From 48e2b1337257bb77573bf76cfffcff3a5efa8704 Mon Sep 17 00:00:00 2001 From: adel boussaken Date: Sat, 20 Jan 2024 09:05:43 +0100 Subject: [PATCH 01/66] Add a dart/flutter binding to README.md (#4882) --- README.md | 1 + 1 file changed, 1 insertion(+) diff --git a/README.md b/README.md index 866aa87b4..cbfba01bc 100644 --- a/README.md +++ b/README.md @@ -128,6 +128,7 @@ as the main playground for developing new features for the [ggml](https://github - React Native: [mybigday/llama.rn](https://github.com/mybigday/llama.rn) - Java: [kherud/java-llama.cpp](https://github.com/kherud/java-llama.cpp) - Zig: [deins/llama.cpp.zig](https://github.com/Deins/llama.cpp.zig) +- Flutter/Dart: [netdur/llama_cpp_dart](https://github.com/netdur/llama_cpp_dart) **UI:** From 77bc1bbd05f0c31cb45773eb5eb59b9ff2b07e1b Mon Sep 17 00:00:00 2001 From: Herman Semenov Date: Sat, 20 Jan 2024 08:11:31 +0000 Subject: [PATCH 02/66] cmake : add support for ccache (#5002) * Added support ccache for speedup recompilation * cmake : option to disable ccache --------- Co-authored-by: Georgi Gerganov --- CMakeLists.txt | 12 ++++++++++++ 1 file changed, 12 insertions(+) diff --git a/CMakeLists.txt b/CMakeLists.txt index 3fc65eaf2..6b3b1396b 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -47,6 +47,7 @@ option(BUILD_SHARED_LIBS "build shared libraries" option(LLAMA_STATIC "llama: static link libraries" OFF) option(LLAMA_NATIVE "llama: enable -march=native flag" ON) option(LLAMA_LTO "llama: enable link time optimization" OFF) +option(LLAMA_CCACHE "llama: use ccache if available" ON) # debug option(LLAMA_ALL_WARNINGS "llama: enable all compiler warnings" ON) @@ -561,6 +562,17 @@ if (LLAMA_LTO) endif() endif() +if (LLAMA_CCACHE) + find_program(LLAMA_CCACHE_FOUND ccache) + if (LLAMA_CCACHE_FOUND) + set_property(GLOBAL PROPERTY RULE_LAUNCH_COMPILE ccache) + set(ENV{CCACHE_SLOPPINESS} time_macros) + message(STATUS "Using ccache") + else() + message(STATUS "Warning: ccache not found - consider installing it or use LLAMA_CCACHE=OFF") + endif () +endif() + # this version of Apple ld64 is buggy execute_process( COMMAND ${CMAKE_C_COMPILER} ${CMAKE_EXE_LINKER_FLAGS} -Wl,-v From 6df465a91d402370bcba6676b19fad85b06ce7e0 Mon Sep 17 00:00:00 2001 From: slaren Date: Sat, 20 Jan 2024 16:05:49 +0100 Subject: [PATCH 03/66] llama : run all KQV ops on the CPU with no KV offload (#5049) ggml-ci --- ggml-backend.c | 34 +++++++----- llama.cpp | 145 +++++++++++++++++++++++++++---------------------- 2 files changed, 99 insertions(+), 80 deletions(-) diff --git a/ggml-backend.c b/ggml-backend.c index ef518dae0..423512def 100644 --- a/ggml-backend.c +++ b/ggml-backend.c @@ -1191,6 +1191,24 @@ static void sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgraph * g ggml_tallocr_t src_allocr = node_allocr(src); GGML_ASSERT(src_allocr != NULL); // all inputs should be assigned by now if (src_allocr != node_allocr) { + // create a copy of the input in the split's backend + size_t id = hash_id(src); + if (sched->node_copies[id][cur_backend_id] == NULL) { + ggml_backend_t backend = get_allocr_backend(sched, cur_allocr); + struct ggml_tensor * tensor_copy = ggml_dup_tensor_layout(sched->ctx, src); + ggml_format_name(tensor_copy, "%s#%s", ggml_backend_name(backend), src->name); + + sched->node_copies[id][cur_backend_id] = tensor_copy; + node_allocr(tensor_copy) = cur_allocr; + SET_CAUSE(tensor_copy, "4.cpy"); + + int n_inputs = sched->splits[cur_split].n_inputs++; + GGML_ASSERT(n_inputs < GGML_MAX_SPLIT_INPUTS); + sched->splits[cur_split].inputs[n_inputs] = src; + } + node->src[j] = sched->node_copies[id][cur_backend_id]; + +#if 0 // check if the input is already in the split bool found = false; for (int k = 0; k < sched->splits[cur_split].n_inputs; k++) { @@ -1206,19 +1224,7 @@ static void sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgraph * g GGML_ASSERT(n_inputs < GGML_MAX_SPLIT_INPUTS); sched->splits[cur_split].inputs[n_inputs] = src; } - - // create a copy of the input in the split's backend - size_t id = hash_id(src); - if (sched->node_copies[id][cur_backend_id] == NULL) { - ggml_backend_t backend = get_allocr_backend(sched, cur_allocr); - struct ggml_tensor * tensor_copy = ggml_dup_tensor_layout(sched->ctx, src); - ggml_format_name(tensor_copy, "%s#%s", ggml_backend_name(backend), src->name); - - sched->node_copies[id][cur_backend_id] = tensor_copy; - node_allocr(tensor_copy) = cur_allocr; - SET_CAUSE(tensor_copy, "4.cpy"); - } - node->src[j] = sched->node_copies[id][cur_backend_id]; +#endif } } } @@ -1333,7 +1339,7 @@ static void sched_compute_splits(ggml_backend_sched_t sched) { uint64_t compute_start_us = ggml_time_us(); if (!sched->callback_eval) { ggml_backend_graph_compute(split_backend, &split->graph); - //ggml_backend_synchronize(split_backend); // necessary to measure compute time + //ggml_backend_synchronize(split_backend); // necessary to measure compute time } else { // similar to ggml_backend_compare_graph_backend for (int j0 = 0; j0 < split->graph.n_nodes; j0++) { diff --git a/llama.cpp b/llama.cpp index 90579ac85..909ad4ad8 100644 --- a/llama.cpp +++ b/llama.cpp @@ -4315,6 +4315,7 @@ static struct ggml_tensor * llm_build_kqv( const llama_model & model, const llama_hparams & hparams, const llama_kv_cache & kv, + struct ggml_cgraph * graph, struct ggml_tensor * wo, struct ggml_tensor * wo_b, struct ggml_tensor * q_cur, @@ -4393,6 +4394,8 @@ static struct ggml_tensor * llm_build_kqv( struct ggml_tensor * cur = ggml_cont_2d(ctx, kqv_merged, n_embd_head_k*n_head, n_tokens); cb(cur, "kqv_merged_cont", il); + ggml_build_forward_expand(graph, cur); + cur = ggml_mul_mat(ctx, wo, cur); if (wo_b) { cb(cur, "kqv_wo", il); @@ -4405,6 +4408,44 @@ static struct ggml_tensor * llm_build_kqv( return cur; } +static struct ggml_tensor * llm_build_kv( + struct ggml_context * ctx, + const llama_model & model, + const llama_hparams & hparams, + const llama_kv_cache & kv, + struct ggml_cgraph * graph, + struct ggml_tensor * wo, + struct ggml_tensor * wo_b, + struct ggml_tensor * k_cur, + struct ggml_tensor * v_cur, + struct ggml_tensor * q_cur, + struct ggml_tensor * kq_mask, + int64_t n_ctx, + int32_t n_tokens, + int32_t kv_head, + int32_t n_kv, + float max_alibi_bias, + float kq_scale, + const llm_build_cb & cb, + int il) { + + // these nodes are added to the graph together so that they are not reordered + // by doing so, the number of splits in the graph is reduced + ggml_build_forward_expand(graph, k_cur); + ggml_build_forward_expand(graph, v_cur); + ggml_build_forward_expand(graph, q_cur); + + llm_build_kv_store(ctx, hparams, kv, graph, k_cur, v_cur, n_ctx, n_tokens, kv_head, cb, il); + + struct ggml_tensor * cur; + cur = llm_build_kqv(ctx, model, hparams, kv, graph, + wo, wo_b, + q_cur, kq_mask, n_ctx, n_tokens, n_kv, max_alibi_bias, kq_scale, cb, il); + cb(cur, "kqv_out", il); + + return cur; +} + struct llm_build_context { const llama_model & model; const llama_hparams & hparams; @@ -4562,12 +4603,6 @@ struct llm_build_context { cb(Vcur, "Vcur", il); } - // these nodes are added to the graph together so that they are not reordered - // by doing so, the number of splits in the graph is reduced - ggml_build_forward_expand(gf, Qcur); - ggml_build_forward_expand(gf, Kcur); - ggml_build_forward_expand(gf, Vcur); - Qcur = ggml_rope_custom( ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, hparams.n_rot, 0, 0, n_orig_ctx, freq_base, freq_scale, @@ -4582,11 +4617,9 @@ struct llm_build_context { ); cb(Kcur, "Kcur", il); - llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il); - - cur = llm_build_kqv(ctx0, model, hparams, kv_self, + cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, model.layers[il].wo, model.layers[il].bo, - Qcur, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il); + Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il); cb(cur, "kqv_out", il); } @@ -4763,14 +4796,13 @@ struct llm_build_context { cb(Qcur, "Qcur", il); cb(Kcur, "Kcur", il); - llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il); // apply ALiBi for 13B model const float max_alibi_bias = model.type == MODEL_13B ? 8.0f : -1.0f; - cur = llm_build_kqv(ctx0, model, hparams, kv_self, + cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, model.layers[il].wo, NULL, - Qcur, KQ_mask, n_ctx, n_tokens, n_kv, max_alibi_bias, 1.0f/sqrtf(float(n_embd_head)), cb, il); + Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, max_alibi_bias, 1.0f/sqrtf(float(n_embd_head)), cb, il); cb(cur, "kqv_out", il); } @@ -4892,11 +4924,9 @@ struct llm_build_context { ); cb(Kcur, "Kcur", il); - llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il); - - cur = llm_build_kqv(ctx0, model, hparams, kv_self, + cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, model.layers[il].wo, NULL, - Qcur, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il); + Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il); cb(cur, "kqv_out", il); } @@ -4993,11 +5023,9 @@ struct llm_build_context { Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il); - - cur = llm_build_kqv(ctx0, model, hparams, kv_self, + cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, model.layers[il].wo, model.layers[il].bo, - Qcur, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il); + Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il); cb(cur, "kqv_out", il); } @@ -5200,12 +5228,9 @@ struct llm_build_context { ); cb(Vcur, "Vcur", il); - llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il); - - // TODO: not tested, could be broken - cur = llm_build_kqv(ctx0, model, hparams, kv_self, + cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, model.layers[il].wo, model.layers[il].bo, - Q, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il); + Kcur, Vcur, Q, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il); cb(cur, "kqv_out", il); } @@ -5292,11 +5317,9 @@ struct llm_build_context { Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); cb(Qcur, "Qcur", il); - llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il); - - cur = llm_build_kqv(ctx0, model, hparams, kv_self, + cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, model.layers[il].wo, NULL, - Qcur, KQ_mask, n_ctx, n_tokens, n_kv, 8.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il); + Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, 8.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il); cb(cur, "kqv_out", il); } @@ -5390,11 +5413,9 @@ struct llm_build_context { Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il); - - cur = llm_build_kqv(ctx0, model, hparams, kv_self, + cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, model.layers[il].wo, model.layers[il].bo, - Qcur, KQ_mask, n_ctx, n_tokens, n_kv, 8.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il); + Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, 8.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il); cb(cur, "kqv_out", il); } @@ -5485,11 +5506,9 @@ struct llm_build_context { Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il); - - cur = llm_build_kqv(ctx0, model, hparams, kv_self, + cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, model.layers[il].wo, NULL, - Qcur, KQ_mask, n_ctx, n_tokens, n_kv, hparams.f_max_alibi_bias, 1.0f/sqrtf(float(n_embd_head)), cb, il); + Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, hparams.f_max_alibi_bias, 1.0f/sqrtf(float(n_embd_head)), cb, il); cb(cur, "kqv_out", il); } @@ -5597,11 +5616,9 @@ struct llm_build_context { ); cb(Kcur, "Kcur", il); - llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il); - - cur = llm_build_kqv(ctx0, model, hparams, kv_self, + cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, model.layers[il].wo, NULL, - Qcur, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il); + Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il); cb(cur, "kqv_out", il); } @@ -5714,11 +5731,9 @@ struct llm_build_context { ); cb(Kcur, "Kcur", il); - llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il); - - cur = llm_build_kqv(ctx0, model, hparams, kv_self, + cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, model.layers[il].wo, NULL, - Qcur, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il); + Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il); cb(cur, "kqv_out", il); } @@ -5837,11 +5852,9 @@ struct llm_build_context { ); cb(Kcur, "Kcur", il); - llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il); - - cur = llm_build_kqv(ctx0, model, hparams, kv_self, + cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, model.layers[il].wo, model.layers[il].bo, - Qcur, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il); + Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il); cb(cur, "kqv_out", il); } @@ -5966,11 +5979,9 @@ struct llm_build_context { ); cb(Kcur, "Kcur", il); - llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il); - - cur = llm_build_kqv(ctx0, model, hparams, kv_self, + cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, model.layers[il].wo, model.layers[il].bo, - Qcur, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, 1.0f, cb, il); + Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, -1.0f, 1.0f, cb, il); cb(cur, "kqv_out", il); } @@ -6071,11 +6082,9 @@ struct llm_build_context { ext_factor, attn_factor, beta_fast, beta_slow); cb(Kcur, "Kcur", il); - llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il); - - cur = llm_build_kqv(ctx0, model, hparams, kv_self, + cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, model.layers[il].wo, NULL, - Qcur, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il); + Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il); cb(cur, "kqv_out", il); } struct ggml_tensor * sa_out = cur; @@ -6172,11 +6181,9 @@ struct llm_build_context { Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il); - - cur = llm_build_kqv(ctx0, model, hparams, kv_self, + cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, model.layers[il].wo, model.layers[il].bo, - Qcur, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il); + Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il); cb(cur, "kqv_out", il); } @@ -6283,11 +6290,9 @@ struct llm_build_context { ); cb(Kcur, "Kcur", il); - llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il); - - cur = llm_build_kqv(ctx0, model, hparams, kv_self, + cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, model.layers[il].wo, model.layers[il].bo, - Qcur, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il); + Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il); cb(cur, "kqv_out", il); } @@ -6355,6 +6360,14 @@ static struct ggml_cgraph * llama_build_graph( ggml_set_name(cur, name); } + + if (!lctx.cparams.offload_kqv) { + if (strcmp(name, "kqv_merged_cont") == 0) { + // all nodes between the KV store and the attention output are run on the CPU + ggml_backend_sched_set_node_backend(lctx.sched, cur, lctx.backend_cpu); + } + } + // // allocate input tensors and set input data // From 97c1549808d2742d37584a3c9df28154bdf34417 Mon Sep 17 00:00:00 2001 From: Jared Van Bortel Date: Sat, 20 Jan 2024 10:08:08 -0500 Subject: [PATCH 04/66] perplexity : fix MSVC build after #5020 (#5043) * perplexity : fix MSVC build after #5020 * try a differerent fix --- examples/perplexity/perplexity.cpp | 12 ++++++------ 1 file changed, 6 insertions(+), 6 deletions(-) diff --git a/examples/perplexity/perplexity.cpp b/examples/perplexity/perplexity.cpp index b07320190..f91f5795a 100644 --- a/examples/perplexity/perplexity.cpp +++ b/examples/perplexity/perplexity.cpp @@ -458,23 +458,24 @@ static bool decode_helper(llama_context * ctx, llama_batch & batch, std::vector< return true; } +#define K_TOKEN_CHUNK 4 + static void compute_logprobs(const float * batch_logits, int n_vocab, std::vector& workers, const std::vector>& eval_pairs, std::vector& eval_results) { - constexpr int k_token_chunk = 4; if (eval_results.size() != eval_pairs.size()) { eval_results.resize(eval_pairs.size()); } if (eval_pairs.empty()) return; - size_t max_threads = std::min((eval_pairs.size() + k_token_chunk - 1)/k_token_chunk, workers.size()); + size_t max_threads = std::min((eval_pairs.size() + K_TOKEN_CHUNK - 1)/K_TOKEN_CHUNK, workers.size()); std::atomic counter(0); auto compute = [&counter, &eval_pairs, &eval_results, batch_logits, n_vocab] () { - float local_logprobs[k_token_chunk]; + float local_logprobs[K_TOKEN_CHUNK]; while (true) { - size_t first = counter.fetch_add(k_token_chunk, std::memory_order_relaxed); + size_t first = counter.fetch_add(K_TOKEN_CHUNK, std::memory_order_relaxed); if (first >= eval_results.size()) break; - size_t last = std::min(first + k_token_chunk, eval_results.size()); + size_t last = std::min(first + K_TOKEN_CHUNK, eval_results.size()); for (size_t i = first; i < last; ++i) { auto logits = batch_logits + eval_pairs[i].first * n_vocab; float max_logit = logits[0]; @@ -497,7 +498,6 @@ static void compute_logprobs(const float * batch_logits, int n_vocab, std::vecto for (size_t it = 0; it < max_threads; ++it) { workers[it].join(); } - } static void hellaswag_score(llama_context * ctx, const gpt_params & params) { From b43ebde3b0ccbc42d9dd782b32e2fd8eb35b43b5 Mon Sep 17 00:00:00 2001 From: Jared Van Bortel Date: Sat, 20 Jan 2024 18:14:18 -0500 Subject: [PATCH 05/66] convert : partially revert PR #4818 (#5041) --- convert-hf-to-gguf.py | 9 +- convert-llama-ggml-to-gguf.py | 14 +- convert-lora-to-ggml.py | 3 +- convert-persimmon-to-gguf.py | 12 +- convert.py | 627 ++++++++++++---------------------- mypy.ini | 1 + 6 files changed, 237 insertions(+), 429 deletions(-) diff --git a/convert-hf-to-gguf.py b/convert-hf-to-gguf.py index 5cb3e63fb..4d995ef78 100755 --- a/convert-hf-to-gguf.py +++ b/convert-hf-to-gguf.py @@ -10,7 +10,7 @@ import re import sys from enum import IntEnum from pathlib import Path -from typing import TYPE_CHECKING, Any, ContextManager, Iterator, cast, Optional +from typing import TYPE_CHECKING, Any, ContextManager, Iterator, cast import numpy as np import torch @@ -487,7 +487,8 @@ class MPTModel(Model): # map tensor names if "scales" in name: new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias", ".scales")) - new_name = new_name.replace("scales", "act.scales") + if new_name is not None: + new_name = new_name.replace("scales", "act.scales") else: new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias")) if new_name is None: @@ -904,7 +905,7 @@ class QwenModel(Model): return ''.join([byte_encoder[ord(char)] for char in b.decode('latin-1')]) @staticmethod - def bpe(mergeable_ranks: dict[bytes, int], token: bytes, max_rank: Optional[int] = None) -> list[bytes]: + def bpe(mergeable_ranks: dict[bytes, int], token: bytes, max_rank: int | None = None) -> list[bytes]: parts = [bytes([b]) for b in token] while True: min_idx = None @@ -1285,7 +1286,7 @@ def main() -> None: if args.awq_path: sys.path.insert(1, str(Path(__file__).parent / 'awq-py')) - from awq.apply_awq import add_scale_weights + from awq.apply_awq import add_scale_weights # type: ignore[import-not-found] tmp_model_path = args.model / "weighted_model" dir_model = tmp_model_path if tmp_model_path.is_dir(): diff --git a/convert-llama-ggml-to-gguf.py b/convert-llama-ggml-to-gguf.py index e359330af..b33108062 100755 --- a/convert-llama-ggml-to-gguf.py +++ b/convert-llama-ggml-to-gguf.py @@ -2,6 +2,7 @@ from __future__ import annotations import argparse +import os import struct import sys from enum import IntEnum @@ -9,7 +10,6 @@ from pathlib import Path import numpy as np -import os if 'NO_LOCAL_GGUF' not in os.environ: sys.path.insert(1, str(Path(__file__).parent / 'gguf-py')) import gguf @@ -371,15 +371,11 @@ def handle_metadata(cfg, hp): params = convert.Params.loadOriginalParamsJson(fakemodel, orig_config_path) else: raise ValueError('Unable to load metadata') - vocab = convert.load_vocab( - cfg.vocab_dir if cfg.vocab_dir is not None else cfg.model_metadata_dir, - cfg.vocabtype) - # FIXME: Respect cfg.vocab_dir? - svocab = gguf.SpecialVocab(cfg.model_metadata_dir, - load_merges = cfg.vocabtype == 'bpe', - n_vocab = vocab.vocab_size) + vocab_path = Path(cfg.vocab_dir if cfg.vocab_dir is not None else cfg.model_metadata_dir) + vocab_factory = convert.VocabFactory(vocab_path) + vocab, special_vocab = vocab_factory.load_vocab(cfg.vocabtype, cfg.model_metadata_dir) convert.check_vocab_size(params, vocab) - return (params, vocab, svocab) + return params, vocab, special_vocab def handle_args(): diff --git a/convert-lora-to-ggml.py b/convert-lora-to-ggml.py index 35ce152f4..4904bf128 100755 --- a/convert-lora-to-ggml.py +++ b/convert-lora-to-ggml.py @@ -5,17 +5,16 @@ import json import os import struct import sys +from pathlib import Path from typing import Any, BinaryIO, Sequence import numpy as np import torch -from pathlib import Path if 'NO_LOCAL_GGUF' not in os.environ: sys.path.insert(1, str(Path(__file__).parent / 'gguf-py' / 'gguf')) import gguf - NUMPY_TYPE_TO_FTYPE: dict[str, int] = {"float32": 0, "float16": 1} diff --git a/convert-persimmon-to-gguf.py b/convert-persimmon-to-gguf.py index 1ba5864dc..d2be805d1 100755 --- a/convert-persimmon-to-gguf.py +++ b/convert-persimmon-to-gguf.py @@ -1,11 +1,13 @@ #!/usr/bin/env python3 -import torch -import os -from pprint import pprint -import sys import argparse +import os +import sys from pathlib import Path +from pprint import pprint + +import torch from sentencepiece import SentencePieceProcessor + if 'NO_LOCAL_GGUF' not in os.environ: sys.path.insert(1, str(Path(__file__).parent / 'gguf-py')) import gguf @@ -69,7 +71,7 @@ def main(): persimmon_model = torch.load(args.ckpt_path) hparams = persimmon_model['args'] pprint(hparams) - tensors = {} + tensors: dict[str, torch.Tensor] = {} _flatten_dict(persimmon_model['model'], tensors, None) arch = gguf.MODEL_ARCH.PERSIMMON diff --git a/convert.py b/convert.py index 980e6fc72..06768033d 100755 --- a/convert.py +++ b/convert.py @@ -17,58 +17,28 @@ import signal import struct import sys import time -import warnings import zipfile from abc import ABCMeta, abstractmethod -from argparse import ArgumentParser from concurrent.futures import ProcessPoolExecutor, ThreadPoolExecutor from dataclasses import dataclass from pathlib import Path -from typing import ( - IO, - TYPE_CHECKING, - Any, - Callable, - Iterable, - Literal, - Optional, - Tuple, - TypeVar, -) +from typing import IO, TYPE_CHECKING, Any, Callable, Iterable, Literal, TypeVar import numpy as np from sentencepiece import SentencePieceProcessor -try: - from transformers import AutoTokenizer -except ModuleNotFoundError as e: - warnings.warn(f"Could not import AutoTokenizer from transformers: {e}") +if 'NO_LOCAL_GGUF' not in os.environ: + sys.path.insert(1, str(Path(__file__).parent / 'gguf-py')) +import gguf -# If NO_LOCAL_GGUF is not set, try to import gguf from the local gguf-py directory -if "NO_LOCAL_GGUF" not in os.environ: - # Use absolute path to the gguf-py directory - gguf_py_dir = str(Path(__file__).resolve().parent / "gguf-py") - print(gguf_py_dir) # NOTE: Remove this once path is verified after changes are completed - if gguf_py_dir not in sys.path: - sys.path.insert(1, gguf_py_dir) +if TYPE_CHECKING: + from typing import TypeAlias -# Import gguf module -try: - import gguf -except ModuleNotFoundError as e: - print(f"Could not import gguf: {e}") - sys.exit(1) - -if TYPE_CHECKING: # NOTE: This isn't necessary. - from typing import TypeAlias # This can technically be omitted. - -if hasattr(faulthandler, "register") and hasattr(signal, "SIGUSR1"): +if hasattr(faulthandler, 'register') and hasattr(signal, 'SIGUSR1'): faulthandler.register(signal.SIGUSR1) -# NOTE: n-dimensional arrays should be directly referenced -NDArray: TypeAlias = "np.ndarray[Any, Any]" +NDArray: TypeAlias = 'np.ndarray[Any, Any]' -# Why is this here? LLAMA and GPT are technically the only compatible ARCHs. ARCH = gguf.MODEL_ARCH.LLAMA DEFAULT_CONCURRENCY = 8 @@ -78,7 +48,6 @@ DEFAULT_CONCURRENCY = 8 # -# TODO: Clean up and refactor data types @dataclass(frozen=True) class DataType: name: str @@ -183,85 +152,65 @@ GGML_FILE_TYPE_TO_DATA_TYPE: dict[GGMLFileType, DataType] = { @dataclass class Params: - n_vocab: int - n_embd: int - n_layer: int - n_ctx: int - n_ff: int - n_head: int - n_head_kv: int - f_norm_eps: Optional[float] = None - n_experts: Optional[int] = None - n_experts_used: Optional[int] = None + n_vocab: int + n_embd: int + n_layer: int + n_ctx: int + n_ff: int + n_head: int + n_head_kv: int + n_experts: int | None = None + n_experts_used: int | None = None + f_norm_eps: float | None = None - rope_scaling_type: Optional[gguf.RopeScalingType] = None - f_rope_freq_base: Optional[float] = None - f_rope_scale: Optional[float] = None - n_orig_ctx: Optional[int] = None - rope_finetuned: Optional[bool] = None + rope_scaling_type: gguf.RopeScalingType | None = None + f_rope_freq_base: float | None = None + f_rope_scale: float | None = None + n_orig_ctx: int | None = None + rope_finetuned: bool | None = None - ftype: Optional[GGMLFileType] = None + ftype: GGMLFileType | None = None # path to the directory containing the model files - path_model: Optional[Path] = None + path_model: Path | None = None @staticmethod - def guessed(model: LazyModel) -> "Params": + def guessed(model: LazyModel) -> Params: # try transformer naming first - n_vocab, n_embd = ( - model["model.embed_tokens.weight"].shape - if "model.embed_tokens.weight" in model - else model["tok_embeddings.weight"].shape - ) + n_vocab, n_embd = model["model.embed_tokens.weight"].shape if "model.embed_tokens.weight" in model else model["tok_embeddings.weight"].shape # try transformer naming first if "model.layers.0.self_attn.q_proj.weight" in model: - n_layer = next( - i - for i in itertools.count() - if f"model.layers.{i}.self_attn.q_proj.weight" not in model - ) - elif ( - "model.layers.0.self_attn.W_pack.weight" in model - ): # next: try baichuan naming - n_layer = next( - i - for i in itertools.count() - if f"model.layers.{i}.self_attn.W_pack.weight" not in model - ) + n_layer = next(i for i in itertools.count() if f"model.layers.{i}.self_attn.q_proj.weight" not in model) + elif "model.layers.0.self_attn.W_pack.weight" in model: # next: try baichuan naming + n_layer = next(i for i in itertools.count() if f"model.layers.{i}.self_attn.W_pack.weight" not in model) else: - n_layer = next( - i - for i in itertools.count() - if f"layers.{i}.attention.wq.weight" not in model - ) + n_layer = next(i for i in itertools.count() if f"layers.{i}.attention.wq.weight" not in model) if n_layer < 1: - raise Exception( - "failed to guess 'n_layer'. This model is unknown or unsupported.\n" - "Suggestion: provide 'config.json' of the model in the same directory containing model files." - ) + raise Exception("failed to guess 'n_layer'. This model is unknown or unsupported.\n" + "Suggestion: provide 'config.json' of the model in the same directory containing model files.") - n_head = n_embd // 128 # guessed - n_mult = 256 # guessed + n_head = n_embd // 128 # guessed + n_mult = 256 # guessed # TODO: verify this n_ff = int(2 * (4 * n_embd) / 3) n_ff = n_mult * ((n_ff + n_mult - 1) // n_mult) return Params( - n_vocab=n_vocab, - n_embd=n_embd, - n_layer=n_layer, - n_ctx=-1, - n_ff=n_ff, - n_head=n_head, - n_head_kv=n_head, - f_norm_eps=1e-5, + n_vocab = n_vocab, + n_embd = n_embd, + n_layer = n_layer, + n_ctx = -1, + n_ff = n_ff, + n_head = n_head, + n_head_kv = n_head, + f_norm_eps = 1e-5, ) @staticmethod - def load_transformers_config(model: LazyModel, config_path: Path) -> "Params": + def loadHFTransformerJson(model: LazyModel, config_path: Path) -> Params: config = json.load(open(config_path)) rope_scaling_type = f_rope_scale = n_orig_ctx = rope_finetuned = None @@ -274,22 +223,20 @@ class Params: rope_scaling_type = gguf.RopeScalingType.LINEAR elif typ == "yarn": rope_scaling_type = gguf.RopeScalingType.YARN - n_orig_ctx = rope_scaling["original_max_position_embeddings"] - rope_finetuned = rope_scaling["finetuned"] + n_orig_ctx = rope_scaling['original_max_position_embeddings'] + rope_finetuned = rope_scaling['finetuned'] else: - raise NotImplementedError(f"Unknown rope scaling type: {typ}") + raise NotImplementedError(f'Unknown rope scaling type: {typ}') if "max_sequence_length" in config: n_ctx = config["max_sequence_length"] elif "max_position_embeddings" in config: n_ctx = config["max_position_embeddings"] else: - raise Exception( - "failed to guess 'n_ctx'. This model is unknown or unsupported.\n" - "Suggestion: provide 'config.json' of the model in the same directory containing model files." - ) + raise Exception("failed to guess 'n_ctx'. This model is unknown or unsupported.\n" + "Suggestion: provide 'config.json' of the model in the same directory containing model files.") - n_experts = None + n_experts = None n_experts_used = None if "num_local_experts" in config: @@ -297,30 +244,30 @@ class Params: n_experts_used = config["num_experts_per_tok"] return Params( - n_vocab=config["vocab_size"], - n_embd=config["hidden_size"], - n_layer=config["num_hidden_layers"], - n_ctx=n_ctx, - n_ff=config["intermediate_size"], - n_head=(n_head := config["num_attention_heads"]), - n_head_kv=config.get("num_key_value_heads", n_head), - n_experts=n_experts, - n_experts_used=n_experts_used, - f_norm_eps=config["rms_norm_eps"], - f_rope_freq_base=config.get("rope_theta"), - rope_scaling_type=rope_scaling_type, - f_rope_scale=f_rope_scale, - n_orig_ctx=n_orig_ctx, - rope_finetuned=rope_finetuned, + n_vocab = config["vocab_size"], + n_embd = config["hidden_size"], + n_layer = config["num_hidden_layers"], + n_ctx = n_ctx, + n_ff = config["intermediate_size"], + n_head = (n_head := config["num_attention_heads"]), + n_head_kv = config.get("num_key_value_heads", n_head), + n_experts = n_experts, + n_experts_used = n_experts_used, + f_norm_eps = config["rms_norm_eps"], + f_rope_freq_base = config.get("rope_theta"), + rope_scaling_type = rope_scaling_type, + f_rope_scale = f_rope_scale, + n_orig_ctx = n_orig_ctx, + rope_finetuned = rope_finetuned, ) # LLaMA v2 70B params.json # {"dim": 8192, "multiple_of": 4096, "ffn_dim_multiplier": 1.3, "n_heads": 64, "n_kv_heads": 8, "n_layers": 80, "norm_eps": 1e-05, "vocab_size": -1} @staticmethod - def load_torch_params(model: LazyModel, config_path: Path) -> "Params": + def loadOriginalParamsJson(model: LazyModel, config_path: Path) -> Params: config = json.load(open(config_path)) - n_experts = None + n_experts = None n_experts_used = None f_rope_freq_base = None @@ -343,50 +290,50 @@ class Params: if config.get("moe"): n_ff = model["layers.0.feed_forward.experts.0.w1.weight"].shape[0] - n_experts = config["moe"]["num_experts"] + n_experts = config["moe"]["num_experts"] n_experts_used = config["moe"]["num_experts_per_tok"] f_rope_freq_base = 1e6 return Params( - n_vocab=model["tok_embeddings.weight"].shape[0], - n_embd=config["dim"], - n_layer=config["n_layers"], - n_ctx=n_ctx, - n_ff=n_ff, - n_head=(n_head := config["n_heads"]), - n_head_kv=config.get("n_kv_heads", n_head), - n_experts=n_experts, - n_experts_used=n_experts_used, - f_norm_eps=config["norm_eps"], - f_rope_freq_base=config.get("rope_theta", f_rope_freq_base), + n_vocab = model["tok_embeddings.weight"].shape[0], + n_embd = config["dim"], + n_layer = config["n_layers"], + n_ctx = n_ctx, + n_ff = n_ff, + n_head = (n_head := config["n_heads"]), + n_head_kv = config.get("n_kv_heads", n_head), + n_experts = n_experts, + n_experts_used = n_experts_used, + f_norm_eps = config["norm_eps"], + f_rope_freq_base = config.get("rope_theta", f_rope_freq_base), ) @staticmethod - def load(model_plus: ModelPlus) -> "Params": - hf_config_path = model_plus.paths[0].parent / "config.json" + def load(model_plus: ModelPlus) -> Params: + hf_config_path = model_plus.paths[0].parent / "config.json" orig_config_path = model_plus.paths[0].parent / "params.json" if hf_config_path.exists(): - params = Params.load_transformers_config(model_plus.model, hf_config_path) + params = Params.loadHFTransformerJson(model_plus.model, hf_config_path) elif orig_config_path.exists(): - params = Params.load_torch_params(model_plus.model, orig_config_path) - elif model_plus.format != "none": + params = Params.loadOriginalParamsJson(model_plus.model, orig_config_path) + elif model_plus.format != 'none': params = Params.guessed(model_plus.model) else: - raise ValueError("Cannot guess params when model format is none") + raise ValueError('Cannot guess params when model format is none') params.path_model = model_plus.paths[0].parent return params -class BpeVocab: # GPT - def __init__( - self, fname_tokenizer: Path, fname_added_tokens: Optional[Path] - ) -> None: - self.bpe_tokenizer = json.loads( - open(str(fname_tokenizer), encoding="utf-8").read() - ) +# +# vocab +# + +class BpeVocab: + def __init__(self, fname_tokenizer: Path, fname_added_tokens: Path | None) -> None: + self.bpe_tokenizer = json.loads(open(str(fname_tokenizer), encoding="utf-8").read()) self.vocab = self.bpe_tokenizer["model"]["vocab"] added_tokens: dict[str, int] if fname_added_tokens is not None: @@ -394,34 +341,31 @@ class BpeVocab: # GPT added_tokens = json.load(open(fname_added_tokens, encoding="utf-8")) else: # Fall back to trying to find the added tokens in tokenizer.json - tokenizer_json_file = fname_tokenizer.parent / "tokenizer.json" + tokenizer_json_file = fname_tokenizer.parent / 'tokenizer.json' if not tokenizer_json_file.is_file(): added_tokens = {} else: tokenizer_json = json.load(open(tokenizer_json_file, encoding="utf-8")) added_tokens = dict( - (item["content"], item["id"]) - for item in tokenizer_json.get("added_tokens", []) + (item['content'], item['id']) + for item in tokenizer_json.get('added_tokens', []) # Added tokens here can be duplicates of the main vocabulary. - if item["content"] not in self.bpe_tokenizer - ) + if item['content'] not in self.bpe_tokenizer) vocab_size: int = len(self.vocab) - expected_ids = list(range(vocab_size, vocab_size + len(added_tokens))) - actual_ids = sorted(added_tokens.values()) + expected_ids = list(range(vocab_size, vocab_size + len(added_tokens))) + actual_ids = sorted(added_tokens.values()) if expected_ids != actual_ids: expected_end_id = vocab_size + len(actual_ids) - 1 - raise Exception( - f"Expected the {len(actual_ids)} added token ID(s) to be sequential in the range {vocab_size} - {expected_end_id}; got {actual_ids}" - ) + raise Exception(f"Expected the {len(actual_ids)} added token ID(s) to be sequential in the range {vocab_size} - {expected_end_id}; got {actual_ids}") items = sorted(added_tokens.items(), key=lambda text_idx: text_idx[1]) - self.added_tokens_dict = added_tokens - self.added_tokens_list = [text for (text, idx) in items] + self.added_tokens_dict = added_tokens + self.added_tokens_list = [text for (text, idx) in items] self.vocab_size_base: int = vocab_size - self.vocab_size: int = self.vocab_size_base + len(self.added_tokens_list) - self.fname_tokenizer = fname_tokenizer - self.fname_added_tokens = fname_added_tokens + self.vocab_size: int = self.vocab_size_base + len(self.added_tokens_list) + self.fname_tokenizer = fname_tokenizer + self.fname_added_tokens = fname_added_tokens def bpe_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]: reverse_vocab = {id: encoded_tok for encoded_tok, id in self.vocab.items()} @@ -442,10 +386,8 @@ class BpeVocab: # GPT return f"" -class SentencePieceVocab: # LlaMa - def __init__( - self, fname_tokenizer: Path, fname_added_tokens: Optional[Path] - ) -> None: +class SentencePieceVocab: + def __init__(self, fname_tokenizer: Path, fname_added_tokens: Path | None) -> None: self.sentencepiece_tokenizer = SentencePieceProcessor(str(fname_tokenizer)) added_tokens: dict[str, int] if fname_added_tokens is not None: @@ -455,23 +397,19 @@ class SentencePieceVocab: # LlaMa vocab_size: int = self.sentencepiece_tokenizer.vocab_size() - new_tokens = { - id: piece for piece, id in added_tokens.items() if id >= vocab_size - } + new_tokens = {id: piece for piece, id in added_tokens.items() if id >= vocab_size} expected_new_ids = list(range(vocab_size, vocab_size + len(new_tokens))) - actual_new_ids = sorted(new_tokens.keys()) + actual_new_ids = sorted(new_tokens.keys()) if expected_new_ids != actual_new_ids: - raise ValueError( - f"Expected new token IDs {expected_new_ids} to be sequential; got {actual_new_ids}" - ) + raise ValueError(f"Expected new token IDs {expected_new_ids} to be sequential; got {actual_new_ids}") # Token pieces that were added to the base vocabulary. self.added_tokens_dict = added_tokens - self.added_tokens_list = [new_tokens[id] for id in actual_new_ids] - self.vocab_size_base = vocab_size - self.vocab_size = self.vocab_size_base + len(self.added_tokens_list) - self.fname_tokenizer = fname_tokenizer + self.added_tokens_list = [new_tokens[id] for id in actual_new_ids] + self.vocab_size_base = vocab_size + self.vocab_size = self.vocab_size_base + len(self.added_tokens_list) + self.fname_tokenizer = fname_tokenizer self.fname_added_tokens = fname_added_tokens def sentencepiece_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]: @@ -512,11 +450,15 @@ class SentencePieceVocab: # LlaMa class HfVocab: - def __init__( - self, - fname_tokenizer: Path, - fname_added_tokens: Optional[Path] = None, - ) -> None: + def __init__(self, fname_tokenizer: Path, fname_added_tokens: Path | None = None) -> None: + try: + from transformers import AutoTokenizer + except ImportError as e: + raise ImportError( + "To use HfVocab, please install the `transformers` package. " + "You can install it with `pip install transformers`." + ) from e + print("fname_tokenizer:", fname_tokenizer) # Allow the tokenizer to default to slow or fast versions. # Explicitly set tokenizer to use local paths. @@ -529,7 +471,7 @@ class HfVocab: # Initialize lists and dictionaries for added tokens self.added_tokens_list = [] self.added_tokens_dict = dict() - self.added_tokens_ids = set() + self.added_tokens_ids = set() # Process added tokens for tok, tokidx in sorted( @@ -550,12 +492,12 @@ class HfVocab: # Set vocabulary sizes self.vocab_size_base = self.tokenizer.vocab_size - self.vocab_size = self.vocab_size_base + len(self.added_tokens_list) + self.vocab_size = self.vocab_size_base + len(self.added_tokens_list) - self.fname_tokenizer = fname_tokenizer + self.fname_tokenizer = fname_tokenizer self.fname_added_tokens = fname_added_tokens - def hf_tokens(self) -> Iterable[Tuple[bytes, float, gguf.TokenType]]: + def hf_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]: reverse_vocab = { id: encoded_tok for encoded_tok, id in self.tokenizer.get_vocab().items() } @@ -573,11 +515,9 @@ class HfVocab: token_id, self.special_ids # Reuse already stored special IDs ) - def get_token_type(self, token_id: int, special_ids: set) -> gguf.TokenType: + def get_token_type(self, token_id: int, special_ids: set[int]) -> gguf.TokenType: # Determine token type based on whether it's a special token - return ( - gguf.TokenType.CONTROL if token_id in special_ids else gguf.TokenType.NORMAL - ) + return gguf.TokenType.CONTROL if token_id in special_ids else gguf.TokenType.NORMAL def get_token_score(self, token_id: int) -> float: # Placeholder for actual logic to determine the token's score @@ -589,7 +529,6 @@ class HfVocab: if text in self.specials: toktype = self.get_token_type(self.specials[text], self.special_ids) score = self.get_token_score(self.specials[text]) - else: toktype = gguf.TokenType.USER_DEFINED score = -1000.0 @@ -783,7 +722,7 @@ def merge_multifile_models(models_plus: list[ModelPlus]) -> ModelPlus: else: model = merge_sharded([mp.model for mp in models_plus]) - return ModelPlus(model, paths, format, vocab) + return ModelPlus(model, paths, format, vocab) # pytype: disable=wrong-arg-types def permute_lazy(lazy_tensor: LazyTensor, n_head: int, n_head_kv: int) -> LazyTensor: @@ -871,17 +810,13 @@ class LazyUnpickler(pickle.Unpickler): CLASSES: dict[tuple[str, str], Any] = { # getattr used here as a workaround for mypy not being smart enough to determine # the staticmethods have a __func__ attribute. - ("torch._tensor", "_rebuild_from_type_v2"): getattr( - rebuild_from_type_v2, "__func__" - ), - ("torch._utils", "_rebuild_tensor_v2"): getattr( - lazy_rebuild_tensor_v2, "__func__" - ), - ("torch", "BFloat16Storage"): LazyStorageKind(DT_BF16), - ("torch", "HalfStorage"): LazyStorageKind(DT_F16), - ("torch", "FloatStorage"): LazyStorageKind(DT_F32), - ("torch", "IntStorage"): LazyStorageKind(DT_I32), - ("torch", "Tensor"): LazyTensor, + ('torch._tensor', '_rebuild_from_type_v2'): getattr(rebuild_from_type_v2, '__func__'), + ('torch._utils', '_rebuild_tensor_v2'): getattr(lazy_rebuild_tensor_v2, '__func__'), + ('torch', 'BFloat16Storage'): LazyStorageKind(DT_BF16), + ('torch', 'HalfStorage'): LazyStorageKind(DT_F16), + ('torch', 'FloatStorage'): LazyStorageKind(DT_F32), + ('torch', 'IntStorage'): LazyStorageKind(DT_I32), + ('torch', 'Tensor'): LazyTensor, } def find_class(self, module: str, name: str) -> Any: @@ -968,7 +903,7 @@ def bounded_parallel_map(func: Callable[[In], Out], iterable: Iterable[In], conc executor_class = ProcessPoolExecutor else: executor_class = ThreadPoolExecutor - with executor_class(max_workers = max_workers) as executor: + with executor_class(max_workers=max_workers) as executor: futures: list[concurrent.futures.Future[Out]] = [] done = False for _ in range(concurrency): @@ -1022,12 +957,8 @@ def check_vocab_size(params: Params, vocab: Vocab, pad_vocab: bool = False) -> N class OutputFile: - def __init__( - self, fname_out: Path, endianess: gguf.GGUFEndian = gguf.GGUFEndian.LITTLE - ) -> None: - self.gguf = gguf.GGUFWriter( - fname_out, gguf.MODEL_ARCH_NAMES[ARCH], endianess=endianess - ) + def __init__(self, fname_out: Path, endianess:gguf.GGUFEndian = gguf.GGUFEndian.LITTLE) -> None: + self.gguf = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH], endianess=endianess) def add_meta_arch(self, params: Params) -> None: name = "LLaMA" @@ -1036,21 +967,16 @@ class OutputFile: if params.n_ctx == 4096: name = "LLaMA v2" elif params.path_model is not None: - name = str(params.path_model.parent).split("/")[-1] + name = str(params.path_model.parent).split('/')[-1] - self.gguf.add_name(name) - self.gguf.add_context_length(params.n_ctx) - self.gguf.add_embedding_length(params.n_embd) - self.gguf.add_block_count(params.n_layer) - self.gguf.add_feed_forward_length(params.n_ff) + self.gguf.add_name (name) + self.gguf.add_context_length (params.n_ctx) + self.gguf.add_embedding_length (params.n_embd) + self.gguf.add_block_count (params.n_layer) + self.gguf.add_feed_forward_length (params.n_ff) self.gguf.add_rope_dimension_count(params.n_embd // params.n_head) - self.gguf.add_head_count(params.n_head) - self.gguf.add_head_count_kv(params.n_head_kv) - - if params.f_norm_eps is None: - raise ValueError("f_norm_eps is None") - - self.gguf.add_layer_norm_rms_eps(params.f_norm_eps) + self.gguf.add_head_count (params.n_head) + self.gguf.add_head_count_kv (params.n_head_kv) if params.n_experts: self.gguf.add_expert_count(params.n_experts) @@ -1058,6 +984,11 @@ class OutputFile: if params.n_experts_used: self.gguf.add_expert_used_count(params.n_experts_used) + if params.f_norm_eps: + self.gguf.add_layer_norm_rms_eps(params.f_norm_eps) + else: + raise ValueError('f_norm_eps is None') + if params.f_rope_freq_base is not None: self.gguf.add_rope_freq_base(params.f_rope_freq_base) @@ -1089,7 +1020,7 @@ class OutputFile: return tokenizer_model - def extract_vocabulary_from_model(self, vocab: Vocab) -> Tuple[list, list, list]: + def extract_vocabulary_from_model(self, vocab: Vocab) -> tuple[list[bytes], list[float], list[gguf.TokenType]]: tokens = [] scores = [] toktypes = [] @@ -1124,14 +1055,10 @@ class OutputFile: def add_tensor_info(self, name: str, tensor: LazyTensor) -> None: n_elements = int(np.prod(tensor.shape)) - raw_dtype = getattr(tensor.data_type, "ggml_type", None) - data_type = ( - getattr(tensor.data_type, "quantized_type", None) or tensor.data_type.dtype - ) + raw_dtype = getattr(tensor.data_type, 'ggml_type', None) + data_type = getattr(tensor.data_type, 'quantized_type', None) or tensor.data_type.dtype data_nbytes = tensor.data_type.elements_to_bytes(n_elements) - self.gguf.add_tensor_info( - name, tensor.shape, data_type, data_nbytes, raw_dtype=raw_dtype - ) + self.gguf.add_tensor_info(name, tensor.shape, data_type, data_nbytes, raw_dtype=raw_dtype) def write_meta(self) -> None: self.gguf.write_header_to_file() @@ -1145,14 +1072,10 @@ class OutputFile: @staticmethod def write_vocab_only( - fname_out: Path, - params: Params, - vocab: Vocab, - svocab: gguf.SpecialVocab, - endianess: gguf.GGUFEndian = gguf.GGUFEndian.LITTLE, - pad_vocab: bool = False, + fname_out: Path, params: Params, vocab: Vocab, svocab: gguf.SpecialVocab, + endianess: gguf.GGUFEndian = gguf.GGUFEndian.LITTLE, pad_vocab: bool = False, ) -> None: - check_vocab_size(params, vocab, pad_vocab=pad_vocab) + check_vocab_size(params, vocab, pad_vocab = pad_vocab) of = OutputFile(fname_out, endianess=endianess) @@ -1180,14 +1103,8 @@ class OutputFile: @staticmethod def write_all( - fname_out: Path, - ftype: GGMLFileType, - params: Params, - model: LazyModel, - vocab: Vocab, - svocab: gguf.SpecialVocab, - concurrency: int = DEFAULT_CONCURRENCY, - endianess: gguf.GGUFEndian = gguf.GGUFEndian.LITTLE, + fname_out: Path, ftype: GGMLFileType, params: Params, model: LazyModel, vocab: Vocab, svocab: gguf.SpecialVocab, + concurrency: int = DEFAULT_CONCURRENCY, endianess: gguf.GGUFEndian = gguf.GGUFEndian.LITTLE, pad_vocab: bool = False, ) -> None: check_vocab_size(params, vocab, pad_vocab=pad_vocab) @@ -1207,26 +1124,19 @@ class OutputFile: of.write_tensor_info() # tensor data - ndarrays_inner = bounded_parallel_map( - OutputFile.do_item, model.items(), concurrency=concurrency - ) + ndarrays_inner = bounded_parallel_map(OutputFile.do_item, model.items(), concurrency = concurrency) if ftype == GGMLFileType.MostlyQ8_0: ndarrays = bounded_parallel_map( - OutputFile.maybe_do_quantize, - ndarrays_inner, - concurrency=concurrency, - max_workers=concurrency, + OutputFile.maybe_do_quantize, ndarrays_inner, concurrency=concurrency, max_workers=concurrency, use_processpool_executor=True, ) else: ndarrays = map(OutputFile.maybe_do_quantize, ndarrays_inner) start = time.time() - for i, ((name, lazy_tensor), ndarray) in enumerate( - zip(model.items(), ndarrays) - ): + for i, ((name, lazy_tensor), ndarray) in enumerate(zip(model.items(), ndarrays)): elapsed = time.time() - start - size = " x ".join(f"{dim:6d}" for dim in lazy_tensor.shape) + size = ' x '.join(f"{dim:6d}" for dim in lazy_tensor.shape) padi = len(str(len(model))) print( f"[{i+1:{padi}d}/{len(model)}] Writing tensor {name:38s} | size {size:16} | type {lazy_tensor.data_type.name:4} | T+{int(elapsed):4}" @@ -1363,7 +1273,7 @@ def load_some_model(path: Path) -> ModelPlus: class VocabFactory: def __init__(self, path: Path): self.path = path - self.files = { + self.files: dict[str, Path | None] = { "tokenizer.model": None, "vocab.json": None, "tokenizer.json": None, @@ -1380,24 +1290,18 @@ class VocabFactory: self.files[file] = parent_file_path print(f"Found vocab files: {self.files}") - def _select_file(self, vocabtype: Optional[str]) -> Path: + def _select_file(self, vocabtype: str | None) -> Path: if vocabtype in ["spm", "bpe"]: for file_key in self.files.keys(): - if self.files[file_key]: - return self.files[file_key] + if (file := self.files[file_key]) is not None: + return file raise FileNotFoundError(f"{vocabtype} vocab not found.") - elif vocabtype == "hfft": + if vocabtype == "hfft": # For Hugging Face Fast Tokenizer, return the directory path instead of a specific file return self.path - else: - raise ValueError(f"Unsupported vocabulary type {vocabtype}") + raise ValueError(f"Unsupported vocabulary type {vocabtype}") - def _create_special_vocab( - self, - vocab: Vocab, - vocabtype: str, - model_parent_path: Path, - ) -> gguf.SpecialVocab: + def _create_special_vocab(self, vocab: Vocab, vocabtype: str, model_parent_path: Path) -> gguf.SpecialVocab: load_merges = vocabtype == "bpe" n_vocab = vocab.vocab_size if hasattr(vocab, "vocab_size") else None return gguf.SpecialVocab( @@ -1407,13 +1311,12 @@ class VocabFactory: n_vocab=n_vocab, ) - def load_vocab( - self, vocabtype: str, model_parent_path: Path - ) -> Tuple[Vocab, gguf.SpecialVocab]: + def load_vocab(self, vocabtype: str, model_parent_path: Path) -> tuple[Vocab, gguf.SpecialVocab]: path = self._select_file(vocabtype) print(f"Loading vocab file '{path}', type '{vocabtype}'") added_tokens_path = path.parent / "added_tokens.json" + vocab: Vocab if vocabtype == "bpe": vocab = BpeVocab( path, added_tokens_path if added_tokens_path.exists() else None @@ -1428,6 +1331,7 @@ class VocabFactory: ) else: raise ValueError(f"Unsupported vocabulary type {vocabtype}") + # FIXME: Respect --vocab-dir? special_vocab = self._create_special_vocab( vocab, vocabtype, @@ -1436,18 +1340,17 @@ class VocabFactory: return vocab, special_vocab -def default_output_file(model_paths: list[Path], file_type: GGMLFileType) -> Path: +def default_outfile(model_paths: list[Path], file_type: GGMLFileType) -> Path: namestr = { - GGMLFileType.AllF32: "f32", + GGMLFileType.AllF32: "f32", GGMLFileType.MostlyF16: "f16", - GGMLFileType.MostlyQ8_0: "q8_0", + GGMLFileType.MostlyQ8_0:"q8_0", }[file_type] ret = model_paths[0].parent / f"ggml-model-{namestr}.gguf" if ret in model_paths: sys.stderr.write( f"Error: Default output path ({ret}) would overwrite the input. " - "Please explicitly specify a path using --outfile.