 = 16;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_S_F32].pipeline;
} break;
case GGML_TYPE_IQ1_S:
{
nth0 = 4;
nth1 = 16;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_S_F32].pipeline;
} break;
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
case GGML_TYPE_IQ1_M:
{
nth0 = 4;
nth1 = 16;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_M_F32].pipeline;
} break;
case GGML_TYPE_IQ4_NL:
{
nth0 = 4;
nth1 = 16;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_NL_F32].pipeline;
} break;
case GGML_TYPE_IQ4_XS:
{
nth0 = 4;
nth1 = 16;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_XS_F32].pipeline;
} break;
default:
{
GGML_METAL_LOG_ERROR("Asserting on type %d\n", (int)src0t);
GGML_ASSERT(false && "not implemented");
}
};
if (ggml_is_quantized(src0t)) {
GGML_ASSERT(ne00 >= nth0*nth1);
}
[encoder setComputePipelineState:pipeline];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
[encoder setBuffer:id_dst offset:offs_dst atIndex:2];
[encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3];
[encoder setBytes:&ne01 length:sizeof(ne01) atIndex:4];
[encoder setBytes:&ne02 length:sizeof(ne02) atIndex:5];
[encoder setBytes:&nb00 length:sizeof(nb00) atIndex:6];
[encoder setBytes:&nb01 length:sizeof(nb01) atIndex:7];
[encoder setBytes:&nb02 length:sizeof(nb02) atIndex:8];
[encoder setBytes:&ne10 length:sizeof(ne10) atIndex:9];
[encoder setBytes:&ne11 length:sizeof(ne11) atIndex:10];
[encoder setBytes:&ne12 length:sizeof(ne12) atIndex:11];
[encoder setBytes:&nb10 length:sizeof(nb10) atIndex:12];
[encoder setBytes:&nb11 length:sizeof(nb11) atIndex:13];
[encoder setBytes:&nb12 length:sizeof(nb12) atIndex:14];
[encoder setBytes:&ne0 length:sizeof(ne0) atIndex:15];
[encoder setBytes:&ne1 length:sizeof(ne1) atIndex:16];
[encoder setBytes:&r2 length:sizeof(r2) atIndex:17];
[encoder setBytes:&r3 length:sizeof(r3) atIndex:18];
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
if (src0t == GGML_TYPE_Q4_0 || src0t == GGML_TYPE_Q4_1 || src0t == GGML_TYPE_Q5_0 ||
src0t == GGML_TYPE_Q5_1 || src0t == GGML_TYPE_Q8_0 || src0t == GGML_TYPE_Q2_K ||
src0t == GGML_TYPE_IQ1_S || src0t == GGML_TYPE_IQ1_M || src0t == GGML_TYPE_IQ2_S) {
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
}
else if (src0t == GGML_TYPE_IQ2_XXS || src0t == GGML_TYPE_IQ2_XS) {
const int mem_size = src0t == GGML_TYPE_IQ2_XXS ? 256*8+128 : 512*8+128;
[encoder setThreadgroupMemoryLength:mem_size atIndex:0];
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 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
else if (src0t == GGML_TYPE_IQ3_XXS || src0t == GGML_TYPE_IQ3_S) {
const int mem_size = src0t == GGML_TYPE_IQ3_XXS ? 256*4+128 : 512*4;
[encoder setThreadgroupMemoryLength:mem_size atIndex:0];
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
}
else if (src0t == GGML_TYPE_IQ4_NL || src0t == GGML_TYPE_IQ4_XS) {
const int mem_size = 32*sizeof(float);
[encoder setThreadgroupMemoryLength:mem_size atIndex:0];
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
}
else if (src0t == GGML_TYPE_Q4_K) {
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
}
else if (src0t == GGML_TYPE_Q3_K) {
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
}
else if (src0t == GGML_TYPE_Q5_K) {
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
}
else if (src0t == GGML_TYPE_Q6_K) {
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 1)/2, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
} else {
const int64_t ny = (ne11 + nrows - 1)/nrows;
[encoder dispatchThreadgroups:MTLSizeMake(ne01, ny, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
}
}
} break;
case GGML_OP_MUL_MAT_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
const int n_as = src0->ne[2];
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
// src2 = ids
const enum ggml_type src2t = src2->type; GGML_UNUSED(src2t);
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_ASSERT(src2t == GGML_TYPE_I32);
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_ASSERT(!ggml_is_transposed(src0));
GGML_ASSERT(!ggml_is_transposed(src1));
GGML_ASSERT(src1t == GGML_TYPE_F32);
// find the break-even point where the matrix-matrix kernel becomes more efficient compared
// to the matrix-vector kernel
// ne20 = n_used_experts
// ne21 = n_rows
const int dst_rows = ne20*ne21;
const int dst_rows_min = n_as;
const int dst_rows_max = (ctx->device.maxThreadgroupMemoryLength - 32 - 8192)/4;
// max size of the rowids array in the kernel shared buffer
GGML_ASSERT(dst_rows <= dst_rows_max);
// for now the matrix-matrix multiplication kernel only works on A14+/M1+ SoCs
// AMD GPU and older A-chips will reuse matrix-vector multiplication kernel
// !!!
// TODO: for now, always use mat-vec kernels until we figure out how to improve the
// indirect matrix multiplication
// !!!
if ([ctx->device supportsFamily:MTLGPUFamilyApple7] &&
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
ne00 % 32 == 0 && ne00 >= 64 &&
dst_rows > dst_rows_min) {
// some Metal matrix data types require aligned pointers
// ref: https://developer.apple.com/metal/Metal-Shading-Language-Specification.pdf (Table 2.5)
switch (src0->type) {
case GGML_TYPE_F32: GGML_ASSERT(nb01 % 16 == 0); break;
case GGML_TYPE_F16: GGML_ASSERT(nb01 % 8 == 0); break;
default: break;
}
id<MTLComputePipelineState> pipeline = nil;
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
switch (src0->type) {
case GGML_TYPE_F32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F32_F32 ].pipeline; break;
case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F16_F32 ].pipeline; break;
case GGML_TYPE_Q4_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_0_F32 ].pipeline; break;
case GGML_TYPE_Q4_1: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_1_F32 ].pipeline; break;
case GGML_TYPE_Q5_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_0_F32 ].pipeline; break;
case GGML_TYPE_Q5_1: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_1_F32 ].pipeline; break;
case GGML_TYPE_Q8_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q8_0_F32 ].pipeline; break;
case GGML_TYPE_Q2_K: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q2_K_F32 ].pipeline; break;
case GGML_TYPE_Q3_K: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q3_K_F32 ].pipeline; break;
case GGML_TYPE_Q4_K: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_K_F32 ].pipeline; break;
case GGML_TYPE_Q5_K: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_K_F32 ].pipeline; break;
case GGML_TYPE_Q6_K: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q6_K_F32 ].pipeline; break;
case GGML_TYPE_IQ2_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XXS_F32].pipeline; break;
case GGML_TYPE_IQ2_XS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XS_F32 ].pipeline; break;
case GGML_TYPE_IQ3_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_XXS_F32].pipeline; break;
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
case GGML_TYPE_IQ3_S: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_S_F32 ].pipeline; break;
case GGML_TYPE_IQ2_S: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_S_F32 ].pipeline; break;
case GGML_TYPE_IQ1_S: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_S_F32 ].pipeline; break;
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
case GGML_TYPE_IQ1_M: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_M_F32 ].pipeline; break;
case GGML_TYPE_IQ4_NL: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_NL_F32 ].pipeline; break;
case GGML_TYPE_IQ4_XS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_XS_F32 ].pipeline; break;
default: GGML_ASSERT(false && "MUL_MAT_ID not implemented");
}
[encoder setComputePipelineState:pipeline];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
[encoder setBuffer:id_dst offset:offs_dst atIndex:2];
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
[encoder setBuffer:id_src2 offset:offs_src2 atIndex:3];
[encoder setBytes:&ne20 length:sizeof(ne20) atIndex:4];
[encoder setBytes:&ne21 length:sizeof(ne21) atIndex:5];
[encoder setBytes:&nb21 length:sizeof(nb21) atIndex:6];
[encoder setBytes:&ne00 length:sizeof(ne00) atIndex:7];
[encoder setBytes:&ne02 length:sizeof(ne02) atIndex:8];
[encoder setBytes:&nb01 length:sizeof(nb01) atIndex:9];
[encoder setBytes:&nb02 length:sizeof(nb02) atIndex:10];
[encoder setBytes:&ne11 length:sizeof(ne11) atIndex:11];
[encoder setBytes:&ne12 length:sizeof(ne12) atIndex:12];
[encoder setBytes:&ne13 length:sizeof(ne13) atIndex:13];
[encoder setBytes:&nb10 length:sizeof(nb10) atIndex:14];
[encoder setBytes:&nb11 length:sizeof(nb11) atIndex:15];
[encoder setBytes:&nb12 length:sizeof(nb12) atIndex:16];
[encoder setBytes:&ne0 length:sizeof(ne0) atIndex:17];
[encoder setBytes:&ne1 length:sizeof(ne1) atIndex:18];
[encoder setBytes:&nb1 length:sizeof(nb1) atIndex:19];
[encoder setThreadgroupMemoryLength:GGML_PAD(8192 + dst_rows*4/*sizeof(ushort2)*/, 16) atIndex:0];
[encoder dispatchThreadgroups:MTLSizeMake((ne21 + 31)/32, (ne01 + 63)/64, n_as) threadsPerThreadgroup:MTLSizeMake(128, 1, 1)];
} else {
int nth0 = 32;
int nth1 = 1;
int nrows = 1;
//printf("vector: ne00 = %6d, ne01 = %6d, ne02 = %6d, ne11 = %6d, ne12 = %6d\n", ne00, ne01, ne02, ne11, ne12);
id<MTLComputePipelineState> pipeline = nil;
// use custom matrix x vector kernel
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
switch (src0t) {
case GGML_TYPE_F32:
{
GGML_ASSERT(src1t == GGML_TYPE_F32);
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F32_F32].pipeline;
} break;
case GGML_TYPE_F16:
{
GGML_ASSERT(src1t == GGML_TYPE_F32);
nth0 = 32;
nth1 = 1;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32].pipeline;
} break;
case GGML_TYPE_Q4_0:
{
nth0 = 8;
nth1 = 8;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_0_F32].pipeline;
} break;
case GGML_TYPE_Q4_1:
{
nth0 = 8;
nth1 = 8;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_1_F32].pipeline;
} break;
case GGML_TYPE_Q5_0:
{
nth0 = 8;
nth1 = 8;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_0_F32].pipeline;
} break;
case GGML_TYPE_Q5_1:
{
nth0 = 8;