\n" - ) + "Please explicitly specify a path using --outfile.\n") sys.exit(1) return ret @@ -1457,111 +1360,34 @@ def do_dump_model(model_plus: ModelPlus) -> None: print(f"model_plus.format = {model_plus.format!r}") print(f"model_plus.vocab = {model_plus.vocab!r}") for name, lazy_tensor in model_plus.model.items(): - print( - f"{name}: shape={lazy_tensor.shape} type={lazy_tensor.data_type}; {lazy_tensor.description}" - ) + print(f"{name}: shape={lazy_tensor.shape} type={lazy_tensor.data_type}; {lazy_tensor.description}") -def get_argument_parser() -> ArgumentParser: +def main(args_in: list[str] | None = None) -> None: output_choices = ["f32", "f16"] if np.uint32(1) == np.uint32(1).newbyteorder("<"): # We currently only support Q8_0 output on little endian systems. output_choices.append("q8_0") + vocab_types = ["spm", "bpe", "hfft"] + parser = argparse.ArgumentParser(description="Convert a LLaMa model to a GGML compatible file") + parser.add_argument("--awq-path", type=Path, help="Path to scale awq cache file", default=None) + parser.add_argument("--dump", action="store_true", help="don't convert, just show what's in the model") + parser.add_argument("--dump-single", action="store_true", help="don't convert, just show what's in a single model file") + parser.add_argument("--vocab-only", action="store_true", help="extract only the vocab") + parser.add_argument("--outtype", choices=output_choices, help="output format - note: q8_0 may be very slow (default: f16 or f32 based on input)") + parser.add_argument("--vocab-dir", type=Path, help="directory containing tokenizer.model, if separate from model file") + parser.add_argument("--vocab-type", choices=vocab_types, help="The vocabulary format used to define the tokenizer model (default: spm)", default="spm") + parser.add_argument("--outfile", type=Path, help="path to write to; default: based on input") + parser.add_argument("model", type=Path, help="directory containing model file, or model file itself (*.pth, *.pt, *.bin)") + parser.add_argument("--ctx", type=int, help="model training context (default: based on input)") + parser.add_argument("--concurrency", type=int, help=f"concurrency used for conversion (default: {DEFAULT_CONCURRENCY})", default=DEFAULT_CONCURRENCY) + parser.add_argument("--big-endian", action="store_true", help="model is executed on big endian machine") + parser.add_argument("--pad-vocab", action="store_true", help="add pad tokens when model vocab expects more than tokenizer metadata provides") - parser = argparse.ArgumentParser( - description="Convert a LLaMa model to a GGML compatible file" - ) - - parser.add_argument( - "model", - type=Path, - help="Directory containing the model file or the model file itself (*.pth, *.pt, *.bin)", - ) - - parser.add_argument( - "--awq-path", - type=Path, - help="Path to the Activation-aware Weight Quantization cache file", - default=None, - ) - - parser.add_argument( - "--dump", - action="store_true", - help="Display the model content without converting it", - ) - - parser.add_argument( - "--dump-single", - action="store_true", - help="Display the content of a single model file without conversion", - ) - - parser.add_argument( - "--vocab-only", - action="store_true", - help="Extract and output only the vocabulary", - ) - - parser.add_argument( - "--outtype", - choices=output_choices, - help="Output format - note: q8_0 may be very slow (default: f16 or f32 based on input)", - ) - - parser.add_argument( - "--vocab-dir", - type=Path, - help="Directory containing the tokenizer.model, if separate from the model file", - ) - - parser.add_argument( - "--vocab-type", - choices=["spm", "bpe", "hfft"], # hfft: Hugging Face Fast Tokenizer - default="spm", - help="The vocabulary format used to define the tokenizer model (default: spm)", - ) - - parser.add_argument( - "--pad-vocab", - action="store_true", - help="Add padding tokens when the model's vocabulary size exceeds the tokenizer metadata", - ) - - parser.add_argument( - "--outfile", - type=Path, - help="Specify the path for the output file (default is based on input)", - ) - - parser.add_argument( - "--ctx", type=int, help="Model training context (default is based on input)" - ) - - parser.add_argument( - "--concurrency", - type=int, - help=f"Concurrency used for conversion (default: {DEFAULT_CONCURRENCY})", - default=DEFAULT_CONCURRENCY, - ) - - parser.add_argument( - "--big-endian", - action="store_true", - help="Indicate that the model is executed on a big-endian machine", - ) - - return parser - - -def main(argv: Optional[list[str]] = None) -> None: - parser = get_argument_parser() - args = parser.parse_args(argv) - + args = parser.parse_args(args_in) if args.awq_path: - sys.path.insert(1, str(Path(__file__).resolve().parent / "awq-py")) - from awq.apply_awq import add_scale_weights - + sys.path.insert(1, str(Path(__file__).parent / 'awq-py')) + from awq.apply_awq import add_scale_weights # type: ignore[import-not-found] tmp_model_path = args.model / "weighted_model" if tmp_model_path.is_dir(): print(f"{tmp_model_path} exists as a weighted model.") @@ -1580,14 +1406,11 @@ def main(argv: Optional[list[str]] = None) -> None: if not args.vocab_only: model_plus = load_some_model(args.model) else: - model_plus = ModelPlus( - model={}, paths=[args.model / "dummy"], format="none", vocab=None - ) + model_plus = ModelPlus(model = {}, paths = [args.model / 'dummy'], format = 'none', vocab = None) if args.dump: do_dump_model(model_plus) return - endianess = gguf.GGUFEndian.LITTLE if args.big_endian: endianess = gguf.GGUFEndian.BIG @@ -1595,12 +1418,10 @@ def main(argv: Optional[list[str]] = None) -> None: params = Params.load(model_plus) if params.n_ctx == -1: if args.ctx is None: - raise Exception( - "The model doesn't have a context size, and you didn't specify one with --ctx\n" - "Please specify one with --ctx:\n" - " - LLaMA v1: --ctx 2048\n" - " - LLaMA v2: --ctx 4096\n" - ) + raise Exception("The model doesn't have a context size, and you didn't specify one with --ctx\n" + "Please specify one with --ctx:\n" + " - LLaMA v1: --ctx 2048\n" + " - LLaMA v2: --ctx 4096\n") params.n_ctx = args.ctx if args.outtype: @@ -1621,42 +1442,30 @@ def main(argv: Optional[list[str]] = None) -> None: if not args.outfile: raise ValueError("need --outfile if using --vocab-only") outfile = args.outfile - OutputFile.write_vocab_only( - outfile, - params, - vocab, - special_vocab, - endianess=endianess, - pad_vocab=args.pad_vocab, - ) + OutputFile.write_vocab_only(outfile, params, vocab, special_vocab, + endianess=endianess, pad_vocab=args.pad_vocab) print(f"Wrote {outfile}") return if model_plus.vocab is not None and args.vocab_dir is None: vocab = model_plus.vocab - model = model_plus.model - model = convert_model_names(model, params) - ftype = pick_output_type(model, args.outtype) - model = convert_to_output_type(model, ftype) - outfile = args.outfile or default_output_file(model_plus.paths, ftype) + print(f"Vocab info: {vocab}") + print(f"Special vocab info: {special_vocab}") + + model = model_plus.model + model = convert_model_names(model, params) + ftype = pick_output_type(model, args.outtype) + model = convert_to_output_type(model, ftype) + outfile = args.outfile or default_outfile(model_plus.paths, ftype) params.ftype = ftype print(f"Writing {outfile}, format {ftype}") - OutputFile.write_all( - outfile, - ftype, - params, - model, - vocab, - special_vocab, - concurrency=args.concurrency, - endianess=endianess, - pad_vocab=args.pad_vocab, - ) + OutputFile.write_all(outfile, ftype, params, model, vocab, special_vocab, + concurrency=args.concurrency, endianess=endianess, pad_vocab=args.pad_vocab) print(f"Wrote {outfile}") -if __name__ == "__main__": - main(sys.argv[1:]) # Exclude the first element (script name) from sys.argv +if __name__ == '__main__': + main() diff --git a/mypy.ini b/mypy.ini index 7215a05dd..e51910ca7 100644 --- a/mypy.ini +++ b/mypy.ini @@ -4,3 +4,4 @@ allow_untyped_calls = true allow_untyped_defs = true allow_incomplete_defs = true disable_error_code = import-untyped +warn_return_any = false From 942c0107a7301434c0a5e7da46bc4cf2393aa556 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Sun, 21 Jan 2024 05:17:27 +0200 Subject: [PATCH 06/66] flake.lock: Update (#5054) MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Flake lock file updates: • Updated input 'nixpkgs': 'github:NixOS/nixpkgs/9b19f5e77dd906cb52dade0b7bd280339d2a1f3d' (2024-01-13) → 'github:NixOS/nixpkgs/bbe7d8f876fbbe7c959c90ba2ae2852220573261' (2024-01-19) Co-authored-by: github-actions[bot] --- flake.lock | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) diff --git a/flake.lock b/flake.lock index cd532ef4f..1b253cb44 100644 --- a/flake.lock +++ b/flake.lock @@ -20,11 +20,11 @@ }, "nixpkgs": { "locked": { - "lastModified": 1705133751, - "narHash": "sha256-rCIsyE80jgiOU78gCWN3A0wE0tR2GI5nH6MlS+HaaSQ=", + "lastModified": 1705677747, + "narHash": "sha256-eyM3okYtMgYDgmYukoUzrmuoY4xl4FUujnsv/P6I/zI=", "owner": "NixOS", "repo": "nixpkgs", - "rev": "9b19f5e77dd906cb52dade0b7bd280339d2a1f3d", + "rev": "bbe7d8f876fbbe7c959c90ba2ae2852220573261", "type": "github" }, "original": { From 726c0fa9a2da976e9c5d5c51e185d9dd453fc9e5 Mon Sep 17 00:00:00 2001 From: Kawrakow <48489457+ikawrakow@users.noreply.github.com> Date: Sun, 21 Jan 2024 08:01:20 +0200 Subject: [PATCH 07/66] Slightly faster imatrix (#5050) * imatrix: speedup by avoiding unnecessary allocations and copies * imatrix: add --no-ppl option to skip PPL calculations altogether --------- Co-authored-by: Iwan Kawrakow --- examples/imatrix/imatrix.cpp | 74 ++++++++++++++++++++++++------------ 1 file changed, 49 insertions(+), 25 deletions(-) diff --git a/examples/imatrix/imatrix.cpp b/examples/imatrix/imatrix.cpp index 5a3d30b88..5687476cd 100644 --- a/examples/imatrix/imatrix.cpp +++ b/examples/imatrix/imatrix.cpp @@ -248,7 +248,7 @@ static void process_logits( } } -static bool compute_imatrix(llama_context * ctx, const gpt_params & params) { +static bool compute_imatrix(llama_context * ctx, const gpt_params & params, bool compute_ppl) { const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx)); const int n_ctx = llama_n_ctx(ctx); @@ -269,10 +269,12 @@ static bool compute_imatrix(llama_context * ctx, const gpt_params & params) { } std::vector logit_history; - logit_history.resize(tokens.size()); - std::vector prob_history; - prob_history.resize(tokens.size()); + + if (compute_ppl) { + logit_history.resize(tokens.size()); + prob_history.resize(tokens.size()); + } const int n_chunk_max = tokens.size() / n_ctx; @@ -288,12 +290,17 @@ static bool compute_imatrix(llama_context * ctx, const gpt_params & params) { std::vector workers(std::thread::hardware_concurrency() - 1); + const int num_batches = (n_ctx + n_batch - 1) / n_batch; + + std::vector logits; + if (compute_ppl && num_batches > 1) { + logits.reserve((size_t)n_ctx * n_vocab); + } + for (int i = 0; i < n_chunk; ++i) { const int start = i * n_ctx; const int end = start + n_ctx; - const int num_batches = (n_ctx + n_batch - 1) / n_batch; - std::vector logits; const auto t_start = std::chrono::high_resolution_clock::now(); @@ -321,8 +328,10 @@ static bool compute_imatrix(llama_context * ctx, const gpt_params & params) { // restore the original token in case it was set to BOS tokens[batch_start] = token_org; - const auto * batch_logits = llama_get_logits(ctx); - logits.insert(logits.end(), batch_logits, batch_logits + batch_size * n_vocab); + if (compute_ppl && num_batches > 1) { + const auto * batch_logits = llama_get_logits(ctx); + logits.insert(logits.end(), batch_logits, batch_logits + batch_size * n_vocab); + } } const auto t_end = std::chrono::high_resolution_clock::now(); @@ -338,25 +347,32 @@ static bool compute_imatrix(llama_context * ctx, const gpt_params & params) { fprintf(stderr, "%.2f minutes\n", total_seconds / 60.0); } - const int first = n_ctx/2; - process_logits(n_vocab, logits.data() + first*n_vocab, tokens.data() + start + first, n_ctx - 1 - first, - workers, nll, nll2, logit_history.data() + start + first, prob_history.data() + start + first); - count += n_ctx - first - 1; + if (compute_ppl) { + const int first = n_ctx/2; + const auto all_logits = num_batches > 1 ? logits.data() : llama_get_logits(ctx); + process_logits(n_vocab, all_logits + first*n_vocab, tokens.data() + start + first, n_ctx - 1 - first, + workers, nll, nll2, logit_history.data() + start + first, prob_history.data() + start + first); + count += n_ctx - first - 1; - printf("[%d]%.4lf,", i + 1, std::exp(nll / count)); - fflush(stdout); + printf("[%d]%.4lf,", i + 1, std::exp(nll / count)); + fflush(stdout); + + logits.clear(); + } } printf("\n"); - nll2 /= count; - nll /= count; - const double ppl = exp(nll); - nll2 -= nll * nll; - if (nll2 > 0) { - nll2 = sqrt(nll2/(count-1)); - printf("Final estimate: PPL = %.4lf +/- %.5lf\n", ppl, nll2*ppl); - } else { - printf("Unexpected negative standard deviation of log(prob)\n"); + if (compute_ppl) { + nll2 /= count; + nll /= count; + const double ppl = exp(nll); + nll2 -= nll * nll; + if (nll2 > 0) { + nll2 = sqrt(nll2/(count-1)); + printf("Final estimate: PPL = %.4lf +/- %.5lf\n", ppl, nll2*ppl); + } else { + printf("Unexpected negative standard deviation of log(prob)\n"); + } } return true; @@ -365,6 +381,7 @@ static bool compute_imatrix(llama_context * ctx, const gpt_params & params) { int main(int argc, char ** argv) { StatParams sparams; + bool compute_ppl = true; std::vector args; args.push_back(argv[0]); int iarg = 1; @@ -381,12 +398,19 @@ int main(int argc, char ** argv) { } else if (arg == "--verbosity") { sparams.verbosity = std::stoi(argv[++iarg]); + } else if (arg == "--no-ppl") { + compute_ppl = false; } else { args.push_back(argv[iarg]); } } if (iarg < argc) { - args.push_back(argv[iarg]); + std::string arg{argv[iarg]}; + if (arg == "--no-ppl") { + compute_ppl = false; + } else { + args.push_back(argv[iarg]); + } } gpt_params params; @@ -448,7 +472,7 @@ int main(int argc, char ** argv) { fprintf(stderr, "%s\n", get_system_info(params).c_str()); } - bool OK = compute_imatrix(ctx, params); + bool OK = compute_imatrix(ctx, params, compute_ppl); if (!OK) { return 1; } From 7dcbe39d36b76389f6c5cd3b151928472b7e22ff Mon Sep 17 00:00:00 2001 From: Kawrakow <48489457+ikawrakow@users.noreply.github.com> Date: Sun, 21 Jan 2024 14:42:44 +0200 Subject: [PATCH 08/66] Add ability to evauate multiple choice tasks (#5047) * TruthfulQA: 1st attempt, does not look like it is working The same implementation can be used for HellaSwag as well, so I converted a HellaSwag validation dataset to the binary format used here and tested with that. The score is only around 50, so something is not quite right. * TruthfulQA: works but the result is bad I know it works because if I convert the HellaSwag validation data to the binary format used in the truthful_qa_score() function I get the exact same result as from the hellaswag_score() function. But I guess, the questions are tricky and the way I have done the combination of question + answer is very likely not the best. The TruthfulQA validation dataset contains 817 questions, with random chance result around 19%. With this version I get 29.1% for Mistral-7B and 55.2% for Mistral-7B-Instruct-v0.2. The HF leader board results for these two models are 42.2% and 68.3%, respectively. * TruthfulQA: fix random sample * TruthfulQA: prepare tasks in parallel for large test datasets * Rename truthful_qa to multiple_choice * Make MSVC happy I had forgotten that MSVC does not make constexpr's available inside a lambda. --------- Co-authored-by: Iwan Kawrakow --- common/common.cpp | 31 +++ common/common.h | 3 + examples/perplexity/perplexity.cpp | 391 ++++++++++++++++++++++++++++- 3 files changed, 422 insertions(+), 3 deletions(-) diff --git a/common/common.cpp b/common/common.cpp index ce20360a4..0e4b8bab2 100644 --- a/common/common.cpp +++ b/common/common.cpp @@ -203,6 +203,25 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) { params.prompt_cache_all = true; } else if (arg == "--prompt-cache-ro") { params.prompt_cache_ro = true; + } else if (arg == "-bf" || arg == "--binary-file") { + if (++i >= argc) { + invalid_param = true; + break; + } + std::ifstream file(argv[i], std::ios::binary); + if (!file) { + fprintf(stderr, "error: failed to open file '%s'\n", argv[i]); + invalid_param = true; + break; + } + // store the external file name in params + params.prompt_file = argv[i]; + file.seekg(0, std::ios::end); + size_t size = file.tellg(); + file.seekg(0, std::ios::beg); + params.prompt.resize(size); + file.read((char *)params.prompt.data(), size); + fprintf(stderr, "Read %zu bytes from binary file %s\n", size, argv[i]); } else if (arg == "-f" || arg == "--file") { if (++i >= argc) { invalid_param = true; @@ -689,6 +708,14 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) { break; } params.winogrande_tasks = std::stoi(argv[i]); + } else if (arg == "--multiple-choice") { + params.multiple_choice = true; + } else if (arg == "--multiple-choice-tasks") { + if (++i >= argc) { + invalid_param = true; + break; + } + params.multiple_choice_tasks = std::stoi(argv[i]); } else if (arg == "--ignore-eos") { params.ignore_eos = true; } else if (arg == "--no-penalize-nl") { @@ -888,6 +915,8 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) { printf(" --in-suffix STRING string to suffix after user inputs with (default: empty)\n"); printf(" -f FNAME, --file FNAME\n"); printf(" prompt file to start generation.\n"); + printf(" -bf FNAME, --binary-file FNAME\n"); + printf(" binary file containing multiple choice tasks.\n"); printf(" -n N, --n-predict N number of tokens to predict (default: %d, -1 = infinity, -2 = until context filled)\n", params.n_predict); printf(" -c N, --ctx-size N size of the prompt context (default: %d, 0 = loaded from model)\n", params.n_ctx); printf(" -b N, --batch-size N batch size for prompt processing (default: %d)\n", params.n_batch); @@ -936,6 +965,8 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) { printf(" --hellaswag-tasks N number of tasks to use when computing the HellaSwag score (default: %zu)\n", params.hellaswag_tasks); printf(" --winogrande compute Winogrande score over random tasks from datafile supplied with -f\n"); printf(" --winogrande-tasks N number of tasks to use when computing the Winogrande score (default: %zu)\n", params.winogrande_tasks); + printf(" --multiple-choice compute multiple choice score over random tasks from datafile supplied with -f\n"); + printf(" --multiple-choice-tasks N number of tasks to use when computing the multiple choice score (default: %zu)\n", params.winogrande_tasks); printf(" --keep N number of tokens to keep from the initial prompt (default: %d, -1 = all)\n", params.n_keep); printf(" --draft N number of tokens to draft for speculative decoding (default: %d)\n", params.n_draft); printf(" --chunks N max number of chunks to process (default: %d, -1 = all)\n", params.n_chunks); diff --git a/common/common.h b/common/common.h index 0ae9c18b3..c69ad7e94 100644 --- a/common/common.h +++ b/common/common.h @@ -108,6 +108,9 @@ struct gpt_params { bool winogrande = false; // compute Winogrande score over random tasks from datafile supplied in prompt size_t winogrande_tasks= 0; // number of tasks to use when computing the Winogrande score. If 0, all tasks will be computed + bool multiple_choice = false; // compute TruthfulQA score over random tasks from datafile supplied in prompt + size_t multiple_choice_tasks = 0; // number of tasks to use when computing the TruthfulQA score. If 0, all tasks will be computed + bool mul_mat_q = true; // if true, use mul_mat_q kernels instead of cuBLAS bool random_prompt = false; // do not randomize prompt if none provided bool use_color = false; // use color to distinguish generations and inputs diff --git a/examples/perplexity/perplexity.cpp b/examples/perplexity/perplexity.cpp index f91f5795a..b7ef9a084 100644 --- a/examples/perplexity/perplexity.cpp +++ b/examples/perplexity/perplexity.cpp @@ -540,14 +540,14 @@ static void hellaswag_score(llama_context * ctx, const gpt_params & params) { // This is needed as usual for LLaMA models const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx)); + // The tasks should be randomized so the score stabilizes quickly. + bool randomize_tasks = true; + // Number of tasks to use when computing the score if (params.hellaswag_tasks < hs_task_count) { hs_task_count = params.hellaswag_tasks; } - // The tasks should be randomized so the score stabilizes quickly. - bool randomize_tasks = true; - // The random seed should not impact the final result if the computation is done over enough tasks, so kept hardcoded for now std::mt19937 rng(1); @@ -1031,6 +1031,389 @@ static void winogrande_score(llama_context * ctx, const gpt_params & params) { printf("Final Winogrande score(%d tasks): %.4lf +/- %.4lf\n", n_done, 100*p, sigma); } +static bool deserialize_string(std::istream& in, std::string& str) { + uint32_t size; + if (!in.read((char *)&size, sizeof(size)).fail()) { + str.resize(size); + if (!in.read((char *)str.data(), size).fail()) return true; + } + return false; +} + +struct multiple_choice_answers { + std::vector answers; + std::vector labels; + bool deserialize(std::istream& in) { + uint32_t n; + in.read((char *)&n, sizeof(n)); + if (in.fail() || n > 100) return false; // 100 as max. number of answers should be good enough for any practical purpose + answers.resize(n); + labels.resize(n); + for (auto& a : answers) { + if (!deserialize_string(in, a)) return false; + } + in.read((char *)labels.data(), n*sizeof(int)); + return !in.fail(); + } +}; + +struct multiple_choice_task { + std::string question; // the question (or context that needs to be continued) + multiple_choice_answers mc1; // possible answers (continuations) with a single correct answer + multiple_choice_answers mc2; // possible answers (continuations) with multiple correct answers - not handled yet + bool deserialize(std::istream& in) { + if (!deserialize_string(in, question)) return false; + return mc1.deserialize(in) && mc2.deserialize(in); + } + + // For evaluation + size_t i_batch; // starting index in the llama_batch + size_t common_prefix; // max number of initial tokens that are the same in all sentences + size_t required_tokens; // needed number of tokens to evaluate all answers + std::vector> seq_tokens; + std::vector log_probs; +}; + +static bool multiple_choice_prepare_one_task(llama_context * ctx, bool add_bos, multiple_choice_task& task, bool log_error) { + if (task.question.empty() || task.mc1.answers.empty()) { + if (log_error) { + printf("%s: found bad task with empty question and/or answers\n", __func__); + } + return false; + } + task.seq_tokens.reserve(task.mc1.answers.size()); + for (auto& answer : task.mc1.answers) { + if (answer.empty()) { + if (log_error) { + printf("%s: found empty answer\n", __func__); + } + return false; + } + task.seq_tokens.emplace_back(::llama_tokenize(ctx, task.question + " " + answer, add_bos)); + } + auto min_len = task.seq_tokens.front().size(); + for (auto& seq : task.seq_tokens) { + min_len = std::min(min_len, seq.size()); + } + task.common_prefix = 0; + for (size_t k = 0; k < min_len; ++k) { + auto token = task.seq_tokens[0][k]; + bool all_same = true; + for (size_t i = 1; i < task.seq_tokens.size(); ++i) { + if (task.seq_tokens[i][k] != token) { + all_same = false; + break; + } + } + if (!all_same) { + break; + } + ++task.common_prefix; + } + task.required_tokens = task.common_prefix; + for (auto& seq : task.seq_tokens) { + task.required_tokens += seq.size() - task.common_prefix; + } + return true; +} + +// +// Calculates score for multiple choice tasks with single correct answer from prompt. +// Commonly used LLM evaluation metrics of this type are +// * ARC +// * HellaSwag +// * MMLU +// * TruthfulQA +// +// Validation datasets for these 4 tests can be found at +// https://huggingface.co/datasets/ikawrakow/validation-datasets-for-llama.cpp +// The data for these datasets was extracted from +// git@hf.co:datasets/allenai/ai2_arc +// https://github.com/rowanz/hellaswag/blob/master/data/hellaswag_val.jsonl +// git@hf.co:datasets/Stevross/mmlu +// https://huggingface.co/datasets/truthful_qa +// +static void multiple_choice_score(llama_context * ctx, const gpt_params & params) { + + std::istringstream strstream(params.prompt); + uint32_t n_task; + strstream.read((char *)&n_task, sizeof(n_task)); + if (strstream.fail() || n_task == 0) { + printf("%s: no tasks\n", __func__); + return; + } + printf("%s: there are %u tasks in prompt\n", __func__, n_task); + std::vector task_pos(n_task); + strstream.read((char *)task_pos.data(), task_pos.size()*sizeof(uint32_t)); + if (strstream.fail()) { + printf("%s: failed to raad task positions from prompt\n", __func__); + return; + } + + std::vector tasks; + if (params.multiple_choice_tasks == 0 || params.multiple_choice_tasks >= (size_t)n_task) { + // Use all tasks + tasks.resize(n_task); + printf("%s: reading tasks", __func__); + int n_dot = n_task/100; + int i = 0; + for (auto& task : tasks) { + ++i; + if (!task.deserialize(strstream)) { + printf("%s: failed to read task %d of %u\n", __func__, i, n_task); + return; + } + if (i%n_dot == 0) printf("."); + } + printf("done\n"); + } + else { + printf("%s: selecting %zu random tasks from %u tasks available\n", __func__, params.multiple_choice_tasks, n_task); + std::mt19937 rng(1); + std::vector aux(n_task); + for (uint32_t i = 0; i < n_task; ++i) aux[i] = i; + float scale = 1.f/(1.f + (float)std::mt19937::max()); + tasks.resize(params.multiple_choice_tasks); + for (auto& task : tasks) { + int j = (int)(scale * rng() * aux.size()); + int idx = aux[j]; + aux[j] = aux.back(); + aux.pop_back(); + strstream.seekg(task_pos[idx], std::ios::beg); + if (!task.deserialize(strstream)) { + printf("%s: failed to read task %d at position %u\n", __func__, idx, task_pos[idx]); + return; + } + } + n_task = params.multiple_choice_tasks; + } + + // This is needed as usual for LLaMA models + const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx)); + + printf("%s: preparing task data", __func__); + fflush(stdout); + if (n_task > 500) { + printf("..."); + fflush(stdout); + std::atomic counter(0); + std::atomic n_bad(0); + auto prepare = [&counter, &n_bad, &tasks, ctx, add_bos] () { + int num_tasks = tasks.size(); + int n_bad_local = 0; + while (true) { + int first = counter.fetch_add(K_TOKEN_CHUNK); + if (first >= num_tasks) { + if (n_bad_local > 0) n_bad += n_bad_local; + break; + } + int last = std::min(first + K_TOKEN_CHUNK, num_tasks); + for (int i = first; i < last; ++i) { + if (!multiple_choice_prepare_one_task(ctx, add_bos, tasks[i], false)) ++n_bad_local; + } + } + }; + size_t max_thread = std::thread::hardware_concurrency(); + max_thread = std::min(max_thread, (tasks.size() + K_TOKEN_CHUNK - 1)/K_TOKEN_CHUNK); + std::vector workers(max_thread-1); + for (auto& w : workers) w = std::thread(prepare); + prepare(); + for (auto& w : workers) w.join(); + printf("done\n"); + fflush(stdout); + int nbad = n_bad; + if (nbad > 0) { + printf("%s: found %d malformed tasks\n", __func__, nbad); + return; + } + } else { + int n_dot = n_task/100; + int i_task = 0; + for (auto& task : tasks) { + ++i_task; + if (!multiple_choice_prepare_one_task(ctx, add_bos, task, true)) { + return; + } + if (i_task%n_dot == 0) { + printf("."); + fflush(stdout); + } + } + printf("done\n"); + } + + printf("%s : calculating TruthfulQA score over %zu tasks.\n", __func__, tasks.size()); + + printf("\ntask\tacc_norm\n"); + + const int n_vocab = llama_n_vocab(llama_get_model(ctx)); + const int n_ctx = llama_n_ctx(ctx); + const int n_batch = params.n_batch; + + const int max_tasks_per_batch = 32; + const int max_seq = 4*max_tasks_per_batch; + + llama_batch batch = llama_batch_init(n_ctx, 0, max_seq); + + std::vector tok_logits(n_vocab); + std::vector batch_logits(n_vocab*n_ctx); + + std::vector> eval_pairs; + std::vector eval_results; + std::vector workers(std::thread::hardware_concurrency()); + std::vector batch_indeces; + + int n_done = 0; + int n_correct = 0; + int n_tot_answers = 0; + + for (size_t i0 = 0; i0 < tasks.size(); i0++) { + int n_cur = 0; + + size_t i1 = i0; + size_t i_batch = 0; // this tells us where in `llama_batch` we are currently + + llama_batch_clear(batch); + + // batch as much tasks as possible into the available context + // each task has 4 unique seuqnce ids - one for each ending + // the common prefix is shared among the 4 sequences to save tokens + // we extract logits only from the last common token and from all ending tokens of each sequence + int s0 = 0; + while (n_cur + (int) tasks[i1].required_tokens <= n_ctx) { + auto& cur_task = tasks[i1]; + + int num_answers = cur_task.seq_tokens.size(); + if (s0 + num_answers > max_seq) { + break; + } + + if (int(batch_indeces.size()) != num_answers) { + batch_indeces.resize(num_answers); + } + for (int s = 0; s < num_answers; ++s) batch_indeces[s] = s0 + s; + + for (size_t i = 0; i < cur_task.common_prefix; ++i) { + //llama_batch_add(batch, cur_task.seq_tokens[0][i], i, { s0 + 0, s0 + 1, s0 + 2, s0 + 3}, false); + llama_batch_add(batch, cur_task.seq_tokens[0][i], i, batch_indeces, false); + } + batch.logits[batch.n_tokens - 1] = true; // we need logits for the last token of the common prefix + + for (int s = 0; s < int(cur_task.seq_tokens.size()); ++s) { + for (size_t i = cur_task.common_prefix; i < cur_task.seq_tokens[s].size(); ++i) { + llama_batch_add(batch, cur_task.seq_tokens[s][i], i, { s0 + s }, true); + } + } + + s0 += num_answers; + + cur_task.i_batch = i_batch; + i_batch += cur_task.required_tokens; + + n_cur += cur_task.required_tokens; + if (++i1 == tasks.size()) { + break; + } + } + + if (i0 == i1) { + fprintf(stderr, "%s : task %zu does not fit in the context window\n", __func__, i0); + return; + } + + llama_kv_cache_clear(ctx); + + // decode all tasks [i0, i1) + if (!decode_helper(ctx, batch, batch_logits, n_batch, n_vocab)) { + fprintf(stderr, "%s: llama_decode() failed\n", __func__); + return; + } + + // Compute log-probs in parallel + // First we collect all tasks + eval_pairs.clear(); + for (size_t i = i0; i < i1; ++i) { + auto& cur_task = tasks[i]; + size_t li = cur_task.common_prefix; + for (int s = 0; s < int(cur_task.seq_tokens.size()); ++s) { + for (size_t j = cur_task.common_prefix; j < cur_task.seq_tokens[s].size() - 1; j++) { + eval_pairs.push_back(std::make_pair(cur_task.i_batch + li++, cur_task.seq_tokens[s][j + 1])); + } + ++li; + } + } + // Then we do the actual calculation + compute_logprobs(batch_logits.data(), n_vocab, workers, eval_pairs, eval_results); + + size_t ir = 0; + + // compute the logprobs for each ending of the decoded tasks + for (size_t i = i0; i < i1; ++i) { + auto & cur_task = tasks[i]; + //printf("==== Evaluating <%s> with correct answer ", cur_task.question.c_str()); + //for (int j = 0; j < int(cur_task.mc1.labels.size()); ++j) { + // if (cur_task.mc1.labels[j] == 1) { + // printf("%d", j+1); + // } + //} + //printf("\n common_prefix: %zu\n", cur_task.common_prefix); + + std::memcpy(tok_logits.data(), batch_logits.data() + n_vocab*(cur_task.i_batch + cur_task.common_prefix - 1), n_vocab*sizeof(float)); + + const auto first_probs = softmax(tok_logits); + + cur_task.log_probs.resize(cur_task.seq_tokens.size()); + for (int s = 0; s < int(cur_task.seq_tokens.size()); ++s) { + size_t count = 1; + float log_prob = std::log(first_probs[cur_task.seq_tokens[s][cur_task.common_prefix]]); + for (size_t j = cur_task.common_prefix; j < cur_task.seq_tokens[s].size() - 1; j++) { + //printf(" %zu %g\n", ir, eval_results[ir]); + ++count; + log_prob += eval_results[ir++]; + } + cur_task.log_probs[s] = log_prob / count; + //printf(" Final: %g\n", log_prob / count); + //printf(" <%s> : %g\n", cur_task.mc1.answers[s].c_str(), log_prob/count); + } + + // Find the ending with maximum logprob + size_t logprob_max_idx = 0; + float logprob_max_val = cur_task.log_probs[0]; + for (size_t s = 1; s < cur_task.log_probs.size(); s++) { + if (cur_task.log_probs[s] > logprob_max_val) { + logprob_max_val = cur_task.log_probs[s]; + logprob_max_idx = s; + } + } + + n_tot_answers += cur_task.log_probs.size(); + if (cur_task.mc1.labels[logprob_max_idx] == 1) { + ++n_correct; + } + ++n_done; + + // Print the accumulated accuracy mean x 100 + printf("%d\t%.8lf\n", n_done, 100.*n_correct/n_done); + fflush(stdout); + } + + i0 = i1 - 1; + } + + llama_batch_free(batch); + + if (n_done < 100) return; + + float p = 1.f*n_correct/n_done; + float sigma = sqrt(p*(1-p)/(n_done-1)); + printf("\n Final result: %.4f +/- %.4f\n", 100.f*p, 100.f*sigma); + p = 1.f*n_done/n_tot_answers; + sigma = sqrt(p*(1-p)/(n_done-1)); + printf("Random chance: %.4f +/- %.4f\n", 100.f*p, 100.f*sigma); + + printf("\n"); +} + int main(int argc, char ** argv) { gpt_params params; @@ -1091,6 +1474,8 @@ int main(int argc, char ** argv) { hellaswag_score(ctx, params); } else if (params.winogrande) { winogrande_score(ctx, params); + } else if (params.multiple_choice) { + multiple_choice_score(ctx, params); } else { results = perplexity(ctx, params); } From 6c5629d4d2d15557c38a0e609b30c1a42abad80d Mon Sep 17 00:00:00 2001 From: bobqianic <129547291+bobqianic@users.noreply.github.com> Date: Sun, 21 Jan 2024 15:17:35 +0000 Subject: [PATCH 09/66] add `#include ` to unicode.h (#5051) Co-authored-by: Jared Van Bortel --- unicode.h | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/unicode.h b/unicode.h index aeca879ea..844eff3da 100644 --- a/unicode.h +++ b/unicode.h @@ -2,8 +2,9 @@ #include #include -#include +#include #include +#include static const std::vector> digit_ranges = { {0x30, 0x39}, {0xB2, 0xB3}, {0xB9, 0xB9}, {0x660, 0x669}, {0x6F0, 0x6F9}, {0x7C0, 0x7C9}, {0x966, 0x96F}, {0x9E6, 0x9EF}, {0xA66, 0xA6F}, {0xAE6, 0xAEF}, {0xB66, 0xB6F}, {0xBE6, 0xBEF}, {0xC66, 0xC6F}, From 05490fad7f7f60ff2bed9ad05cd81b44e82ccde3 Mon Sep 17 00:00:00 2001 From: kuronekosaiko Date: Mon, 22 Jan 2024 00:28:14 +0800 Subject: [PATCH 10/66] add safetensors support to convert-lora-to-ggml.py (#5062) * add safetensors support to convert-lora-to-ggml.py * Update convert-lora-to-ggml.py Remove white space in line 69. --- convert-lora-to-ggml.py | 9 ++++++++- 1 file changed, 8 insertions(+), 1 deletion(-) diff --git a/convert-lora-to-ggml.py b/convert-lora-to-ggml.py index 4904bf128..9a9936dec 100755 --- a/convert-lora-to-ggml.py +++ b/convert-lora-to-ggml.py @@ -59,7 +59,14 @@ if __name__ == '__main__': input_model = os.path.join(sys.argv[1], "adapter_model.bin") output_path = os.path.join(sys.argv[1], "ggml-adapter-model.bin") - model = torch.load(input_model, map_location="cpu") + if os.path.exists(input_model): + model = torch.load(input_model, map_location="cpu") + else: + input_model = os.path.join(sys.argv[1], "adapter_model.safetensors") + # lazy import load_file only if lora is in safetensors format. + from safetensors.torch import load_file + model = load_file(input_model, device="cpu") + arch_name = sys.argv[2] if len(sys.argv) == 3 else "llama" if arch_name not in gguf.MODEL_ARCH_NAMES.values(): From 504dc37be8446fb09b1ede70300250ad41be32a2 Mon Sep 17 00:00:00 2001 From: iSma Date: Sun, 21 Jan 2024 22:37:13 +0100 Subject: [PATCH 11/66] Revert LLAMA_NATIVE to OFF in flake.nix (#5066) --- .devops/nix/package.nix | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/.devops/nix/package.nix b/.devops/nix/package.nix index 43bdbd755..1ef2c6cd9 100644 --- a/.devops/nix/package.nix +++ b/.devops/nix/package.nix @@ -159,7 +159,7 @@ effectiveStdenv.mkDerivation ( cmakeFlags = [ - (cmakeBool "LLAMA_NATIVE" true) + (cmakeBool "LLAMA_NATIVE" false) (cmakeBool "LLAMA_BUILD_SERVER" true) (cmakeBool "BUILD_SHARED_LIBS" true) (cmakeBool "CMAKE_SKIP_BUILD_RPATH" true) From 3466c6ebcf668cac453f891b493ead19283347a8 Mon Sep 17 00:00:00 2001 From: Shijie <821898965@qq.com> Date: Mon, 22 Jan 2024 15:33:19 +0800 Subject: [PATCH 12/66] llama : add more qwen2 models (#5071) --- llama.cpp | 6 ++++-- 1 file changed, 4 insertions(+), 2 deletions(-) diff --git a/llama.cpp b/llama.cpp index 909ad4ad8..9ad74d735 100644 --- a/llama.cpp +++ b/llama.cpp @@ -1325,8 +1325,10 @@ static llama_state g_state; // available llama models enum e_model { MODEL_UNKNOWN, + MODEL_0_5B, MODEL_1B, MODEL_3B, + MODEL_4B, MODEL_7B, MODEL_8B, MODEL_13B, @@ -2892,9 +2894,9 @@ static void llm_load_hparams( { ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); switch (hparams.n_layer) { - case 24: model.type = e_model::MODEL_1B; break; + case 24: model.type = hparams.n_embd == 1024 ? e_model::MODEL_0_5B : e_model::MODEL_1B; break; case 32: model.type = e_model::MODEL_7B; break; - case 40: model.type = e_model::MODEL_13B; break; + case 40: model.type = hparams.n_head == 20 ? e_model::MODEL_4B : e_model::MODEL_13B; break; case 80: model.type = e_model::MODEL_70B; break; default: model.type = e_model::MODEL_UNKNOWN; } From 57744932c64266359ee905518de7e096c0295d8c Mon Sep 17 00:00:00 2001 From: bobqianic <129547291+bobqianic@users.noreply.github.com> Date: Mon, 22 Jan 2024 08:55:05 +0000 Subject: [PATCH 13/66] ci : fix Windows CI by updating Intel SDE version (#5053) --- .github/workflows/build.yml | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/.github/workflows/build.yml b/.github/workflows/build.yml index 367df07a7..c3aa6f992 100644 --- a/.github/workflows/build.yml +++ b/.github/workflows/build.yml @@ -295,7 +295,7 @@ jobs: OPENBLAS_VERSION: 0.3.23 OPENCL_VERSION: 2023.04.17 CLBLAST_VERSION: 1.6.0 - SDE_VERSION: 9.21.1-2023-04-24 + SDE_VERSION: 9.33.0-2024-01-07 strategy: matrix: @@ -400,7 +400,7 @@ jobs: id: cmake_test_sde if: ${{ matrix.build == 'avx512' && env.HAS_AVX512F == '0' }} # use Intel SDE for AVX-512 emulation run: | - curl.exe -o $env:RUNNER_TEMP/sde.tar.xz -L "https://downloadmirror.intel.com/777395/sde-external-${env:SDE_VERSION}-win.tar.xz" + curl.exe -o $env:RUNNER_TEMP/sde.tar.xz -L "https://downloadmirror.intel.com/813591/sde-external-${env:SDE_VERSION}-win.tar.xz" # for some weird reason windows tar doesn't like sde tar.xz 7z x "-o${env:RUNNER_TEMP}" $env:RUNNER_TEMP/sde.tar.xz 7z x "-o${env:RUNNER_TEMP}" $env:RUNNER_TEMP/sde.tar From 66d575c45c5a370d668f9c3283cdf348e2329fa2 Mon Sep 17 00:00:00 2001 From: Kawrakow <48489457+ikawrakow@users.noreply.github.com> Date: Mon, 22 Jan 2024 12:43:33 +0200 Subject: [PATCH 14/66] llama : add Q3_K_XS (#5060) * Add Q3_K_XS - intermediate size between Q2_K and Q3_K_S * Q3_K_XS: quanize first 1/8 of ffn_down layers with Q4_K Together with an importance matrix, this brings perplexity for LLaMA-v2-70B below the perplexity of the former Q2_K with a 800 MB smaller quantized model size. --------- Co-authored-by: Iwan Kawrakow --- examples/quantize/quantize.cpp | 1 + llama.cpp | 62 +++++++++++++++++++++++++--------- llama.h | 1 + 3 files changed, 48 insertions(+), 16 deletions(-) diff --git a/examples/quantize/quantize.cpp b/examples/quantize/quantize.cpp index 2ae046933..f4786157e 100644 --- a/examples/quantize/quantize.cpp +++ b/examples/quantize/quantize.cpp @@ -26,6 +26,7 @@ static const std::vector QUANT_OPTIONS = { { "Q2_K", LLAMA_FTYPE_MOSTLY_Q2_K, " 2.63G, +0.6717 ppl @ LLaMA-v1-7B", }, { "Q2_K_S", LLAMA_FTYPE_MOSTLY_Q2_K_S, " 2.16G, +9.0634 ppl @ LLaMA-v1-7B", }, { "Q3_K", LLAMA_FTYPE_MOSTLY_Q3_K_M, "alias for Q3_K_M" }, + { "Q3_K_XS",LLAMA_FTYPE_MOSTLY_Q3_K_XS,"3-bit extra small quantization" , }, { "Q3_K_S", LLAMA_FTYPE_MOSTLY_Q3_K_S, " 2.75G, +0.5551 ppl @ LLaMA-v1-7B", }, { "Q3_K_M", LLAMA_FTYPE_MOSTLY_Q3_K_M, " 3.07G, +0.2496 ppl @ LLaMA-v1-7B", }, { "Q3_K_L", LLAMA_FTYPE_MOSTLY_Q3_K_L, " 3.35G, +0.1764 ppl @ LLaMA-v1-7B", }, diff --git a/llama.cpp b/llama.cpp index 9ad74d735..c56d31163 100644 --- a/llama.cpp +++ b/llama.cpp @@ -2661,6 +2661,7 @@ static std::string llama_model_ftype_name(llama_ftype ftype) { case LLAMA_FTYPE_MOSTLY_Q6_K: return "Q6_K"; case LLAMA_FTYPE_MOSTLY_IQ2_XXS:return "IQ2_XSS - 2.0625 bpw"; case LLAMA_FTYPE_MOSTLY_IQ2_XS: return "IQ2_XS - 2.3125 bpw"; + case LLAMA_FTYPE_MOSTLY_Q3_K_XS:return "Q3_K - Extra small"; default: return "unknown, may not work"; } @@ -8765,9 +8766,13 @@ struct quantize_state_internal { const llama_model_quantize_params * params; int n_attention_wv = 0; - int n_feed_forward_w2 = 0; + int n_ffn_down = 0; + int n_ffn_gate = 0; + int n_ffn_up = 0; int i_attention_wv = 0; - int i_feed_forward_w2 = 0; + int i_ffn_down = 0; + int i_ffn_gate = 0; + int i_ffn_up = 0; int n_k_quantized = 0; int n_fallback = 0; @@ -8870,8 +8875,8 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty ++qs.i_attention_wv; } else if (name.find("ffn_down") != std::string::npos) { - if (qs.i_feed_forward_w2 < qs.n_feed_forward_w2/8) new_type = GGML_TYPE_Q2_K; - ++qs.i_feed_forward_w2; + if (qs.i_ffn_down < qs.n_ffn_down/8) new_type = GGML_TYPE_Q2_K; + ++qs.i_ffn_down; } else if (name == "token_embd.weight") new_type = GGML_TYPE_Q2_K; } else if (name.find("attn_v.weight") != std::string::npos) { @@ -8908,18 +8913,21 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty // TODO: explore better strategies new_type = GGML_TYPE_Q8_0; } + else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_XS) { + new_type = GGML_TYPE_Q2_K; + } } else if (name.find("ffn_down") != std::string::npos) { const int n_expert = std::max(1, (int)qs.model.hparams.n_expert); int i_layer, n_layer; if (n_expert == 1) { - i_layer = qs.i_feed_forward_w2; - n_layer = qs.n_feed_forward_w2; + i_layer = qs.i_ffn_down; + n_layer = qs.n_ffn_down; } else { // Believe it or not, "experts" in the FFN of Mixtral-8x7B are not consecutive, but iccasionally randomly - // sprinkled in the model. Hence, simply dividing i_feed_forward_w2 by n_expert does not work + // sprinkled in the model. Hence, simply dividing i_ffn_down by n_expert does not work // for getting the current layer as I initially thought, and we need to resort to parsing the // tensor name. - n_layer = qs.n_feed_forward_w2 / n_expert; + n_layer = qs.n_ffn_down / n_expert; if (sscanf(name.c_str(), "blk.%d.ffn_down", &i_layer) != 1) { throw std::runtime_error(format("Failed to determine layer for tensor %s", name.c_str())); } @@ -8928,7 +8936,7 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty } } if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q3_K; - else if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K_S) { + else if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K_S || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_XS) { if (i_layer < n_layer/8) new_type = GGML_TYPE_Q4_K; } else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M) { @@ -8958,11 +8966,12 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty // same quantization as before imatrix stuff, and b) Q4_1/Q5_1 do go crazy on ffn_down without an imatrix. new_type = ftype == LLAMA_FTYPE_MOSTLY_Q4_0 ? GGML_TYPE_Q4_1 : GGML_TYPE_Q5_1; } - ++qs.i_feed_forward_w2; + ++qs.i_ffn_down; } else if (name.find("attn_output.weight") != std::string::npos) { if (arch != LLM_ARCH_FALCON) { if (qs.model.hparams.n_expert == 8) { - if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_S || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M || + if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_XS || + ftype == LLAMA_FTYPE_MOSTLY_Q3_K_S || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S || ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M) { new_type = GGML_TYPE_Q5_K; } @@ -8980,6 +8989,20 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M) new_type = GGML_TYPE_Q5_K; else if (ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M) new_type = GGML_TYPE_Q6_K; } + else if (name.find("ffn_gate") != std::string::npos) { + if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_XS && !use_more_bits(qs.i_ffn_gate, qs.n_ffn_gate)) { + new_type = GGML_TYPE_Q2_K; + } + ++qs.i_ffn_gate; + } + else if (name.find("ffn_up") != std::string::npos) { + if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_XS && !use_more_bits(qs.i_ffn_up, qs.n_ffn_up)) { + new_type = GGML_TYPE_Q2_K; + } + ++qs.i_ffn_up; + } + // if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q3_K; + //} // IK: let's remove this, else Q2_K is almost the same as Q3_K_S //else if (name.find("ffn_gate") != std::string::npos || name.find("ffn_up") != std::string::npos) { // if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q3_K; @@ -9034,8 +9057,9 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s case LLAMA_FTYPE_ALL_F32: quantized_type = GGML_TYPE_F32; break; // K-quants + case LLAMA_FTYPE_MOSTLY_Q2_K_S: case LLAMA_FTYPE_MOSTLY_Q2_K: quantized_type = GGML_TYPE_Q2_K; break; - case LLAMA_FTYPE_MOSTLY_Q2_K_S: quantized_type = GGML_TYPE_Q2_K; break; + case LLAMA_FTYPE_MOSTLY_Q3_K_XS: case LLAMA_FTYPE_MOSTLY_Q3_K_S: case LLAMA_FTYPE_MOSTLY_Q3_K_M: case LLAMA_FTYPE_MOSTLY_Q3_K_L: quantized_type = GGML_TYPE_Q3_K; break; @@ -9103,12 +9127,18 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s ++qs.n_attention_wv; } else if (name.find("ffn_down") != std::string::npos) { - ++qs.n_feed_forward_w2; + ++qs.n_ffn_down; + } + else if (name.find("ffn_gate") != std::string::npos) { + ++qs.n_ffn_gate; + } + else if (name.find("ffn_up") != std::string::npos) { + ++qs.n_ffn_up; } } - if (qs.n_attention_wv != qs.n_feed_forward_w2 || (uint32_t)qs.n_attention_wv != model.hparams.n_layer) { - LLAMA_LOG_WARN("%s ============ Strange model: n_attention_wv = %d, n_feed_forward_w2 = %d, hparams.n_layer = %d\n", - __func__, qs.n_attention_wv, qs.n_feed_forward_w2, model.hparams.n_layer); + if (qs.n_attention_wv != qs.n_ffn_down || (uint32_t)qs.n_attention_wv != model.hparams.n_layer) { + LLAMA_LOG_WARN("%s ============ Strange model: n_attention_wv = %d, n_ffn_down = %d, hparams.n_layer = %d\n", + __func__, qs.n_attention_wv, qs.n_ffn_down, model.hparams.n_layer); } size_t total_size_org = 0; diff --git a/llama.h b/llama.h index e268d7a1d..bb6054557 100644 --- a/llama.h +++ b/llama.h @@ -107,6 +107,7 @@ extern "C" { LLAMA_FTYPE_MOSTLY_IQ2_XXS = 19, // except 1d tensors LLAMA_FTYPE_MOSTLY_IQ2_XS = 20, // except 1d tensors LLAMA_FTYPE_MOSTLY_Q2_K_S = 21, // except 1d tensors + LLAMA_FTYPE_MOSTLY_Q3_K_XS = 22, // except 1d tensors LLAMA_FTYPE_GUESSED = 1024, // not specified in the model file }; From 152d9d05e097e35f1cac21262946d57faec7542a Mon Sep 17 00:00:00 2001 From: Daniel Bevenius Date: Mon, 22 Jan 2024 12:11:01 +0100 Subject: [PATCH 15/66] finetune : print sample-start/include-sample-start (#5072) This commit adds `--sample-start` and `--include-sample-start` to the output from the main function in finetune.cpp. The motivation for this is that even though these are set explicitly by the user via the command line, if one forgets to set them then it is useful to have their values printed out. Otherwise it is possible to go through the whole training process before realizing that the values are not what one expected. Signed-off-by: Daniel Bevenius --- examples/finetune/finetune.cpp | 2 ++ 1 file changed, 2 insertions(+) diff --git a/examples/finetune/finetune.cpp b/examples/finetune/finetune.cpp index 11fcbf443..b7e19c5fe 100644 --- a/examples/finetune/finetune.cpp +++ b/examples/finetune/finetune.cpp @@ -1800,6 +1800,8 @@ int main(int argc, char ** argv) { std::vector train_samples_begin; std::vector train_samples_size; printf("%s: tokenize training data from %s\n", __func__, params.common.fn_train_data); + printf("%s: sample-start: %s\n", __func__, params.common.sample_start.c_str()); + printf("%s: include-sample-start: %s\n", __func__, params.common.include_sample_start ? "true" : "false"); tokenize_file(lctx, params.common.fn_train_data, params.common.sample_start, From d6bd4d46ddb6926087c11e0f6633ab1c81da58c3 Mon Sep 17 00:00:00 2001 From: compilade <113953597+compilade@users.noreply.github.com> Date: Mon, 22 Jan 2024 06:21:52 -0500 Subject: [PATCH 16/66] llama : support StableLM 2 1.6B (#5052) * llama : support StableLM 2 1.6B * convert : fix Qwen's set_vocab wrongly naming all special tokens [PAD{id}] * convert : refactor Qwen's set_vocab to use it for StableLM 2 too * nix : add tiktoken to llama-python-extra * convert : use presence of tokenizer.json to determine StableLM tokenizer loader It's a less arbitrary heuristic than the vocab size. --- .devops/nix/package.nix | 1 + convert-hf-to-gguf.py | 106 +++++++++++++++++++++++----------------- llama.cpp | 18 +++++++ 3 files changed, 79 insertions(+), 46 deletions(-) diff --git a/.devops/nix/package.nix b/.devops/nix/package.nix index 1ef2c6cd9..e8534956f 100644 --- a/.devops/nix/package.nix +++ b/.devops/nix/package.nix @@ -73,6 +73,7 @@ let ps: [ ps.numpy ps.sentencepiece + ps.tiktoken ps.torchWithoutCuda ps.transformers ] diff --git a/convert-hf-to-gguf.py b/convert-hf-to-gguf.py index 4d995ef78..7a0a8c3db 100755 --- a/convert-hf-to-gguf.py +++ b/convert-hf-to-gguf.py @@ -289,6 +289,58 @@ class Model: special_vocab = gguf.SpecialVocab(dir_model, load_merges=True) special_vocab.add_to_gguf(self.gguf_writer) + def _set_vocab_qwen(self): + dir_model = self.dir_model + hparams = self.hparams + tokens: list[bytearray] = [] + toktypes: list[int] = [] + + from transformers import AutoTokenizer + tokenizer = AutoTokenizer.from_pretrained(dir_model, trust_remote_code=True) + vocab_size = hparams["vocab_size"] + assert max(tokenizer.get_vocab().values()) < vocab_size + + merges = [] + vocab = {} + mergeable_ranks = tokenizer.mergeable_ranks + for token, rank in mergeable_ranks.items(): + vocab[QwenModel.token_bytes_to_string(token)] = rank + if len(token) == 1: + continue + merged = QwenModel.bpe(mergeable_ranks, token, max_rank=rank) + assert len(merged) == 2 + merges.append(' '.join(map(QwenModel.token_bytes_to_string, merged))) + + # for this kind of tokenizer, added_vocab is not a subset of vocab, so they need to be combined + added_vocab = tokenizer.special_tokens + reverse_vocab = {id_ : encoded_tok for encoded_tok, id_ in (vocab | added_vocab).items()} + + for i in range(vocab_size): + if i not in reverse_vocab: + pad_token = f"[PAD{i}]".encode("utf-8") + tokens.append(bytearray(pad_token)) + toktypes.append(gguf.TokenType.USER_DEFINED) + elif reverse_vocab[i] in added_vocab: + tokens.append(reverse_vocab[i]) + toktypes.append(gguf.TokenType.CONTROL) + else: + tokens.append(reverse_vocab[i]) + toktypes.append(gguf.TokenType.NORMAL) + + self.gguf_writer.add_tokenizer_model("gpt2") + self.gguf_writer.add_token_list(tokens) + self.gguf_writer.add_token_types(toktypes) + + special_vocab = gguf.SpecialVocab(dir_model, load_merges=False) + special_vocab.merges = merges + # only add special tokens when they were not already loaded from config.json + if len(special_vocab.special_token_ids) == 0: + special_vocab._set_special_token("bos", tokenizer.special_tokens["<|endoftext|>"]) + special_vocab._set_special_token("eos", tokenizer.special_tokens["<|endoftext|>"]) + # this one is usually not in config.json anyway + special_vocab._set_special_token("unk", tokenizer.special_tokens["<|endoftext|>"]) + special_vocab.add_to_gguf(self.gguf_writer) + def _set_vocab_sentencepiece(self): from sentencepiece import SentencePieceProcessor @@ -877,6 +929,13 @@ class PersimmonModel(Model): class StableLMModel(Model): + def set_vocab(self): + if (self.dir_model / "tokenizer.json").is_file(): + self._set_vocab_gpt2() + else: + # StableLM 2 1.6B uses a vocab in a similar format to Qwen's vocab + self._set_vocab_qwen() + def set_gguf_parameters(self): hparams = self.hparams block_count = hparams["num_hidden_layers"] @@ -922,52 +981,7 @@ class QwenModel(Model): return parts def set_vocab(self): - dir_model = self.dir_model - hparams = self.hparams - tokens: list[bytearray] = [] - toktypes: list[int] = [] - - from transformers import AutoTokenizer - tokenizer = AutoTokenizer.from_pretrained(dir_model, trust_remote_code=True) - vocab_size = hparams["vocab_size"] - assert max(tokenizer.get_vocab().values()) < vocab_size - - merges = [] - vocab = {} - mergeable_ranks = tokenizer.mergeable_ranks - for token, rank in mergeable_ranks.items(): - vocab[self.token_bytes_to_string(token)] = rank - if len(token) == 1: - continue - merged = QwenModel.bpe(mergeable_ranks, token, max_rank=rank) - assert len(merged) == 2 - merges.append(' '.join(map(self.token_bytes_to_string, merged))) - - reverse_vocab = {id_ : encoded_tok for encoded_tok, id_ in vocab.items()} - added_vocab = tokenizer.special_tokens - - for i in range(vocab_size): - if i not in reverse_vocab: - pad_token = f"[PAD{i}]".encode("utf-8") - tokens.append(bytearray(pad_token)) - toktypes.append(gguf.TokenType.USER_DEFINED) - elif reverse_vocab[i] in added_vocab: - tokens.append(reverse_vocab[i]) - toktypes.append(gguf.TokenType.CONTROL) - else: - tokens.append(reverse_vocab[i]) - toktypes.append(gguf.TokenType.NORMAL) - - self.gguf_writer.add_tokenizer_model("gpt2") - self.gguf_writer.add_token_list(tokens) - self.gguf_writer.add_token_types(toktypes) - - special_vocab = gguf.SpecialVocab(dir_model, load_merges=False) - special_vocab.merges = merges - special_vocab._set_special_token("bos", tokenizer.special_tokens["<|endoftext|>"]) - special_vocab._set_special_token("eos", tokenizer.special_tokens["<|endoftext|>"]) - special_vocab._set_special_token("unk", tokenizer.special_tokens["<|endoftext|>"]) - special_vocab.add_to_gguf(self.gguf_writer) + self._set_vocab_qwen() def set_gguf_parameters(self): self.gguf_writer.add_name("Qwen") diff --git a/llama.cpp b/llama.cpp index c56d31163..8c906a22f 100644 --- a/llama.cpp +++ b/llama.cpp @@ -2877,6 +2877,7 @@ static void llm_load_hparams( ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps); switch (hparams.n_layer) { + case 24: model.type = e_model::MODEL_1B; break; case 32: model.type = e_model::MODEL_3B; break; default: model.type = e_model::MODEL_UNKNOWN; } @@ -3700,6 +3701,11 @@ static bool llm_load_tensors( layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_gqa}); layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}); + // optional bias tensors, present in Stable LM 2 1.6B + layer.bq = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q, "bias", i), {n_embd}, false); + layer.bk = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K, "bias", i), {n_embd_gqa}, false); + layer.bv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_V, "bias", i), {n_embd_gqa}, false); + layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}); layer.ffn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "bias", i), {n_embd}); @@ -5598,12 +5604,24 @@ struct llm_build_context { // compute Q and K and RoPE them struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].wq, cur); cb(Qcur, "Qcur", il); + if (model.layers[il].bq) { + Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + cb(Qcur, "Qcur", il); + } struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, model.layers[il].wk, cur); cb(Kcur, "Kcur", il); + if (model.layers[il].bk) { + Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + cb(Kcur, "Kcur", il); + } struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv, cur); cb(Vcur, "Vcur", il); + if (model.layers[il].bv) { + Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + cb(Vcur, "Vcur", il); + } Qcur = ggml_rope_custom( ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, From 15bceec2d73d4166340b46b27677c45ac1b4cdad Mon Sep 17 00:00:00 2001 From: Kawrakow <48489457+ikawrakow@users.noreply.github.com> Date: Mon, 22 Jan 2024 14:18:43 +0200 Subject: [PATCH 17/66] imatrix : keep intermediate imatrix results (#5077) Co-authored-by: Iwan Kawrakow --- examples/imatrix/imatrix.cpp | 25 ++++++++++++++++++++++++- 1 file changed, 24 insertions(+), 1 deletion(-) diff --git a/examples/imatrix/imatrix.cpp b/examples/imatrix/imatrix.cpp index 5687476cd..ea06fcdbf 100644 --- a/examples/imatrix/imatrix.cpp +++ b/examples/imatrix/imatrix.cpp @@ -26,6 +26,7 @@ struct StatParams { std::string ofile = "imatrix.dat"; int n_output_frequency = 10; int verbosity = 1; + int keep_every = 0; bool collect_output_weight = false; }; @@ -42,6 +43,9 @@ private: int m_last_call = 0; std::vector m_src1_data; std::vector m_ids; // the expert ids from ggml_mul_mat_id + // + void save_imatrix(const char * file_name) const; + void keep_imatrix(int ncall) const; }; bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void * user_data) { @@ -117,6 +121,9 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void * if (m_last_call % m_params.n_output_frequency == 0) { save_imatrix(); } + if (m_params.keep_every > 0 && m_last_call%m_params.keep_every == 0) { + keep_imatrix(m_last_call); + } } } } else { @@ -143,6 +150,9 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void * if (m_last_call % m_params.n_output_frequency == 0) { save_imatrix(); } + if (m_params.keep_every > 0 && m_last_call%m_params.keep_every == 0) { + keep_imatrix(m_last_call); + } } } @@ -150,7 +160,18 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void * } void IMatrixCollector::save_imatrix() const { - const char * fname = m_params.ofile.empty() ? "imatrix.dat" : m_params.ofile.c_str(); + save_imatrix(m_params.ofile.empty() ? "imatrix.dat" : m_params.ofile.c_str()); +} + +void IMatrixCollector::keep_imatrix(int ncall) const { + auto file_name = m_params.ofile; + if (file_name.empty()) file_name = "imatrix.dat"; + file_name += ".at_"; + file_name += std::to_string(ncall); + save_imatrix(file_name.c_str()); +} + +void IMatrixCollector::save_imatrix(const char * fname) const { std::ofstream out(fname, std::ios::binary); int n_entries = m_stats.size(); out.write((const char*)&n_entries, sizeof(n_entries)); @@ -400,6 +421,8 @@ int main(int argc, char ** argv) { sparams.verbosity = std::stoi(argv[++iarg]); } else if (arg == "--no-ppl") { compute_ppl = false; + } else if (arg == "--keep-imatrix") { + sparams.keep_every = std::stoi(argv[++iarg]); } else { args.push_back(argv[iarg]); } From f7276f7500f7ea588836dd1fc6f126334c517878 Mon Sep 17 00:00:00 2001 From: Someone Serge Date: Sat, 13 Jan 2024 17:10:19 +0000 Subject: [PATCH 18/66] workflows: nix-ci: rebuild on flake.lock updates --- .github/workflows/nix-ci.yml | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/.github/workflows/nix-ci.yml b/.github/workflows/nix-ci.yml index 845b93bfb..4592b9607 100644 --- a/.github/workflows/nix-ci.yml +++ b/.github/workflows/nix-ci.yml @@ -5,10 +5,10 @@ on: push: branches: - master - paths: ['.github/workflows/**', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', '**/*.sh', '**/*.py', '**/*.nix'] + paths: ['.github/workflows/**', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', '**/*.sh', '**/*.py', '**/*.nix', 'flake.lock'] pull_request: types: [opened, synchronize, reopened] - paths: ['**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', '**/*.sh', '**/*.py', '**/*.nix'] + paths: ['**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', '**/*.sh', '**/*.py', '**/*.nix', 'flake.lock'] jobs: nix-eval: From f4dd059259d0234913b9e9780e1662811744c09d Mon Sep 17 00:00:00 2001 From: Someone Serge Date: Sat, 13 Jan 2024 17:16:54 +0000 Subject: [PATCH 19/66] workflows: nix-build-aarch64: rate limit --- .github/workflows/nix-ci-aarch64.yml | 11 +++++++++-- 1 file changed, 9 insertions(+), 2 deletions(-) diff --git a/.github/workflows/nix-ci-aarch64.yml b/.github/workflows/nix-ci-aarch64.yml index be7c26d40..0c6cf5f09 100644 --- a/.github/workflows/nix-ci-aarch64.yml +++ b/.github/workflows/nix-ci-aarch64.yml @@ -2,13 +2,20 @@ name: Nix aarch64 builds on: workflow_dispatch: # allows manual triggering + schedule: + # Rebuild daily rather than on every push because QEMU is expensive (e.g. + # 1.5h instead of minutes with the cold cache). + # + # randint(0, 59), randint(0, 23) + - cron: '26 12 * * *' + # But also rebuild if we touched any of the Nix expressions: push: branches: - master - paths: ['.github/workflows/**', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', '**/*.sh', '**/*.py', '**/*.nix'] + paths: ['**/*.nix', 'flake.lock'] pull_request: types: [opened, synchronize, reopened] - paths: ['**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', '**/*.sh', '**/*.py', '**/*.nix'] + paths: ['**/*.nix', 'flake.lock'] jobs: nix-build-aarch64: From fe8b3c0d4b0d806e8b46660e24eaf4b90b8b385f Mon Sep 17 00:00:00 2001 From: Someone Serge Date: Sat, 13 Jan 2024 17:38:32 +0000 Subject: [PATCH 20/66] workflows: nix-ci: drop the redundant "paths" filter --- .github/workflows/nix-ci.yml | 2 -- 1 file changed, 2 deletions(-) diff --git a/.github/workflows/nix-ci.yml b/.github/workflows/nix-ci.yml index 4592b9607..d19c7a576 100644 --- a/.github/workflows/nix-ci.yml +++ b/.github/workflows/nix-ci.yml @@ -5,10 +5,8 @@ on: push: branches: - master - paths: ['.github/workflows/**', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', '**/*.sh', '**/*.py', '**/*.nix', 'flake.lock'] pull_request: types: [opened, synchronize, reopened] - paths: ['**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', '**/*.sh', '**/*.py', '**/*.nix', 'flake.lock'] jobs: nix-eval: From 7251870780e2d572dd6f239d7a0bfe438c82fa74 Mon Sep 17 00:00:00 2001 From: Someone Serge Date: Sat, 13 Jan 2024 17:45:01 +0000 Subject: [PATCH 21/66] nix: refactor the cleanSource rules --- .devops/nix/package.nix | 16 ++++++++++++---- 1 file changed, 12 insertions(+), 4 deletions(-) diff --git a/.devops/nix/package.nix b/.devops/nix/package.nix index e8534956f..c25d99f01 100644 --- a/.devops/nix/package.nix +++ b/.devops/nix/package.nix @@ -115,14 +115,22 @@ effectiveStdenv.mkDerivation ( pname = "llama-cpp${pnameSuffix}"; version = llamaVersion; + # Note: none of the files discarded here are visible in the sandbox or + # affect the output hash. This also means they can be modified without + # triggering a rebuild. src = lib.cleanSourceWith { filter = name: type: - !