nth1 = 8;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_1_F32].pipeline;
} break;
case GGML_TYPE_Q8_0:
{
nth0 = 8;
nth1 = 8;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q8_0_F32].pipeline;
} break;
case GGML_TYPE_Q2_K:
{
nth0 = 2;
nth1 = 32;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q2_K_F32].pipeline;
} break;
case GGML_TYPE_Q3_K:
{
nth0 = 2;
nth1 = 32;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q3_K_F32].pipeline;
} break;
case GGML_TYPE_Q4_K:
{
nth0 = 4; //1;
nth1 = 8; //32;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_K_F32].pipeline;
} break;
case GGML_TYPE_Q5_K:
{
nth0 = 2;
nth1 = 32;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_K_F32].pipeline;
} break;
case GGML_TYPE_Q6_K:
{
nth0 = 2;
nth1 = 32;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q6_K_F32].pipeline;
} break;
case GGML_TYPE_IQ2_XXS:
{
nth0 = 4;
nth1 = 16;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_XXS_F32].pipeline;
} break;
case GGML_TYPE_IQ2_XS:
{
nth0 = 4;
nth1 = 16;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_XS_F32].pipeline;
} break;
case GGML_TYPE_IQ3_XXS:
{
nth0 = 4;
nth1 = 16;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_XXS_F32].pipeline;
} break;
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
case GGML_TYPE_IQ3_S:
{
nth0 = 4;
nth1 = 16;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_S_F32].pipeline;
} break;
case GGML_TYPE_IQ2_S:
{
nth0 = 4;
nth1 = 16;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_S_F32].pipeline;
} break;
case GGML_TYPE_IQ1_S:
{
nth0 = 4;
nth1 = 16;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ1_S_F32].pipeline;
} break;
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
case GGML_TYPE_IQ1_M:
{
nth0 = 4;
nth1 = 16;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ1_M_F32].pipeline;
} break;
case GGML_TYPE_IQ4_NL:
{
nth0 = 4;
nth1 = 16;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_NL_F32].pipeline;
} break;
case GGML_TYPE_IQ4_XS:
{
nth0 = 4;
nth1 = 16;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_XS_F32].pipeline;
} break;
default:
{
GGML_METAL_LOG_ERROR("Asserting on type %d\n", (int)src2t);
GGML_ASSERT(false && "not implemented");
}
};
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 (ggml_is_quantized(src0t)) {
GGML_ASSERT(ne00 >= nth0*nth1);
}
[encoder setComputePipelineState:pipeline];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
[encoder setBuffer:id_dst offset:offs_dst atIndex:2];
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
[encoder setBuffer:id_src2 offset:offs_src2 atIndex:3];
[encoder setBytes:&ne20 length:sizeof(ne20) atIndex:4];
[encoder setBytes:&ne21 length:sizeof(ne21) atIndex:5];
[encoder setBytes:&nb21 length:sizeof(nb21) atIndex:6];
[encoder setBytes:&ne00 length:sizeof(ne00) atIndex:7];
[encoder setBytes:&ne01 length:sizeof(ne01) atIndex:8];
[encoder setBytes:&ne02 length:sizeof(ne02) atIndex:9];
[encoder setBytes:&nb00 length:sizeof(nb00) atIndex:10];
[encoder setBytes:&nb01 length:sizeof(nb01) atIndex:11];
[encoder setBytes:&nb02 length:sizeof(nb02) atIndex:12];
[encoder setBytes:&ne10 length:sizeof(ne10) atIndex:13];
[encoder setBytes:&ne11 length:sizeof(ne11) atIndex:14];
[encoder setBytes:&ne12 length:sizeof(ne12) atIndex:15];
[encoder setBytes:&ne13 length:sizeof(ne13) atIndex:16];
[encoder setBytes:&nb10 length:sizeof(nb10) atIndex:17];
[encoder setBytes:&nb11 length:sizeof(nb11) atIndex:18];
[encoder setBytes:&nb12 length:sizeof(nb12) atIndex:19];
[encoder setBytes:&ne0 length:sizeof(ne0) atIndex:20];
[encoder setBytes:&ne1 length:sizeof(ne1) atIndex:21];
[encoder setBytes:&nb1 length:sizeof(nb1) atIndex:22];
const int64_t _ne1 = 1;
const int tgz = dst_rows;
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 (src0t == GGML_TYPE_Q4_0 || src0t == GGML_TYPE_Q4_1 || src0t == GGML_TYPE_Q5_0 ||
src0t == GGML_TYPE_Q5_1 || src0t == GGML_TYPE_Q8_0 || src0t == GGML_TYPE_Q2_K ||
src0t == GGML_TYPE_IQ1_S || src0t == GGML_TYPE_IQ1_M || src0t == GGML_TYPE_IQ2_S) {
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 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
else if (src0t == GGML_TYPE_IQ2_XXS || src0t == GGML_TYPE_IQ2_XS) {
const int mem_size = src0t == GGML_TYPE_IQ2_XXS ? 256*8+128 : 512*8+128;
[encoder setThreadgroupMemoryLength:mem_size atIndex:0];
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 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
else if (src0t == GGML_TYPE_IQ3_XXS || src0t == GGML_TYPE_IQ3_S) {
const int mem_size = src0t == GGML_TYPE_IQ3_XXS ? 256*4+128 : 512*4;
[encoder setThreadgroupMemoryLength:mem_size atIndex:0];
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 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
else if (src0t == GGML_TYPE_IQ4_NL || src0t == GGML_TYPE_IQ4_XS) {
const int mem_size = 32*sizeof(float);
[encoder setThreadgroupMemoryLength:mem_size atIndex:0];
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 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
else if (src0t == GGML_TYPE_Q4_K) {
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 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
else if (src0t == GGML_TYPE_Q3_K) {
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 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
else if (src0t == GGML_TYPE_Q5_K) {
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 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
else if (src0t == GGML_TYPE_Q6_K) {
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 1)/2, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
} else {
const int64_t ny = (_ne1 + nrows - 1)/nrows; // = _ne1
[encoder dispatchThreadgroups:MTLSizeMake(ne01, ny, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
}
}
} break;
case GGML_OP_GET_ROWS:
{
id<MTLComputePipelineState> pipeline = nil;
switch (src0->type) {
case GGML_TYPE_F32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_F32 ].pipeline; break;
case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_F16 ].pipeline; break;
case GGML_TYPE_Q4_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_Q4_0 ].pipeline; break;
case GGML_TYPE_Q4_1: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_Q4_1 ].pipeline; break;
case GGML_TYPE_Q5_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_Q5_0 ].pipeline; break;
case GGML_TYPE_Q5_1: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_Q5_1 ].pipeline; break;
case GGML_TYPE_Q8_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_Q8_0 ].pipeline; break;
case GGML_TYPE_Q2_K: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_Q2_K ].pipeline; break;
case GGML_TYPE_Q3_K: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_Q3_K ].pipeline; break;
case GGML_TYPE_Q4_K: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_Q4_K ].pipeline; break;
case GGML_TYPE_Q5_K: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_Q5_K ].pipeline; break;
case GGML_TYPE_Q6_K: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_Q6_K ].pipeline; break;
case GGML_TYPE_IQ2_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XXS].pipeline; break;
case GGML_TYPE_IQ2_XS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XS ].pipeline; break;
case GGML_TYPE_IQ3_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ3_XXS].pipeline; break;
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
case GGML_TYPE_IQ3_S: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ3_S ].pipeline; break;
case GGML_TYPE_IQ2_S: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_S ].pipeline; break;
case GGML_TYPE_IQ1_S: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ1_S ].pipeline; break;
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
case GGML_TYPE_IQ1_M: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ1_M ].pipeline; break;
case GGML_TYPE_IQ4_NL: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_NL ].pipeline; break;
case GGML_TYPE_IQ4_XS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_XS ].pipeline; break;
case GGML_TYPE_I32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_I32 ].pipeline; break;
default: GGML_ASSERT(false && "not implemented");
}
[encoder setComputePipelineState:pipeline];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
[encoder setBuffer:id_dst offset:offs_dst atIndex:2];
[encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:3];
[encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:4];
[encoder setBytes:&nb02 length:sizeof(uint64_t) atIndex:5];
[encoder setBytes:&ne10 length:sizeof( int64_t) atIndex:6];
[encoder setBytes:&nb10 length:sizeof( int64_t) atIndex:7];
[encoder setBytes:&nb11 length:sizeof( int64_t) atIndex:8];
[encoder setBytes:&nb1 length:sizeof(uint64_t) atIndex:9];
[encoder setBytes:&nb2 length:sizeof(uint64_t) atIndex:10];
[encoder dispatchThreadgroups:MTLSizeMake(ne10, ne11, 1) threadsPerThreadgroup:MTLSizeMake(32, 1, 1)];
} break;
case GGML_OP_RMS_NORM:
{
GGML_ASSERT(ne00 % 4 == 0);
GGML_ASSERT(ggml_is_contiguous_1(src0));
float eps;
memcpy(&eps, dst->op_params, sizeof(float));
int nth = 32; // SIMD width
while (nth < ne00/4 && nth < 1024) {
nth *= 2;
}
id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_RMS_NORM].pipeline;
[encoder setComputePipelineState:pipeline];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_dst offset:offs_dst atIndex:1];
[encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:2];
[encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:3];
[encoder setBytes:&eps length:sizeof( float) atIndex:4];
[encoder setThreadgroupMemoryLength:32*sizeof(float) atIndex:0];
const int64_t nrows = ggml_nrows(src0);
[encoder dispatchThreadgroups:MTLSizeMake(nrows, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
} break;
case GGML_OP_GROUP_NORM:
{
GGML_ASSERT(ne00 % 4 == 0);
GGML_ASSERT(ggml_is_contiguous(src0));
//float eps;
//memcpy(&eps, dst->op_params, sizeof(float));
const float eps = 1e-6f; // TODO: temporarily hardcoded
const int32_t n_groups = ((int32_t *) dst->op_params)[0];
int nth = 32; // SIMD width
//while (nth < ne00/4 && nth < 1024) {
// nth *= 2;
//}
id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GROUP_NORM].pipeline;
[encoder setComputePipelineState:pipeline];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_dst offset:offs_dst atIndex:1];
[encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:2];
[encoder setBytes:&ne01 length:sizeof( int64_t) atIndex:3];
[encoder setBytes:&ne02 length:sizeof( int64_t) atIndex:4];
[encoder setBytes:&nb00 length:sizeof(uint64_t) atIndex:5];
[encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:6];
[encoder setBytes:&nb02 length:sizeof(uint64_t) atIndex:7];
[encoder setBytes:&n_groups length:sizeof( int32_t) atIndex:8];
[encoder setBytes:&eps length:sizeof( float) atIndex:9];
[encoder setThreadgroupMemoryLength:32*sizeof(float) atIndex:0];
[encoder dispatchThreadgroups:MTLSizeMake(n_groups, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
} break;
case GGML_OP_NORM:
{
GGML_ASSERT(ggml_is_contiguous_1(src0));
float eps;
memcpy(&eps, dst->op_params, sizeof(float));
const int nth = MIN(256, ne00);
id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_NORM].pipeline;
[encoder setComputePipelineState:pipeline];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_dst offset:offs_dst atIndex:1];
[encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:2];
[encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:3];
[encoder setBytes:&eps length:sizeof( float) atIndex:4];
[encoder setThreadgroupMemoryLength:GGML_PAD(nth*sizeof(float), 16) atIndex:0];
const int64_t nrows = ggml_nrows(src0);
[encoder dispatchThreadgroups:MTLSizeMake(nrows, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
} break;
case GGML_OP_ROPE:
{
GGML_ASSERT(ne10 == ne02);
const int nth = MIN(1024, ne00);
const int n_past = ((int32_t *) dst->op_params)[0];
const int n_dims = ((int32_t *) dst->op_params)[1];
const int mode = ((int32_t *) dst->op_params)[2];
// skip 3, n_ctx, used in GLM RoPE, unimplemented in metal
const int n_ctx_orig = ((int32_t *) dst->op_params)[4];
float freq_base;
float freq_scale;
float ext_factor;
float attn_factor;
float beta_fast;
float beta_slow;
memcpy(&freq_base, (int32_t *) dst->op_params + 5, sizeof(float));
memcpy(&freq_scale, (int32_t *) dst->op_params + 6, sizeof(float));
memcpy(&ext_factor, (int32_t *) dst->op_params + 7, sizeof(float));
memcpy(&attn_factor, (int32_t *) dst->op_params + 8, sizeof(float));
memcpy(&beta_fast, (int32_t *) dst->op_params + 9, sizeof(float));
memcpy(&beta_slow, (int32_t *) dst->op_params + 10, sizeof(float));
const bool is_neox = mode & 2;
id<MTLComputePipelineState> pipeline = nil;
if (!is_neox) {
switch (src0->type) {
case GGML_TYPE_F32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ROPE_NORM_F32].pipeline; break;
case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ROPE_NORM_F16].pipeline; break;
default: GGML_ASSERT(false);
};
} else {
switch (src0->type) {
case GGML_TYPE_F32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ROPE_NEOX_F32].pipeline; break;
case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ROPE_NEOX_F16].pipeline; break;
default: GGML_ASSERT(false);
};
}
[encoder setComputePipelineState:pipeline];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
if (id_src2 != nil) {
[encoder setBuffer:id_src2 offset:offs_src2 atIndex:2];
} else {
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:2];
}
[encoder setBuffer:id_dst offset:offs_dst atIndex:3];
[encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:4];
[encoder setBytes:&ne01 length:sizeof( int64_t) atIndex:5];
[encoder setBytes:&ne02 length:sizeof( int64_t) atIndex:6];
[encoder setBytes:&ne03 length:sizeof( int64_t) atIndex:7];
[encoder setBytes:&nb00 length:sizeof(uint64_t) atIndex:8];
[encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:9];
[encoder setBytes:&nb02 length:sizeof(uint64_t) atIndex:10];
[encoder setBytes:&nb03 length:sizeof(uint64_t) atIndex:11];
[encoder setBytes:&ne0 length:sizeof( int64_t) atIndex:12];
[encoder setBytes:&ne1 length:sizeof( int64_t) atIndex:13];
[encoder setBytes:&ne2 length:sizeof( int64_t) atIndex:14];
[encoder setBytes:&ne3 length:sizeof( int64_t) atIndex:15];
[encoder setBytes:&nb0 length:sizeof(uint64_t) atIndex:16];
[encoder setBytes:&nb1 length:sizeof(uint64_t) atIndex:17];
[encoder setBytes:&nb2 length:sizeof(uint64_t) atIndex:18];
[encoder setBytes:&nb3 length:sizeof(uint64_t) atIndex:19];
[encoder setBytes:&n_past length:sizeof( int) atIndex:20];
[encoder setBytes:&n_dims length:sizeof( int) atIndex:21];
[encoder setBytes:&n_ctx_orig length:sizeof( int) atIndex:22];
[encoder setBytes:&freq_base length:sizeof( float) atIndex:23];
[encoder setBytes:&freq_scale length:sizeof( float) atIndex:24];
[encoder setBytes:&ext_factor length:sizeof( float) atIndex:25];
[encoder setBytes:&attn_factor length:sizeof( float) atIndex:26];
[encoder setBytes:&beta_fast length:sizeof( float) atIndex:27];
[encoder setBytes:&beta_slow length:sizeof( float) atIndex:28];
[encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
} break;
case GGML_OP_IM2COL:
{
GGML_ASSERT(src0->type == GGML_TYPE_F16);
GGML_ASSERT(src1->type == GGML_TYPE_F32);
GGML_ASSERT( dst->type == GGML_TYPE_F16 || dst->type == GGML_TYPE_F32);
const int32_t s0 = ((const int32_t *)(dst->op_params))[0];
const int32_t s1 = ((const int32_t *)(dst->op_params))[1];
const int32_t p0 = ((const int32_t *)(dst->op_params))[2];
const int32_t p1 = ((const int32_t *)(dst->op_params))[3];
const int32_t d0 = ((const int32_t *)(dst->op_params))[4];
const int32_t d1 = ((const int32_t *)(dst->op_params))[5];
const bool is_2D = ((const int32_t *)(dst->op_params))[6] == 1;
const int32_t N = src1->ne[is_2D ? 3 : 2];
const int32_t IC = src1->ne[is_2D ? 2 : 1];
const int32_t IH = is_2D ? src1->ne[1] : 1;
const int32_t IW = src1->ne[0];
const int32_t KH = is_2D ? src0->ne[1] : 1;
const int32_t KW = src0->ne[0];
const int32_t OH = is_2D ? dst->ne[2] : 1;
const int32_t OW = dst->ne[1];
const int32_t CHW = IC * KH * KW;
const int32_t ofs0 = src1->nb[is_2D ? 3 : 2] / 4;
const int32_t ofs1 = src1->nb[is_2D ? 2 : 1] / 4;
id<MTLComputePipelineState> pipeline = nil;
switch (dst->type) {
case GGML_TYPE_F32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_IM2COL_F32].pipeline; break;
case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_IM2COL_F16].pipeline; break;
default: GGML_ASSERT(false);
};
[encoder setComputePipelineState:pipeline];
[encoder setBuffer:id_src1 offset:offs_src1 atIndex:0];
[encoder setBuffer:id_dst offset:offs_dst atIndex:1];
[encoder setBytes:&ofs0 length:sizeof( int32_t) atIndex:2];
[encoder setBytes:&ofs1 length:sizeof( int32_t) atIndex:3];
[encoder setBytes:&IW length:sizeof( int32_t) atIndex:4];
[encoder setBytes:&IH length:sizeof( int32_t) atIndex:5];
[encoder setBytes:&CHW length:sizeof( int32_t) atIndex:6];
[encoder setBytes:&s0 length:sizeof( int32_t) atIndex:7];
[encoder setBytes:&s1 length:sizeof( int32_t) atIndex:8];
[encoder setBytes:&p0 length:sizeof( int32_t) atIndex:9];
[encoder setBytes:&p1 length:sizeof( int32_t) atIndex:10];
[encoder setBytes:&d0 length:sizeof( int32_t) atIndex:11];
[encoder setBytes:&d1 length:sizeof( int32_t) atIndex:12];
[encoder dispatchThreadgroups:MTLSizeMake(IC, OH, OW) threadsPerThreadgroup:MTLSizeMake(N, KH, KW)];
} break;
case GGML_OP_UPSCALE:
{
GGML_ASSERT(src0->type == GGML_TYPE_F32);
const float sf0 = (float)ne0/src0->ne[0];
const float sf1 = (float)ne1/src0->ne[1];
const float sf2 = (float)ne2/src0->ne[2];
const float sf3 = (float)ne3/src0->ne[3];
const id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_UPSCALE_F32].pipeline;
[encoder setComputePipelineState:pipeline];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_dst offset:offs_dst atIndex:1];
[encoder setBytes:&ne00 length:sizeof(ne00) atIndex:2];
[encoder setBytes:&ne01 length:sizeof(ne01) atIndex:3];
[encoder setBytes:&ne02 length:sizeof(ne02) atIndex:4];
[encoder setBytes:&ne03 length:sizeof(ne03) atIndex:5];
[encoder setBytes:&nb00 length:sizeof(nb00) atIndex:6];
[encoder setBytes:&nb01 length:sizeof(nb01) atIndex:7];
[encoder setBytes:&nb02 length:sizeof(nb02) atIndex:8];
[encoder setBytes:&nb03 length:sizeof(nb03) atIndex:9];
[encoder setBytes:&ne0 length:sizeof(ne0) atIndex:10];
[encoder setBytes:&ne1 length:sizeof(ne1) atIndex:11];
[encoder setBytes:&ne2 length:sizeof(ne2) atIndex:12];
[encoder setBytes:&ne3 length:sizeof(ne3) atIndex:13];
[encoder setBytes:&nb0 length:sizeof(nb0) atIndex:14];
[encoder setBytes:&nb1 length:sizeof(nb1) atIndex:15];
[encoder setBytes:&nb2 length:sizeof(nb2) atIndex:16];
[encoder setBytes:&nb3 length:sizeof(nb3) atIndex:17];
[encoder setBytes:&sf0 length:sizeof(sf0) atIndex:18];
[encoder setBytes:&sf1 length:sizeof(sf1) atIndex:19];
[encoder setBytes:&sf2 length:sizeof(sf2) atIndex:20];
[encoder setBytes:&sf3 length:sizeof(sf3) atIndex:21];
const int nth = MIN((int) pipeline.maxTotalThreadsPerThreadgroup, ne0);
[encoder dispatchThreadgroups:MTLSizeMake(ne1, ne2, ne3) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
} break;
case GGML_OP_PAD:
{
GGML_ASSERT(src0->type == GGML_TYPE_F32);
id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_PAD_F32].pipeline;
[encoder setComputePipelineState:pipeline];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_dst offset:offs_dst atIndex:1];
[encoder setBytes:&ne00 length:sizeof(ne00) atIndex:2];
[encoder setBytes:&ne01 length:sizeof(ne01) atIndex:3];
[encoder setBytes:&ne02 length:sizeof(ne02) atIndex:4];
[encoder setBytes:&ne03 length:sizeof(ne03) atIndex:5];
[encoder setBytes:&nb00 length:sizeof(nb00) atIndex:6];
[encoder setBytes:&nb01 length:sizeof(nb01) atIndex:7];
[encoder setBytes:&nb02 length:sizeof(nb02) atIndex:8];
[encoder setBytes:&nb03 length:sizeof(nb03) atIndex:9];
[encoder setBytes:&ne0 length:sizeof(ne0) atIndex:10];
[encoder setBytes:&ne1 length:sizeof(ne1) atIndex:11];
[encoder setBytes:&ne2 length:sizeof(ne2) atIndex:12];
[encoder setBytes:&ne3 length:sizeof(ne3) atIndex:13];
[encoder setBytes:&nb0 length:sizeof(nb0) atIndex:14];
[encoder setBytes:&nb1 length:sizeof(nb1) atIndex:15];
[encoder setBytes:&nb2 length:sizeof(nb2) atIndex:16];
[encoder setBytes:&nb3 length:sizeof(nb3) atIndex:17];
const int nth = MIN(1024, ne0);
[encoder dispatchThreadgroups:MTLSizeMake(ne1, ne2, ne3) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
} break;
case GGML_OP_ARANGE:
{
GGML_ASSERT(dst->type == GGML_TYPE_F32);
float start;
float step;
memcpy(&start, ((int32_t *) dst->op_params) + 0, sizeof(float));
memcpy(&step, ((int32_t *) dst->op_params) + 2, sizeof(float));
id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ARANGE_F32].pipeline;
[encoder setComputePipelineState:pipeline];
[encoder setBuffer:id_dst offset:offs_dst atIndex:0];
[encoder setBytes:&ne0 length:sizeof(ne0) atIndex:1];
[encoder setBytes:&start length:sizeof(start) atIndex:2];
[encoder setBytes:&step length:sizeof(step) atIndex:3];
const int nth = MIN(1024, ne0);
[encoder dispatchThreadgroups:MTLSizeMake(1, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
} break;
case GGML_OP_TIMESTEP_EMBEDDING:
{
GGML_ASSERT(src0->type == GGML_TYPE_F32);
const int dim = dst->op_params[0];
const int max_period = dst->op_params[1];
const int half = dim / 2;
id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_TIMESTEP_EMBEDDING_F32].pipeline;
[encoder setComputePipelineState:pipeline];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_dst offset:offs_dst atIndex:1];
[encoder setBytes:&nb1 length:sizeof(nb1) atIndex:2];
[encoder setBytes:&dim length:sizeof(dim) atIndex:3];
[encoder setBytes:&max_period length:sizeof(max_period) atIndex:4];
const int nth = MIN(1024, half);
[encoder dispatchThreadgroups:MTLSizeMake(ne00, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
} break;
case GGML_OP_ARGSORT:
{
GGML_ASSERT(src0->type == GGML_TYPE_F32);
GGML_ASSERT( dst->type == GGML_TYPE_I32);
const int nrows = ggml_nrows(src0);
enum ggml_sort_order order = (enum ggml_sort_order) dst->op_params[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
// bitonic sort requires the number of elements to be power of 2
int64_t ne00_padded = 1;
while (ne00_padded < ne00) {
ne00_padded *= 2;
}
// Metal kernels require the buffer size to be multiple of 16 bytes
// https://developer.apple.com/documentation/metal/mtlcomputecommandencoder/1443142-setthreadgroupmemorylength
const int mem_size = GGML_PAD(ne00_padded*sizeof(int32_t), 16);
id<MTLComputePipelineState> pipeline = nil;
switch (order) {
case GGML_SORT_ORDER_ASC: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_ASC].pipeline; break;
case GGML_SORT_ORDER_DESC: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_DESC].pipeline; break;
default: GGML_ASSERT(false);
};
[encoder setComputePipelineState:pipeline];
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
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_dst offset:offs_dst atIndex:1];
[encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:2];
[encoder setBytes:&ne00_padded length:sizeof( int64_t) atIndex:3];
[encoder setThreadgroupMemoryLength:mem_size atIndex: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
[encoder dispatchThreadgroups:MTLSizeMake(1, nrows, 1) threadsPerThreadgroup:MTLSizeMake(ne00_padded, 1, 1)];
} break;
case GGML_OP_LEAKY_RELU:
{
GGML_ASSERT(src0->type == GGML_TYPE_F32);
float slope;
memcpy(&slope, dst->op_params, sizeof(float));
id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_LEAKY_RELU_F32].pipeline;
[encoder setComputePipelineState:pipeline];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_dst offset:offs_dst atIndex:1];
[encoder setBytes:&slope length:sizeof(slope) atIndex:2];
const int64_t n = ggml_nelements(dst);
[encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
} break;
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
case GGML_OP_FLASH_ATTN_EXT:
{
GGML_ASSERT(ne00 % 4 == 0);
GGML_ASSERT(ne11 % 32 == 0);
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
GGML_ASSERT(src0->type == GGML_TYPE_F32);
GGML_ASSERT(ggml_are_same_shape (src1, src2));
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
struct ggml_tensor * src3 = gf->nodes[i]->src[3];
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
size_t offs_src3 = 0;
id<MTLBuffer> id_src3 = src3 ? ggml_metal_get_buffer(src3, &offs_src3) : nil;
GGML_ASSERT(!src3 || src3->type == GGML_TYPE_F16);
GGML_ASSERT(!src3 || src3->ne[1] >= GGML_PAD(src0->ne[1], 8) &&
"the Flash-Attention Metal kernel requires the mask to be padded to 8 and at least n_queries big");
const int64_t ne30 = src3 ? src3->ne[0] : 0; GGML_UNUSED(ne30);
//const int64_t ne31 = src3 ? src3->ne[1] : 0;
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 int64_t ne32 = src3 ? src3->ne[2] : 0; GGML_UNUSED(ne32);
const int64_t ne33 = src3 ? src3->ne[3] : 0; GGML_UNUSED(ne33);
const uint64_t nb30 = src3 ? src3->nb[0] : 0; GGML_UNUSED(nb30);
const uint64_t nb31 = src3 ? src3->nb[1] : 0;
const uint64_t nb32 = src3 ? src3->nb[2] : 0; GGML_UNUSED(nb32);
const uint64_t nb33 = src3 ? src3->nb[3] : 0; GGML_UNUSED(nb33);
const enum ggml_type src2t = src2 ? src2->type : GGML_TYPE_COUNT; GGML_UNUSED(src2t);
float scale;
float max_bias;
memcpy(&scale, ((int32_t *) dst->op_params) + 0, sizeof(scale));
memcpy(&max_bias, ((int32_t *) dst->op_params) + 1, sizeof(max_bias));
const uint32_t n_head = src0->ne[2];
const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head));
const float m0 = powf(2.0f, -(max_bias ) / n_head_log2);
const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
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
id<MTLComputePipelineState> pipeline = nil;
bool use_vec_kernel = false;
if (ne01 >= 4 || (ne00%128 != 0)) {
switch (ne00) {
case 64: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H64 ].pipeline; break;
case 80: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H80 ].pipeline; break;
case 96: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H96 ].pipeline; break;
case 112: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H112].pipeline; break;
case 128: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H128].pipeline; break;
//case 256: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H256].pipeline; break;
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
default:
{
GGML_METAL_LOG_ERROR("unsupported size: %lld\n", ne00);
GGML_METAL_LOG_ERROR("add template specialization for this size\n");
GGML_ASSERT(false && "add template specialization for this size");
}
}
} else {
use_vec_kernel = true;
switch (ne00) {
case 128: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H128].pipeline; break;
//case 256: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H256].pipeline; break;
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
default:
{
GGML_METAL_LOG_ERROR("unsupported size: %lld\n", ne00);
GGML_METAL_LOG_ERROR("add template specialization for this size\n");
GGML_ASSERT(false && "add template specialization for this size");
}
}
}
[encoder setComputePipelineState:pipeline];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
[encoder setBuffer:id_src2 offset:offs_src2 atIndex:2];
if (id_src3) {
[encoder setBuffer:id_src3 offset:offs_src3 atIndex:3];
} else {
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:3];
}
[encoder setBuffer:id_dst offset:offs_dst atIndex:4];
[encoder setBytes:&ne01 length:sizeof( int64_t) atIndex:5];
[encoder setBytes:&ne02 length:sizeof( int64_t) atIndex:6];
[encoder setBytes:&ne03 length:sizeof( int64_t) atIndex:7];
[encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:8];
[encoder setBytes:&nb02 length:sizeof(uint64_t) atIndex:9];
[encoder setBytes:&nb03 length:sizeof(uint64_t) atIndex:10];
[encoder setBytes:&ne11 length:sizeof( int64_t) atIndex:11];
[encoder setBytes:&ne12 length:sizeof( int64_t) atIndex:12];
[encoder setBytes:&ne13 length:sizeof( int64_t) atIndex:13];
[encoder setBytes:&nb11 length:sizeof(uint64_t) atIndex:14];
[encoder setBytes:&nb12 length:sizeof(uint64_t) atIndex:15];
[encoder setBytes:&nb13 length:sizeof(uint64_t) atIndex:16];
[encoder setBytes:&nb21 length:sizeof(uint64_t) atIndex:17];
[encoder setBytes:&nb22 length:sizeof(uint64_t) atIndex:18];
[encoder setBytes:&nb23 length:sizeof(uint64_t) atIndex:19];
[encoder setBytes:&nb31 length:sizeof(uint64_t) atIndex:20];
[encoder setBytes:&ne1 length:sizeof( int64_t) atIndex:21];
[encoder setBytes:&ne2 length:sizeof( int64_t) atIndex:22];
[encoder setBytes:&scale length:sizeof( float) atIndex:23];
[encoder setBytes:&max_bias length:sizeof( float) atIndex:24];
[encoder setBytes:&m0 length:sizeof(m0) atIndex:25];
[encoder setBytes:&m1 length:sizeof(m1) atIndex:26];
[encoder setBytes:&n_head_log2 length:sizeof(n_head_log2) atIndex:27];
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
if (!use_vec_kernel) {
// half8x8 kernel
const int64_t nqptg = 8; // queries per threadgroup !! sync with kernel template arguments !!
const int64_t ncpsg = 32; // cache values per simdgroup !! sync with kernel template arguments !!
GGML_ASSERT(nqptg <= 32);
GGML_ASSERT(nqptg % 8 == 0);
GGML_ASSERT(ncpsg % 32 == 0);
int64_t nsgmax = 2;
while (true) {
const size_t smem = nqptg*(ne00 + 2*nsgmax*(ncpsg + nqptg))*(sizeof(float)/2);
if (smem > ctx->device.maxThreadgroupMemoryLength) {
break;
}
nsgmax *= 2;
}
nsgmax /= 2;
// simdgroups per threadgroup (a.k.a. warps)
const int64_t nsg = ne01 <= nqptg ? MAX(4, MIN(nsgmax, MIN(ne11/ncpsg, (int64_t) pipeline.maxTotalThreadsPerThreadgroup/32))) : 4;
const size_t smem = nqptg*(ne00 + 2*nsg*(ncpsg + nqptg))*(sizeof(float)/2);
//printf("smem: %zu, max: %zu\n", smem, ctx->device.maxThreadgroupMemoryLength);
GGML_ASSERT(smem <= ctx->device.maxThreadgroupMemoryLength);
[encoder setThreadgroupMemoryLength:GGML_PAD(smem, 16) atIndex:0];
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + nqptg - 1)/nqptg, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(32, nsg, 1)];
} else {
// half1x4 kernel
const int64_t nqptg = 1; // queries per threadgroup !! sync with kernel template arguments !!
const int64_t ncpsg = 32; // cache values per simdgroup !! sync with kernel template arguments !!
GGML_ASSERT(nqptg <= 32);
GGML_ASSERT(nqptg % 1 == 0);
GGML_ASSERT(ncpsg % 32 == 0);
// simdgroups per threadgroup (a.k.a. warps)
const int64_t nsgt = MAX(2, MIN(ne11/ncpsg, (int64_t) pipeline.maxTotalThreadsPerThreadgroup/32));
int64_t nsg = 1;
while (nsg <= nsgt) {
nsg *= 2;
}
nsg /= 2;
const size_t smem = (nqptg*(ne00 + 2*nsg*(ncpsg + nqptg)) + nsg*ne00)*(sizeof(float)/2);
//printf("smem: %zu, max: %zu\n", smem, ctx->device.maxThreadgroupMemoryLength);
GGML_ASSERT(smem <= ctx->device.maxThreadgroupMemoryLength);
[encoder setThreadgroupMemoryLength:GGML_PAD(smem, 16) atIndex:0];
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + nqptg - 1)/nqptg, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(32, nsg, 1)];
}
} break;
case GGML_OP_DUP:
case GGML_OP_CPY:
case GGML_OP_CONT:
{
GGML_ASSERT(ne00 % ggml_blck_size(src0->type) == 0);
int nth = MIN(1024, ne00/ggml_blck_size(src0->type));
id<MTLComputePipelineState> pipeline = nil;
switch (src0t) {
case GGML_TYPE_F32:
{
GGML_ASSERT(ne0 % ggml_blck_size(dst->type) == 0);
switch (dstt) {
case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_F16].pipeline; break;
case GGML_TYPE_F32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_F32].pipeline; break;
case GGML_TYPE_Q8_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_Q8_0].pipeline; break;
case GGML_TYPE_Q4_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_Q4_0].pipeline; break;
case GGML_TYPE_Q4_1: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_Q4_1].pipeline; break;
case GGML_TYPE_Q5_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_Q5_0].pipeline; break;
case GGML_TYPE_Q5_1: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_Q5_1].pipeline; break;
case GGML_TYPE_IQ4_NL: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_IQ4_NL].pipeline; break;
default: GGML_ASSERT(false && "not implemented");
};
} break;
case GGML_TYPE_F16:
{
switch (dstt) {
case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F16_F16].pipeline; break;
case GGML_TYPE_F32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F16_F32].pipeline; break;
default: GGML_ASSERT(false && "not implemented");
};
} break;
default: GGML_ASSERT(false && "not implemented");
}
[encoder setComputePipelineState:pipeline];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_dst offset:offs_dst atIndex:1];
[encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:2];
[encoder setBytes:&ne01 length:sizeof( int64_t) atIndex:3];
[encoder setBytes:&ne02 length:sizeof( int64_t) atIndex:4];
[encoder setBytes:&ne03 length:sizeof( int64_t) atIndex:5];
[encoder setBytes:&nb00 length:sizeof(uint64_t) atIndex:6];
[encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:7];
[encoder setBytes:&nb02 length:sizeof(uint64_t) atIndex:8];
[encoder setBytes:&nb03 length:sizeof(uint64_t) atIndex:9];
[encoder setBytes:&ne0 length:sizeof( int64_t) atIndex:10];
[encoder setBytes:&ne1 length:sizeof( int64_t) atIndex:11];
[encoder setBytes:&ne2 length:sizeof( int64_t) atIndex:12];
[encoder setBytes:&ne3 length:sizeof( int64_t) atIndex:13];
[encoder setBytes:&nb0 length:sizeof(uint64_t) atIndex:14];
[encoder setBytes:&nb1 length:sizeof(uint64_t) atIndex:15];
[encoder setBytes:&nb2 length:sizeof(uint64_t) atIndex:16];
[encoder setBytes:&nb3 length:sizeof(uint64_t) atIndex:17];
[encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
} break;
default:
{
GGML_METAL_LOG_ERROR("%s: error: node %3d, op = %8s not implemented\n", __func__, i, ggml_op_name(dst->op));
GGML_ASSERT(false);
}
}
if (should_capture) {
[encoder popDebugGroup];
}
}
[encoder endEncoding];
[command_buffer commit];
});
// Wait for completion and check status of each command buffer
// needed to detect if the device ran out-of-memory for example (#1881)
for (int i = 0; i < n_cb; ++i) {
id<MTLCommandBuffer> command_buffer = command_buffers[i];
[command_buffer waitUntilCompleted];
MTLCommandBufferStatus status = [command_buffer status];
if (status != MTLCommandBufferStatusCompleted) {
GGML_METAL_LOG_INFO("%s: command buffer %d failed with status %lu\n", __func__, i, status);
2024-04-30 11:34:50 +02:00
if (status == MTLCommandBufferStatusError) {
NSString * error_code = [command_buffer error].localizedDescription;
GGML_METAL_LOG_INFO("error: %s\n", [error_code UTF8String]);
2024-04-30 11:34:50 +02:00
}
return GGML_STATUS_FAILED;
}
}
if (should_capture) {
[[MTLCaptureManager sharedCaptureManager] stopCapture];
}
}
return GGML_STATUS_SUCCESS;
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
}
////////////////////////////////////////////////////////////////////////////////
// backend interface
// default buffer
static id<MTLDevice> g_backend_device = nil;
static int g_backend_device_ref_count = 0;
static id<MTLDevice> ggml_backend_metal_get_device(void) {
if (g_backend_device == nil) {
g_backend_device = MTLCreateSystemDefaultDevice();
}
g_backend_device_ref_count++;
return g_backend_device;
}
static void ggml_backend_metal_free_device(void) {
assert(g_backend_device_ref_count > 0);
g_backend_device_ref_count--;
if (g_backend_device_ref_count == 0) {
[g_backend_device release];
g_backend_device = nil;
}
}
GGML_CALL static const char * ggml_backend_metal_buffer_get_name(ggml_backend_buffer_t buffer) {
llama : ggml-backend integration (#4766) * llama : ggml-backend integration * ggml-backend : add names to buffers * fix unmap after loading * batched-bench : add tensor_split param * llama : check for null tensor_split * ggml-backend : increase GGML_MAX_BACKENDS * improve graph splitting, partial fix for --no-kv-offload * cuda : add ggml-backend split buffer support * cuda : do not create buffer types for devices that don't exist (fixes usage without CUDA devices available) * ggml : fix null backend dereference (#4807) * ggml : fix null backend dereference * ggml : also check ggml_backend_is_cpu * test-backend-ops : check buffer allocation failures * llama : add cparam (split_mode) and command line argument (--split-mode, -sm) to configure the split mode (none, layer or row) * ggml : fix mul_mat_id work size * llama : rewrite session kv load/set without graphs * minor * llama : only initialize used backends, free backends on context free * llama : abort ctx if cuda backend init fails * llama : rewrite lora with ggml-backend and compute on CPU ggml-ci * llama : only map to a backend buffer the region of the file mapping containing the tensors used in the buffer * opencl : add ggml-backend buffer type * cuda : only use batched_cublas with batched mat muls (fixes fp16 tg perf) * llama : on Metal, by default offload the full model ggml-ci * metal : page align the data ptr (#4854) * Apply suggestions from code review Co-authored-by: Johannes Gäßler <johannesg@5d6.de> * cuda : fix split buffer free * address review comments * llama-bench : add split-mode parameter * fix whitespace * opencl : fix double initialization * server : add --split-mode parameter * use async copy and compute to improve multi-gpu performance ggml-ci * use async memcpys to copy the graph outputs to the CPU * fix opencl * use a host buffer for the cpu compute buffer for faster copies to the gpu --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
2024-01-12 20:07:38 +01:00
return "Metal";
llama : ggml-backend integration (#4766) * llama : ggml-backend integration * ggml-backend : add names to buffers * fix unmap after loading * batched-bench : add tensor_split param * llama : check for null tensor_split * ggml-backend : increase GGML_MAX_BACKENDS * improve graph splitting, partial fix for --no-kv-offload * cuda : add ggml-backend split buffer support * cuda : do not create buffer types for devices that don't exist (fixes usage without CUDA devices available) * ggml : fix null backend dereference (#4807) * ggml : fix null backend dereference * ggml : also check ggml_backend_is_cpu * test-backend-ops : check buffer allocation failures * llama : add cparam (split_mode) and command line argument (--split-mode, -sm) to configure the split mode (none, layer or row) * ggml : fix mul_mat_id work size * llama : rewrite session kv load/set without graphs * minor * llama : only initialize used backends, free backends on context free * llama : abort ctx if cuda backend init fails * llama : rewrite lora with ggml-backend and compute on CPU ggml-ci * llama : only map to a backend buffer the region of the file mapping containing the tensors used in the buffer * opencl : add ggml-backend buffer type * cuda : only use batched_cublas with batched mat muls (fixes fp16 tg perf) * llama : on Metal, by default offload the full model ggml-ci * metal : page align the data ptr (#4854) * Apply suggestions from code review Co-authored-by: Johannes Gäßler <johannesg@5d6.de> * cuda : fix split buffer free * address review comments * llama-bench : add split-mode parameter * fix whitespace * opencl : fix double initialization * server : add --split-mode parameter * use async copy and compute to improve multi-gpu performance ggml-ci * use async memcpys to copy the graph outputs to the CPU * fix opencl * use a host buffer for the cpu compute buffer for faster copies to the gpu --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
2024-01-12 20:07:38 +01:00
UNUSED(buffer);
}
GGML_CALL static void ggml_backend_metal_buffer_free_buffer(ggml_backend_buffer_t buffer) {
struct ggml_backend_metal_buffer_context * ctx = (struct ggml_backend_metal_buffer_context *)buffer->context;
for (int i = 0; i < ctx->n_buffers; i++) {
[ctx->buffers[i].metal release];
}
ggml_backend_metal_free_device();
if (ctx->owned) {
#if TARGET_OS_OSX
vm_deallocate((vm_map_t)mach_task_self(), (vm_address_t)ctx->all_data, ctx->all_size);
#else
free(ctx->all_data);
#endif
}
free(ctx);
}
GGML_CALL static void * ggml_backend_metal_buffer_get_base(ggml_backend_buffer_t buffer) {
llama : ggml-backend integration (#4766) * llama : ggml-backend integration * ggml-backend : add names to buffers * fix unmap after loading * batched-bench : add tensor_split param * llama : check for null tensor_split * ggml-backend : increase GGML_MAX_BACKENDS * improve graph splitting, partial fix for --no-kv-offload * cuda : add ggml-backend split buffer support * cuda : do not create buffer types for devices that don't exist (fixes usage without CUDA devices available) * ggml : fix null backend dereference (#4807) * ggml : fix null backend dereference * ggml : also check ggml_backend_is_cpu * test-backend-ops : check buffer allocation failures * llama : add cparam (split_mode) and command line argument (--split-mode, -sm) to configure the split mode (none, layer or row) * ggml : fix mul_mat_id work size * llama : rewrite session kv load/set without graphs * minor * llama : only initialize used backends, free backends on context free * llama : abort ctx if cuda backend init fails * llama : rewrite lora with ggml-backend and compute on CPU ggml-ci * llama : only map to a backend buffer the region of the file mapping containing the tensors used in the buffer * opencl : add ggml-backend buffer type * cuda : only use batched_cublas with batched mat muls (fixes fp16 tg perf) * llama : on Metal, by default offload the full model ggml-ci * metal : page align the data ptr (#4854) * Apply suggestions from code review Co-authored-by: Johannes Gäßler <johannesg@5d6.de> * cuda : fix split buffer free * address review comments * llama-bench : add split-mode parameter * fix whitespace * opencl : fix double initialization * server : add --split-mode parameter * use async copy and compute to improve multi-gpu performance ggml-ci * use async memcpys to copy the graph outputs to the CPU * fix opencl * use a host buffer for the cpu compute buffer for faster copies to the gpu --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
2024-01-12 20:07:38 +01:00
struct ggml_backend_metal_buffer_context * ctx = (struct ggml_backend_metal_buffer_context *)buffer->context;
return ctx->all_data;
}
GGML_CALL static void ggml_backend_metal_buffer_set_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
memcpy((char *)tensor->data + offset, data, size);
UNUSED(buffer);
}
GGML_CALL static void ggml_backend_metal_buffer_get_tensor(ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
memcpy(data, (const char *)tensor->data + offset, size);
UNUSED(buffer);
}
GGML_CALL static bool ggml_backend_metal_buffer_cpy_tensor(ggml_backend_buffer_t buffer, const struct ggml_tensor * src, struct ggml_tensor * dst) {
llama : ggml-backend integration (#4766) * llama : ggml-backend integration * ggml-backend : add names to buffers * fix unmap after loading * batched-bench : add tensor_split param * llama : check for null tensor_split * ggml-backend : increase GGML_MAX_BACKENDS * improve graph splitting, partial fix for --no-kv-offload * cuda : add ggml-backend split buffer support * cuda : do not create buffer types for devices that don't exist (fixes usage without CUDA devices available) * ggml : fix null backend dereference (#4807) * ggml : fix null backend dereference * ggml : also check ggml_backend_is_cpu * test-backend-ops : check buffer allocation failures * llama : add cparam (split_mode) and command line argument (--split-mode, -sm) to configure the split mode (none, layer or row) * ggml : fix mul_mat_id work size * llama : rewrite session kv load/set without graphs * minor * llama : only initialize used backends, free backends on context free * llama : abort ctx if cuda backend init fails * llama : rewrite lora with ggml-backend and compute on CPU ggml-ci * llama : only map to a backend buffer the region of the file mapping containing the tensors used in the buffer * opencl : add ggml-backend buffer type * cuda : only use batched_cublas with batched mat muls (fixes fp16 tg perf) * llama : on Metal, by default offload the full model ggml-ci * metal : page align the data ptr (#4854) * Apply suggestions from code review Co-authored-by: Johannes Gäßler <johannesg@5d6.de> * cuda : fix split buffer free * address review comments * llama-bench : add split-mode parameter * fix whitespace * opencl : fix double initialization * server : add --split-mode parameter * use async copy and compute to improve multi-gpu performance ggml-ci * use async memcpys to copy the graph outputs to the CPU * fix opencl * use a host buffer for the cpu compute buffer for faster copies to the gpu --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
2024-01-12 20:07:38 +01:00
if (ggml_backend_buffer_is_host(src->buffer)) {
memcpy(dst->data, src->data, ggml_nbytes(src));
return true;
}
return false;
UNUSED(buffer);
}
GGML_CALL static void ggml_backend_metal_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
struct ggml_backend_metal_buffer_context * ctx = (struct ggml_backend_metal_buffer_context *)buffer->context;
memset(ctx->all_data, value, ctx->all_size);
}
static struct ggml_backend_buffer_i ggml_backend_metal_buffer_i = {
llama : ggml-backend integration (#4766) * llama : ggml-backend integration * ggml-backend : add names to buffers * fix unmap after loading * batched-bench : add tensor_split param * llama : check for null tensor_split * ggml-backend : increase GGML_MAX_BACKENDS * improve graph splitting, partial fix for --no-kv-offload * cuda : add ggml-backend split buffer support * cuda : do not create buffer types for devices that don't exist (fixes usage without CUDA devices available) * ggml : fix null backend dereference (#4807) * ggml : fix null backend dereference * ggml : also check ggml_backend_is_cpu * test-backend-ops : check buffer allocation failures * llama : add cparam (split_mode) and command line argument (--split-mode, -sm) to configure the split mode (none, layer or row) * ggml : fix mul_mat_id work size * llama : rewrite session kv load/set without graphs * minor * llama : only initialize used backends, free backends on context free * llama : abort ctx if cuda backend init fails * llama : rewrite lora with ggml-backend and compute on CPU ggml-ci * llama : only map to a backend buffer the region of the file mapping containing the tensors used in the buffer * opencl : add ggml-backend buffer type * cuda : only use batched_cublas with batched mat muls (fixes fp16 tg perf) * llama : on Metal, by default offload the full model ggml-ci * metal : page align the data ptr (#4854) * Apply suggestions from code review Co-authored-by: Johannes Gäßler <johannesg@5d6.de> * cuda : fix split buffer free * address review comments * llama-bench : add split-mode parameter * fix whitespace * opencl : fix double initialization * server : add --split-mode parameter * use async copy and compute to improve multi-gpu performance ggml-ci * use async memcpys to copy the graph outputs to the CPU * fix opencl * use a host buffer for the cpu compute buffer for faster copies to the gpu --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
2024-01-12 20:07:38 +01:00
/* .get_name = */ ggml_backend_metal_buffer_get_name,
/* .free_buffer = */ ggml_backend_metal_buffer_free_buffer,
/* .get_base = */ ggml_backend_metal_buffer_get_base,
/* .init_tensor = */ NULL,
/* .set_tensor = */ ggml_backend_metal_buffer_set_tensor,
/* .get_tensor = */ ggml_backend_metal_buffer_get_tensor,
llama : ggml-backend integration (#4766) * llama : ggml-backend integration * ggml-backend : add names to buffers * fix unmap after loading * batched-bench : add tensor_split param * llama : check for null tensor_split * ggml-backend : increase GGML_MAX_BACKENDS * improve graph splitting, partial fix for --no-kv-offload * cuda : add ggml-backend split buffer support * cuda : do not create buffer types for devices that don't exist (fixes usage without CUDA devices available) * ggml : fix null backend dereference (#4807) * ggml : fix null backend dereference * ggml : also check ggml_backend_is_cpu * test-backend-ops : check buffer allocation failures * llama : add cparam (split_mode) and command line argument (--split-mode, -sm) to configure the split mode (none, layer or row) * ggml : fix mul_mat_id work size * llama : rewrite session kv load/set without graphs * minor * llama : only initialize used backends, free backends on context free * llama : abort ctx if cuda backend init fails * llama : rewrite lora with ggml-backend and compute on CPU ggml-ci * llama : only map to a backend buffer the region of the file mapping containing the tensors used in the buffer * opencl : add ggml-backend buffer type * cuda : only use batched_cublas with batched mat muls (fixes fp16 tg perf) * llama : on Metal, by default offload the full model ggml-ci * metal : page align the data ptr (#4854) * Apply suggestions from code review Co-authored-by: Johannes Gäßler <johannesg@5d6.de> * cuda : fix split buffer free * address review comments * llama-bench : add split-mode parameter * fix whitespace * opencl : fix double initialization * server : add --split-mode parameter * use async copy and compute to improve multi-gpu performance ggml-ci * use async memcpys to copy the graph outputs to the CPU * fix opencl * use a host buffer for the cpu compute buffer for faster copies to the gpu --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
2024-01-12 20:07:38 +01:00
/* .cpy_tensor = */ ggml_backend_metal_buffer_cpy_tensor,
/* .clear = */ ggml_backend_metal_buffer_clear,
llama : ggml-backend integration (#4766) * llama : ggml-backend integration * ggml-backend : add names to buffers * fix unmap after loading * batched-bench : add tensor_split param * llama : check for null tensor_split * ggml-backend : increase GGML_MAX_BACKENDS * improve graph splitting, partial fix for --no-kv-offload * cuda : add ggml-backend split buffer support * cuda : do not create buffer types for devices that don't exist (fixes usage without CUDA devices available) * ggml : fix null backend dereference (#4807) * ggml : fix null backend dereference * ggml : also check ggml_backend_is_cpu * test-backend-ops : check buffer allocation failures * llama : add cparam (split_mode) and command line argument (--split-mode, -sm) to configure the split mode (none, layer or row) * ggml : fix mul_mat_id work size * llama : rewrite session kv load/set without graphs * minor * llama : only initialize used backends, free backends on context free * llama : abort ctx if cuda backend init fails * llama : rewrite lora with ggml-backend and compute on CPU ggml-ci * llama : only map to a backend buffer the region of the file mapping containing the tensors used in the buffer * opencl : add ggml-backend buffer type * cuda : only use batched_cublas with batched mat muls (fixes fp16 tg perf) * llama : on Metal, by default offload the full model ggml-ci * metal : page align the data ptr (#4854) * Apply suggestions from code review Co-authored-by: Johannes Gäßler <johannesg@5d6.de> * cuda : fix split buffer free * address review comments * llama-bench : add split-mode parameter * fix whitespace * opencl : fix double initialization * server : add --split-mode parameter * use async copy and compute to improve multi-gpu performance ggml-ci * use async memcpys to copy the graph outputs to the CPU * fix opencl * use a host buffer for the cpu compute buffer for faster copies to the gpu --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
2024-01-12 20:07:38 +01:00
/* .reset = */ NULL,
};
// default buffer type
GGML_CALL static const char * ggml_backend_metal_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
llama : ggml-backend integration (#4766) * llama : ggml-backend integration * ggml-backend : add names to buffers * fix unmap after loading * batched-bench : add tensor_split param * llama : check for null tensor_split * ggml-backend : increase GGML_MAX_BACKENDS * improve graph splitting, partial fix for --no-kv-offload * cuda : add ggml-backend split buffer support * cuda : do not create buffer types for devices that don't exist (fixes usage without CUDA devices available) * ggml : fix null backend dereference (#4807) * ggml : fix null backend dereference * ggml : also check ggml_backend_is_cpu * test-backend-ops : check buffer allocation failures * llama : add cparam (split_mode) and command line argument (--split-mode, -sm) to configure the split mode (none, layer or row) * ggml : fix mul_mat_id work size * llama : rewrite session kv load/set without graphs * minor * llama : only initialize used backends, free backends on context free * llama : abort ctx if cuda backend init fails * llama : rewrite lora with ggml-backend and compute on CPU ggml-ci * llama : only map to a backend buffer the region of the file mapping containing the tensors used in the buffer * opencl : add ggml-backend buffer type * cuda : only use batched_cublas with batched mat muls (fixes fp16 tg perf) * llama : on Metal, by default offload the full model ggml-ci * metal : page align the data ptr (#4854) * Apply suggestions from code review Co-authored-by: Johannes Gäßler <johannesg@5d6.de> * cuda : fix split buffer free * address review comments * llama-bench : add split-mode parameter * fix whitespace * opencl : fix double initialization * server : add --split-mode parameter * use async copy and compute to improve multi-gpu performance ggml-ci * use async memcpys to copy the graph outputs to the CPU * fix opencl * use a host buffer for the cpu compute buffer for faster copies to the gpu --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
2024-01-12 20:07:38 +01:00
return "Metal";
UNUSED(buft);
}
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
static void ggml_backend_metal_log_allocated_size(id<MTLDevice> device, size_t size_aligned) {
#ifndef GGML_METAL_NDEBUG
#if TARGET_OS_OSX || (TARGET_OS_IOS && __clang_major__ >= 15)
if (@available(macOS 10.12, iOS 16.0, *)) {
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
GGML_METAL_LOG_INFO("%s: allocated buffer, size = %8.2f MiB, (%8.2f / %8.2f)",
__func__,
size_aligned / 1024.0 / 1024.0,
device.currentAllocatedSize / 1024.0 / 1024.0,
device.recommendedMaxWorkingSetSize / 1024.0 / 1024.0);
if (device.currentAllocatedSize > device.recommendedMaxWorkingSetSize) {
GGML_METAL_LOG_WARN("%s: warning: current allocated size is greater than the recommended max working set size\n", __func__);
} else {
GGML_METAL_LOG_INFO("\n");
}
} else {
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
GGML_METAL_LOG_INFO("%s: allocated buffer, size = %8.2f MiB, (%8.2f)\n",
__func__,
size_aligned / 1024.0 / 1024.0,
device.currentAllocatedSize / 1024.0 / 1024.0);
}
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
#endif
#endif
UNUSED(device);
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
UNUSED(size_aligned);
}
GGML_CALL static ggml_backend_buffer_t ggml_backend_metal_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
struct ggml_backend_metal_buffer_context * ctx = malloc(sizeof(struct ggml_backend_metal_buffer_context));
const size_t size_page = sysconf(_SC_PAGESIZE);
size_t size_aligned = size;
if ((size_aligned % size_page) != 0) {
size_aligned += (size_page - (size_aligned % size_page));
}
id<MTLDevice> device = ggml_backend_metal_get_device();
ctx->all_data = ggml_metal_host_malloc(size_aligned);
ctx->all_size = size_aligned;
ctx->owned = true;
ctx->n_buffers = 1;
if (ctx->all_data != NULL) {
ctx->buffers[0].data = ctx->all_data;
ctx->buffers[0].size = size;
ctx->buffers[0].metal = [device newBufferWithBytesNoCopy:ctx->all_data
length:size_aligned
options:MTLResourceStorageModeShared
deallocator:nil];
}
if (ctx->all_data == NULL || ctx->buffers[0].metal == nil) {
GGML_METAL_LOG_ERROR("%s: error: failed to allocate buffer, size = %8.2f MiB\n", __func__, size_aligned / 1024.0 / 1024.0);
free(ctx);
ggml_backend_metal_free_device();
return NULL;
}
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
//ggml_backend_metal_log_allocated_size(device, size_aligned);
return ggml_backend_buffer_init(buft, ggml_backend_metal_buffer_i, ctx, size);
}
GGML_CALL static size_t ggml_backend_metal_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
return 32;
UNUSED(buft);
}
GGML_CALL static size_t ggml_backend_metal_buffer_type_get_max_size(ggml_backend_buffer_type_t buft) {
id<MTLDevice> device = ggml_backend_metal_get_device();
size_t max_size = device.maxBufferLength;
ggml_backend_metal_free_device();
return max_size;
UNUSED(buft);
}
GGML_CALL static bool ggml_backend_metal_buffer_type_is_host(ggml_backend_buffer_type_t buft) {
return true;
UNUSED(buft);
}
GGML_CALL ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void) {
static struct ggml_backend_buffer_type ggml_backend_buffer_type_metal = {
/* .iface = */ {
llama : ggml-backend integration (#4766) * llama : ggml-backend integration * ggml-backend : add names to buffers * fix unmap after loading * batched-bench : add tensor_split param * llama : check for null tensor_split * ggml-backend : increase GGML_MAX_BACKENDS * improve graph splitting, partial fix for --no-kv-offload * cuda : add ggml-backend split buffer support * cuda : do not create buffer types for devices that don't exist (fixes usage without CUDA devices available) * ggml : fix null backend dereference (#4807) * ggml : fix null backend dereference * ggml : also check ggml_backend_is_cpu * test-backend-ops : check buffer allocation failures * llama : add cparam (split_mode) and command line argument (--split-mode, -sm) to configure the split mode (none, layer or row) * ggml : fix mul_mat_id work size * llama : rewrite session kv load/set without graphs * minor * llama : only initialize used backends, free backends on context free * llama : abort ctx if cuda backend init fails * llama : rewrite lora with ggml-backend and compute on CPU ggml-ci * llama : only map to a backend buffer the region of the file mapping containing the tensors used in the buffer * opencl : add ggml-backend buffer type * cuda : only use batched_cublas with batched mat muls (fixes fp16 tg perf) * llama : on Metal, by default offload the full model ggml-ci * metal : page align the data ptr (#4854) * Apply suggestions from code review Co-authored-by: Johannes Gäßler <johannesg@5d6.de> * cuda : fix split buffer free * address review comments * llama-bench : add split-mode parameter * fix whitespace * opencl : fix double initialization * server : add --split-mode parameter * use async copy and compute to improve multi-gpu performance ggml-ci * use async memcpys to copy the graph outputs to the CPU * fix opencl * use a host buffer for the cpu compute buffer for faster copies to the gpu --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
2024-01-12 20:07:38 +01:00
/* .get_name = */ ggml_backend_metal_buffer_type_get_name,
/* .alloc_buffer = */ ggml_backend_metal_buffer_type_alloc_buffer,
/* .get_alignment = */ ggml_backend_metal_buffer_type_get_alignment,
/* .get_max_size = */ ggml_backend_metal_buffer_type_get_max_size,
/* .get_alloc_size = */ NULL, // defaults to ggml_nbytes
/* .is_host = */ ggml_backend_metal_buffer_type_is_host,
},
/* .context = */ NULL,
};
return &ggml_backend_buffer_type_metal;
}
// buffer from ptr
GGML_CALL ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, size_t size, size_t max_size) {
struct ggml_backend_metal_buffer_context * ctx = malloc(sizeof(struct ggml_backend_metal_buffer_context));
ctx->all_data = data;
ctx->all_size = size;
ctx->owned = false;
ctx->n_buffers = 0;
const size_t size_page = sysconf(_SC_PAGESIZE);
llama : ggml-backend integration (#4766) * llama : ggml-backend integration * ggml-backend : add names to buffers * fix unmap after loading * batched-bench : add tensor_split param * llama : check for null tensor_split * ggml-backend : increase GGML_MAX_BACKENDS * improve graph splitting, partial fix for --no-kv-offload * cuda : add ggml-backend split buffer support * cuda : do not create buffer types for devices that don't exist (fixes usage without CUDA devices available) * ggml : fix null backend dereference (#4807) * ggml : fix null backend dereference * ggml : also check ggml_backend_is_cpu * test-backend-ops : check buffer allocation failures * llama : add cparam (split_mode) and command line argument (--split-mode, -sm) to configure the split mode (none, layer or row) * ggml : fix mul_mat_id work size * llama : rewrite session kv load/set without graphs * minor * llama : only initialize used backends, free backends on context free * llama : abort ctx if cuda backend init fails * llama : rewrite lora with ggml-backend and compute on CPU ggml-ci * llama : only map to a backend buffer the region of the file mapping containing the tensors used in the buffer * opencl : add ggml-backend buffer type * cuda : only use batched_cublas with batched mat muls (fixes fp16 tg perf) * llama : on Metal, by default offload the full model ggml-ci * metal : page align the data ptr (#4854) * Apply suggestions from code review Co-authored-by: Johannes Gäßler <johannesg@5d6.de> * cuda : fix split buffer free * address review comments * llama-bench : add split-mode parameter * fix whitespace * opencl : fix double initialization * server : add --split-mode parameter * use async copy and compute to improve multi-gpu performance ggml-ci * use async memcpys to copy the graph outputs to the CPU * fix opencl * use a host buffer for the cpu compute buffer for faster copies to the gpu --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
2024-01-12 20:07:38 +01:00
// page-align the data ptr
{
const uintptr_t offs = (uintptr_t) data % size_page;
data = (void *) ((char *) data - offs);
size += offs;
}
size_t size_aligned = size;
if ((size_aligned % size_page) != 0) {
size_aligned += (size_page - (size_aligned % size_page));
}
id<MTLDevice> device = ggml_backend_metal_get_device();
// the buffer fits into the max buffer size allowed by the device
if (size_aligned <= device.maxBufferLength) {
ctx->buffers[ctx->n_buffers].data = data;
ctx->buffers[ctx->n_buffers].size = size;
ctx->buffers[ctx->n_buffers].metal = [device newBufferWithBytesNoCopy:data length:size_aligned options:MTLResourceStorageModeShared deallocator:nil];
if (ctx->buffers[ctx->n_buffers].metal == nil) {
GGML_METAL_LOG_ERROR("%s: error: failed to allocate buffer, size = %8.2f MiB\n", __func__, size_aligned / 1024.0 / 1024.0);
return false;
}
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
ggml_backend_metal_log_allocated_size(device, size_aligned);
++ctx->n_buffers;
} else {
// this overlap between the views will guarantee that the tensor with the maximum size will fully fit into
// one of the views
const size_t size_ovlp = ((max_size + size_page - 1) / size_page + 1) * size_page; // round-up 2 pages just in case
const size_t size_step = device.maxBufferLength - size_ovlp;
const size_t size_view = device.maxBufferLength;
for (size_t i = 0; i < size; i += size_step) {
const size_t size_step_aligned = (i + size_view <= size) ? size_view : (size_aligned - i);
ctx->buffers[ctx->n_buffers].data = (void *) ((uint8_t *) data + i);
ctx->buffers[ctx->n_buffers].size = size_step_aligned;
ctx->buffers[ctx->n_buffers].metal = [device newBufferWithBytesNoCopy:(void *) ((uint8_t *) data + i) length:size_step_aligned options:MTLResourceStorageModeShared deallocator:nil];
if (ctx->buffers[ctx->n_buffers].metal == nil) {
GGML_METAL_LOG_ERROR("%s: error: failed to allocate buffer, size = %8.2f MiB\n", __func__, size_step_aligned / 1024.0 / 1024.0);
return false;
}
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
ggml_backend_metal_log_allocated_size(device, size_step_aligned);
if (i + size_step < size) {
GGML_METAL_LOG_INFO("\n");
}
++ctx->n_buffers;
}
}
return ggml_backend_buffer_init(ggml_backend_metal_buffer_type(), ggml_backend_metal_buffer_i, ctx, size);
}
// backend
GGML_CALL static const char * ggml_backend_metal_name(ggml_backend_t backend) {
return "Metal";
UNUSED(backend);
}
GGML_CALL static void ggml_backend_metal_free(ggml_backend_t backend) {
struct ggml_metal_context * ctx = (struct ggml_metal_context *)backend->context;
ggml_metal_free(ctx);
free(backend);
}
GGML_CALL static ggml_backend_buffer_type_t ggml_backend_metal_get_default_buffer_type(ggml_backend_t backend) {
return ggml_backend_metal_buffer_type();
UNUSED(backend);
}
GGML_CALL static enum ggml_status ggml_backend_metal_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
struct ggml_metal_context * metal_ctx = (struct ggml_metal_context *)backend->context;
return ggml_metal_graph_compute(metal_ctx, cgraph);
}
GGML_CALL static bool ggml_backend_metal_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
struct ggml_metal_context * metal_ctx = (struct ggml_metal_context *)backend->context;
return ggml_metal_supports_op(metal_ctx, op);
}
GGML_CALL static bool ggml_backend_metal_supports_buft(ggml_backend_t backend, ggml_backend_buffer_type_t buft) {
return buft->iface.get_name == ggml_backend_metal_buffer_type_get_name;
UNUSED(backend);
}
llama : ggml-backend integration (#4766) * llama : ggml-backend integration * ggml-backend : add names to buffers * fix unmap after loading * batched-bench : add tensor_split param * llama : check for null tensor_split * ggml-backend : increase GGML_MAX_BACKENDS * improve graph splitting, partial fix for --no-kv-offload * cuda : add ggml-backend split buffer support * cuda : do not create buffer types for devices that don't exist (fixes usage without CUDA devices available) * ggml : fix null backend dereference (#4807) * ggml : fix null backend dereference * ggml : also check ggml_backend_is_cpu * test-backend-ops : check buffer allocation failures * llama : add cparam (split_mode) and command line argument (--split-mode, -sm) to configure the split mode (none, layer or row) * ggml : fix mul_mat_id work size * llama : rewrite session kv load/set without graphs * minor * llama : only initialize used backends, free backends on context free * llama : abort ctx if cuda backend init fails * llama : rewrite lora with ggml-backend and compute on CPU ggml-ci * llama : only map to a backend buffer the region of the file mapping containing the tensors used in the buffer * opencl : add ggml-backend buffer type * cuda : only use batched_cublas with batched mat muls (fixes fp16 tg perf) * llama : on Metal, by default offload the full model ggml-ci * metal : page align the data ptr (#4854) * Apply suggestions from code review Co-authored-by: Johannes Gäßler <johannesg@5d6.de> * cuda : fix split buffer free * address review comments * llama-bench : add split-mode parameter * fix whitespace * opencl : fix double initialization * server : add --split-mode parameter * use async copy and compute to improve multi-gpu performance ggml-ci * use async memcpys to copy the graph outputs to the CPU * fix opencl * use a host buffer for the cpu compute buffer for faster copies to the gpu --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
2024-01-12 20:07:38 +01:00
static struct ggml_backend_i ggml_backend_metal_i = {
/* .get_name = */ ggml_backend_metal_name,
/* .free = */ ggml_backend_metal_free,
/* .get_default_buffer_type = */ ggml_backend_metal_get_default_buffer_type,
/* .set_tensor_async = */ NULL,
/* .get_tensor_async = */ NULL,
llama : ggml-backend integration (#4766) * llama : ggml-backend integration * ggml-backend : add names to buffers * fix unmap after loading * batched-bench : add tensor_split param * llama : check for null tensor_split * ggml-backend : increase GGML_MAX_BACKENDS * improve graph splitting, partial fix for --no-kv-offload * cuda : add ggml-backend split buffer support * cuda : do not create buffer types for devices that don't exist (fixes usage without CUDA devices available) * ggml : fix null backend dereference (#4807) * ggml : fix null backend dereference * ggml : also check ggml_backend_is_cpu * test-backend-ops : check buffer allocation failures * llama : add cparam (split_mode) and command line argument (--split-mode, -sm) to configure the split mode (none, layer or row) * ggml : fix mul_mat_id work size * llama : rewrite session kv load/set without graphs * minor * llama : only initialize used backends, free backends on context free * llama : abort ctx if cuda backend init fails * llama : rewrite lora with ggml-backend and compute on CPU ggml-ci * llama : only map to a backend buffer the region of the file mapping containing the tensors used in the buffer * opencl : add ggml-backend buffer type * cuda : only use batched_cublas with batched mat muls (fixes fp16 tg perf) * llama : on Metal, by default offload the full model ggml-ci * metal : page align the data ptr (#4854) * Apply suggestions from code review Co-authored-by: Johannes Gäßler <johannesg@5d6.de> * cuda : fix split buffer free * address review comments * llama-bench : add split-mode parameter * fix whitespace * opencl : fix double initialization * server : add --split-mode parameter * use async copy and compute to improve multi-gpu performance ggml-ci * use async memcpys to copy the graph outputs to the CPU * fix opencl * use a host buffer for the cpu compute buffer for faster copies to the gpu --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
2024-01-12 20:07:38 +01:00
/* .cpy_tensor_async = */ NULL,
/* .synchronize = */ NULL,
/* .graph_plan_create = */ NULL,
/* .graph_plan_free = */ NULL,
/* .graph_plan_update = */ NULL,
/* .graph_plan_compute = */ NULL,
/* .graph_compute = */ ggml_backend_metal_graph_compute,
/* .supports_op = */ ggml_backend_metal_supports_op,
/* .supports_buft = */ ggml_backend_metal_supports_buft,
/* .offload_op = */ NULL,
/* .event_new = */ NULL,
/* .event_free = */ NULL,
/* .event_record = */ NULL,
/* .event_wait = */ NULL,
/* .event_synchronize = */ NULL,
};
2024-01-13 19:45:45 +01:00
void ggml_backend_metal_log_set_callback(ggml_log_callback log_callback, void * user_data) {
ggml_metal_log_callback = log_callback;
ggml_metal_log_user_data = user_data;
}
static ggml_guid_t ggml_backend_metal_guid(void) {
static ggml_guid guid = { 0x81, 0xa1, 0x8b, 0x1e, 0x71, 0xec, 0x79, 0xed, 0x2b, 0x85, 0xdc, 0x8a, 0x61, 0x98, 0x30, 0xe6 };
return &guid;
}
ggml_backend_t ggml_backend_metal_init(void) {
struct ggml_metal_context * ctx = ggml_metal_init(GGML_DEFAULT_N_THREADS);
if (ctx == NULL) {
return NULL;
}
ggml_backend_t metal_backend = malloc(sizeof(struct ggml_backend));
*metal_backend = (struct ggml_backend) {
/* .guid = */ ggml_backend_metal_guid(),
llama : ggml-backend integration (#4766) * llama : ggml-backend integration * ggml-backend : add names to buffers * fix unmap after loading * batched-bench : add tensor_split param * llama : check for null tensor_split * ggml-backend : increase GGML_MAX_BACKENDS * improve graph splitting, partial fix for --no-kv-offload * cuda : add ggml-backend split buffer support * cuda : do not create buffer types for devices that don't exist (fixes usage without CUDA devices available) * ggml : fix null backend dereference (#4807) * ggml : fix null backend dereference * ggml : also check ggml_backend_is_cpu * test-backend-ops : check buffer allocation failures * llama : add cparam (split_mode) and command line argument (--split-mode, -sm) to configure the split mode (none, layer or row) * ggml : fix mul_mat_id work size * llama : rewrite session kv load/set without graphs * minor * llama : only initialize used backends, free backends on context free * llama : abort ctx if cuda backend init fails * llama : rewrite lora with ggml-backend and compute on CPU ggml-ci * llama : only map to a backend buffer the region of the file mapping containing the tensors used in the buffer * opencl : add ggml-backend buffer type * cuda : only use batched_cublas with batched mat muls (fixes fp16 tg perf) * llama : on Metal, by default offload the full model ggml-ci * metal : page align the data ptr (#4854) * Apply suggestions from code review Co-authored-by: Johannes Gäßler <johannesg@5d6.de> * cuda : fix split buffer free * address review comments * llama-bench : add split-mode parameter * fix whitespace * opencl : fix double initialization * server : add --split-mode parameter * use async copy and compute to improve multi-gpu performance ggml-ci * use async memcpys to copy the graph outputs to the CPU * fix opencl * use a host buffer for the cpu compute buffer for faster copies to the gpu --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
2024-01-12 20:07:38 +01:00
/* .interface = */ ggml_backend_metal_i,
/* .context = */ ctx,
};
return metal_backend;
}
bool ggml_backend_is_metal(ggml_backend_t backend) {
return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_metal_guid());
}
void ggml_backend_metal_set_n_cb(ggml_backend_t backend, int n_cb) {
GGML_ASSERT(ggml_backend_is_metal(backend));
struct ggml_metal_context * ctx = (struct ggml_metal_context *)backend->context;
2024-01-13 19:45:45 +01:00
ctx->n_cb = MIN(n_cb, GGML_METAL_MAX_BUFFERS);
}
bool ggml_backend_metal_supports_family(ggml_backend_t backend, int family) {
GGML_ASSERT(ggml_backend_is_metal(backend));
struct ggml_metal_context * ctx = (struct ggml_metal_context *)backend->context;
return [ctx->device supportsFamily:(MTLGPUFamilyApple1 + family - 1)];
}
void ggml_backend_metal_capture_next_compute(ggml_backend_t backend) {
GGML_ASSERT(ggml_backend_is_metal(backend));
struct ggml_metal_context * ctx = (struct ggml_metal_context *)backend->context;
ctx->should_capture_next_compute = true;
}
GGML_CALL ggml_backend_t ggml_backend_reg_metal_init(const char * params, void * user_data); // silence warning
GGML_CALL ggml_backend_t ggml_backend_reg_metal_init(const char * params, void * user_data) {
return ggml_backend_metal_init();
GGML_UNUSED(params);
GGML_UNUSED(user_data);
}