(builtins.any (_: _) [ + let + noneOf = builtins.all (x: !x); + baseName = baseNameOf name; + in + noneOf [ (lib.hasSuffix ".nix" name) # Ignore *.nix files when computing outPaths - (name == "README.md") # Ignore *.md changes whe computing outPaths - (lib.hasPrefix "." name) # Skip hidden files and directories - ]); + (lib.hasSuffix ".md" name) # Ignore *.md changes whe computing outPaths + (lib.hasPrefix "." baseName) # Skip hidden files and directories + (baseName == "flake.lock") + ]; src = lib.cleanSource ../../.; }; From 5e97ec91ae3038720a5b15cde4c52d2a53ec2137 Mon Sep 17 00:00:00 2001 From: Someone Serge Date: Sun, 21 Jan 2024 03:15:13 +0000 Subject: [PATCH 22/66] nix: add a comment about makeScope --- .devops/nix/scope.nix | 4 ++++ 1 file changed, 4 insertions(+) diff --git a/.devops/nix/scope.nix b/.devops/nix/scope.nix index 7932ac1e8..d295995a4 100644 --- a/.devops/nix/scope.nix +++ b/.devops/nix/scope.nix @@ -4,6 +4,10 @@ llamaVersion ? "0.0.0", }: +# We're using `makeScope` instead of just writing out an attrset +# because it allows users to apply overlays later using `overrideScope'`. +# Cf. https://noogle.dev/f/lib/makeScope + lib.makeScope newScope ( self: { inherit llamaVersion; From 28603cd2833cedd4434f398d847f87fc83546dbb Mon Sep 17 00:00:00 2001 From: Someone Serge Date: Sun, 21 Jan 2024 03:29:38 +0000 Subject: [PATCH 23/66] nix: add a comment on the many nixpkgs-with-cuda instances --- .devops/nix/nixpkgs-instances.nix | 12 ++++++++++++ 1 file changed, 12 insertions(+) diff --git a/.devops/nix/nixpkgs-instances.nix b/.devops/nix/nixpkgs-instances.nix index 6e9872b28..4a2f81c4b 100644 --- a/.devops/nix/nixpkgs-instances.nix +++ b/.devops/nix/nixpkgs-instances.nix @@ -7,6 +7,18 @@ { system, ... }: { _module.args = { + # Note: bringing up https://zimbatm.com/notes/1000-instances-of-nixpkgs + # again, the below creates several nixpkgs instances which the + # flake-centric CLI will be forced to evaluate e.g. on `nix flake show`. + # + # This is currently "slow" and "expensive", on a certain scale. + # This also isn't "right" in that this hinders dependency injection at + # the level of flake inputs. This might get removed in the foreseeable + # future. + # + # Note that you can use these expressions without Nix + # (`pkgs.callPackage ./devops/nix/scope.nix { }` is the entry point). + pkgsCuda = import inputs.nixpkgs { inherit system; # Ensure dependencies use CUDA consistently (e.g. that openmpi, ucc, From b2d80e105a59b54822edf7ce7f3ed5f317e96e21 Mon Sep 17 00:00:00 2001 From: Someone Serge Date: Sun, 21 Jan 2024 03:41:37 +0000 Subject: [PATCH 24/66] flake.nix: add a comment about flakes vs nix --- flake.nix | 14 ++++++++++++++ 1 file changed, 14 insertions(+) diff --git a/flake.nix b/flake.nix index ec62c773a..a776ba024 100644 --- a/flake.nix +++ b/flake.nix @@ -1,3 +1,17 @@ +# The flake interface to llama.cpp's Nix expressions. The flake is used as a +# more discoverable entry-point, as well as a way to pin the dependencies and +# expose default outputs, including the outputs built by the CI. + +# For more serious applications involving some kind of customization you may +# want to consider consuming the overlay, or instantiating `llamaPackages` +# directly: +# +# ```nix +# pkgs.callPackage ${llama-cpp-root}/.devops/nix/scope.nix { }` +# ``` + +# Cf. https://jade.fyi/blog/flakes-arent-real/ for a more detailed exposition +# of the relation between Nix and the Nix Flakes. { description = "Port of Facebook's LLaMA model in C/C++"; From 3ce7e8f8e7ccfce07e5947ac5f1f3f4628cf68ea Mon Sep 17 00:00:00 2001 From: XiaotaoChen Date: Mon, 22 Jan 2024 21:09:35 +0800 Subject: [PATCH 25/66] llava : MobileVLM support (#4954) * MobileVLM native implementation * delete depthwise_conv_2d and permute_cpy relative code, replace the two by the existed functions, and opt ldp definition, support LLAMA_PERF option for CMake * move android script to example/llava directory * Fix the editor config checks --------- Co-authored-by: Chenxiaotao03 --- CMakeLists.txt | 7 + examples/llava/MobileVLM-README.md | 131 ++++++ examples/llava/android/adb_run.sh | 53 +++ examples/llava/android/build_64.sh | 8 + examples/llava/clip.cpp | 391 +++++++++++++++++- .../llava/convert-image-encoder-to-gguf.py | 6 +- ggml.c | 141 ++++++- ggml.h | 24 ++ 8 files changed, 737 insertions(+), 24 deletions(-) create mode 100644 examples/llava/MobileVLM-README.md create mode 100755 examples/llava/android/adb_run.sh create mode 100755 examples/llava/android/build_64.sh diff --git a/CMakeLists.txt b/CMakeLists.txt index 6b3b1396b..5a333ff52 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -108,6 +108,13 @@ option(LLAMA_BUILD_TESTS "llama: build tests" ${LLAMA_STA option(LLAMA_BUILD_EXAMPLES "llama: build examples" ${LLAMA_STANDALONE}) option(LLAMA_BUILD_SERVER "llama: build server example" ON) + +# add perf arguments +option(LLAMA_PERF "llama: enable perf" OFF) +if (LLAMA_PERF) + add_definitions(-DGGML_PERF) +endif() + # Required for relocatable CMake package include(${CMAKE_CURRENT_SOURCE_DIR}/scripts/build-info.cmake) diff --git a/examples/llava/MobileVLM-README.md b/examples/llava/MobileVLM-README.md new file mode 100644 index 000000000..c6258eba6 --- /dev/null +++ b/examples/llava/MobileVLM-README.md @@ -0,0 +1,131 @@ +# MobileVLM + +Currently this implementation supports [MobileVLM-v1.7](https://huggingface.co/mtgv/MobileVLM-1.7B) variants. + +for more information, please go to [Meituan-AutoML/MobileVLM](https://github.com/Meituan-AutoML/MobileVLM) + +The implementation is based on llava, and is compatible with llava and mobileVLM. The usage is basically same as llava. + +## Usage +Build with cmake or run `make llava-cli` to build it. + +After building, run: `./llava-cli` to see the usage. For example: + +```sh +./llava-cli -m MobileVLM-1.7B/ggml-model-q4_k.gguf \ + --mmproj MobileVLM-1.7B/mmproj-model-f16.gguf \ + --image path/to/an/image.jpg \ + -p "A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. USER: \nWho is the author of this book? Answer the question using a single word or phrase. ASSISTANT:" +``` + +## Model conversion + +- Clone `mobileVLM-1.7B` and `clip-vit-large-patch14-336` locally: + +```sh +git clone https://huggingface.co/mtgv/MobileVLM-1.7B + +git clone https://huggingface.co/openai/clip-vit-large-patch14-336 +``` + +2. Use `llava-surgery.py` to split the LLaVA model to LLaMA and multimodel projector constituents: + +```sh +python ./examples/llava/llava-surgery.py -m path/to/MobileVLM-1.7B +``` + +3. Use `convert-image-encoder-to-gguf.py` with `--projector-type ldp` to convert the LLaVA image encoder to GGUF: + +```sh +python ./examples/llava/convert-image-encoder-to-gguf \ + -m path/to/clip-vit-large-patch14-336 \ + --llava-projector path/to/MobileVLM-1.7B/llava.projector \ + --output-dir path/to/MobileVLM-1.7B \ + --projector-type ldp +``` + +4. Use `convert.py` to convert the LLaMA part of LLaVA to GGUF: + +```sh +python ./convert.py path/to/MobileVLM-1.7B +``` + +5. Use `quantize` to convert LLaMA part's DataType from `fp16` to `q4_k` +```sh +./quantize path/to/MobileVLM-1.7B/ggml-model-f16.gguf path/to/MobileVLM-1.7B/ggml-model-q4_k.gguf q4_k_s +``` + +Now both the LLaMA part and the image encoder is in the `MobileVLM-1.7B` directory. + +## Android compile and run +### compile +refer to `examples/llava/android/build_64.sh` +```sh +mkdir examples/llava/android/build_64 +cd examples/llava/android/build_64 +../build_64.sh +``` +### run on Android +refer to `android/adb_run.sh`, modify resources' `name` and `path` + +## some result on Android with `Snapdragon 888` chip +### case 1 +**input** +```sh +/data/local/tmp/llava-cli \ + -m /data/local/tmp/ggml-model-q4_k.gguf \ + --mmproj /data/local/tmp/mmproj-model-f16.gguf \ + -t 4 \ + --image /data/local/tmp/demo.jpg \ + -p "A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. USER: \nWho is the author of this book? \nAnswer the question using a single word or phrase. ASSISTANT:" +``` +**output** +```sh +encode_image_with_clip: image encoded in 21148.71 ms by CLIP ( 146.87 ms per image patch) + Susan Wise Bauer +llama_print_timings: load time = 23574.72 ms +llama_print_timings: sample time = 1.24 ms / 6 runs ( 0.21 ms per token, 4850.44 tokens per second) +llama_print_timings: prompt eval time = 12460.15 ms / 246 tokens ( 50.65 ms per token, 19.74 tokens per second) +llama_print_timings: eval time = 424.86 ms / 6 runs ( 70.81 ms per token, 14.12 tokens per second) +llama_print_timings: total time = 34731.93 ms +``` +### case 2 +**input** +```sh +/data/local/tmp/llava-cli \ + -m /data/local/tmp/ggml-model-q4_k.gguf \ + --mmproj /data/local/tmp/mmproj-model-f16.gguf \ + -t 4 \ + --image /data/local/tmp/cat.jpeg \ + -p "A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. USER: \nWhat is in the image? ASSISTANT:" +``` + +**output** +```sh +encode_image_with_clip: image encoded in 21149.51 ms by CLIP ( 146.87 ms per image patch) + The image depicts a cat sitting in the grass near some tall green plants. +llama_print_timings: load time = 23257.32 ms +llama_print_timings: sample time = 5.25 ms / 18 runs ( 0.29 ms per token, 3430.53 tokens per second) +llama_print_timings: prompt eval time = 11900.73 ms / 232 tokens ( 51.30 ms per token, 19.49 tokens per second) +llama_print_timings: eval time = 1279.03 ms / 18 runs ( 71.06 ms per token, 14.07 tokens per second) +llama_print_timings: total time = 34570.79 ms +``` + +## Minor shortcomings +The `n_patch` of output in `ldp` is 1/4 of the input. In order to implement quickly, we uniformly modified `clip_n_patches` function to a quarter. when counting the time consumption, the calculated time will be 4 times bigger than the real cost. + +## TODO + +- [ ] Support non-CPU backend for the new operators, such as `depthwise`, `hardswish`, `hardsigmoid` +- [ ] Optimize LDP projector performance + + - Optimize the structure definition to avoid unnecessary memory rearrangements, to reduce the use of `ggml_permute_cpy`; + - Optimize operator implementation (ARM CPU/NVIDIA GPU): such as depthwise conv, hardswish, hardsigmoid, etc. +- [ ] run MobileVLM on `Jetson Orin` +- [ ] Support more model variants, such as `MobileVLM-3B`. + + +## contributor +```sh +zhangjidong05, yangyang260, huyiming03, chenxiaotao03 +``` diff --git a/examples/llava/android/adb_run.sh b/examples/llava/android/adb_run.sh new file mode 100755 index 000000000..f73623ae3 --- /dev/null +++ b/examples/llava/android/adb_run.sh @@ -0,0 +1,53 @@ +#!/bin/bash + +model_dir="/Users/cxt/model/llm/mobileVLM/MobileVLM-1.7B_processed" +projector_name="mmproj-model-f16.gguf" +llama_name="ggml-model-q4_k.gguf" +img_dir="/Users/cxt/model/llm" +img_name="demo.jpg" +prompt="A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. USER: \nWho is the author of this book? \nAnswer the question using a single word or phrase. ASSISTANT:" +# img_name="cat.jpeg" +# prompt="A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. USER: \nWhat is in the image? ASSISTANT:" + +program_dir="build_64/bin" +binName="llava-cli" +n_threads=4 + + +deviceDir="/data/local/tmp" +saveDir="output" +if [ ! -d ${saveDir} ]; then + mkdir ${saveDir} +fi + + +function android_run() { + # # copy resource into device + # adb push ${model_dir}/${projector_name} ${deviceDir}/${projector_name} + # adb push ${model_dir}/${llama_name} ${deviceDir}/${llama_name} + adb push ${img_dir}/${img_name} ${deviceDir}/${img_name} + # copy program into device + adb push ${program_dir}/${binName} ${deviceDir}/${binName} + adb shell "chmod 0777 ${deviceDir}/${binName}" + + # run + adb shell "echo cd ${deviceDir} ${deviceDir}/${binName} \ + -m ${deviceDir}/${llama_name} \ + --mmproj ${deviceDir}/${projector_name} \ + -t ${n_threads} \ + --image ${deviceDir}/${img_name} \ + -p \"${prompt}\" \ + > ${deviceDir}/${modelName}_${projector_name}_${n_threads}_${img_name}.txt" + adb shell "cd ${deviceDir}; pwd; ${deviceDir}/${binName} \ + -m ${deviceDir}/${llama_name} \ + --mmproj ${deviceDir}/${projector_name} \ + -t ${n_threads} \ + --image ${deviceDir}/${img_name} \ + -p \"${prompt}\" \ + >> ${deviceDir}/${modelName}_${projector_name}_${n_threads}_${img_name}.txt 2>&1" + adb pull ${deviceDir}/${modelName}_${projector_name}_${n_threads}_${img_name}.txt ${saveDir} +} + +android_run + +echo "android_run is Done!" diff --git a/examples/llava/android/build_64.sh b/examples/llava/android/build_64.sh new file mode 100755 index 000000000..71b6fd3f7 --- /dev/null +++ b/examples/llava/android/build_64.sh @@ -0,0 +1,8 @@ +#!/bin/bash +cmake ../../../../ \ +-DCMAKE_TOOLCHAIN_FILE=$ANDROID_NDK/build/cmake/android.toolchain.cmake \ +-DCMAKE_BUILD_TYPE=Release \ +-DANDROID_ABI="arm64-v8a" \ +-DANDROID_PLATFORM=android-23 $1 + +make -j4 diff --git a/examples/llava/clip.cpp b/examples/llava/clip.cpp index 2ae8853d3..6161fd858 100644 --- a/examples/llava/clip.cpp +++ b/examples/llava/clip.cpp @@ -12,6 +12,7 @@ #include #include #include +#include #include "clip.h" #include "ggml.h" @@ -67,6 +68,7 @@ static std::string format(const char * fmt, ...) { #define KEY_PATCH_SIZE "clip.vision.patch_size" #define KEY_IMAGE_MEAN "clip.vision.image_mean" #define KEY_IMAGE_STD "clip.vision.image_std" +#define KEY_PROJ_TYPE "clip.projector_type" // // tensor name constants @@ -89,6 +91,21 @@ static std::string format(const char * fmt, ...) { #define TN_TEXT_PROJ "text_projection.weight" #define TN_VIS_PROJ "visual_projection.weight" #define TN_LLAVA_PROJ "mm.%d.%s" +#define TN_MVLM_PROJ_MLP "mm.model.mlp.%d.%s" +#define TN_MVLM_PROJ_BLOCK "mm.model.mb_block.%d.block.%d.%s" + + +enum projector_type { + PROJECTOR_TYPE_MLP, + PROJECTOR_TYPE_LDP, + PROJECTOR_TYPE_UNKNOWN, +}; + +static std::map PROJECTOR_TYPE_NAMES = { + { PROJECTOR_TYPE_MLP, "mlp" }, + { PROJECTOR_TYPE_LDP, "ldp" }, +}; + // // utilities to get data from a gguf file @@ -129,6 +146,91 @@ static std::string get_ftype(int ftype) { return ggml_type_name(static_cast(ftype)); } +static std::string gguf_data_to_str(enum gguf_type type, const void * data, int i) { + switch (type) { + case GGUF_TYPE_UINT8: return std::to_string(((const uint8_t *)data)[i]); + case GGUF_TYPE_INT8: return std::to_string(((const int8_t *)data)[i]); + case GGUF_TYPE_UINT16: return std::to_string(((const uint16_t *)data)[i]); + case GGUF_TYPE_INT16: return std::to_string(((const int16_t *)data)[i]); + case GGUF_TYPE_UINT32: return std::to_string(((const uint32_t *)data)[i]); + case GGUF_TYPE_INT32: return std::to_string(((const int32_t *)data)[i]); + case GGUF_TYPE_UINT64: return std::to_string(((const uint64_t *)data)[i]); + case GGUF_TYPE_INT64: return std::to_string(((const int64_t *)data)[i]); + case GGUF_TYPE_FLOAT32: return std::to_string(((const float *)data)[i]); + case GGUF_TYPE_FLOAT64: return std::to_string(((const double *)data)[i]); + case GGUF_TYPE_BOOL: return ((const bool *)data)[i] ? "true" : "false"; + default: return format("unknown type %d", type); + } +} + + +static void replace_all(std::string & s, const std::string & search, const std::string & replace) { + std::string result; + for (size_t pos = 0; ; pos += search.length()) { + auto new_pos = s.find(search, pos); + if (new_pos == std::string::npos) { + result += s.substr(pos, s.size() - pos); + break; + } + result += s.substr(pos, new_pos - pos) + replace; + pos = new_pos; + } + s = std::move(result); +} + +static std::string gguf_kv_to_str(const struct gguf_context * ctx_gguf, int i) { + const enum gguf_type type = gguf_get_kv_type(ctx_gguf, i); + + switch (type) { + case GGUF_TYPE_STRING: + return gguf_get_val_str(ctx_gguf, i); + case GGUF_TYPE_ARRAY: + { + const enum gguf_type arr_type = gguf_get_arr_type(ctx_gguf, i); + int arr_n = gguf_get_arr_n(ctx_gguf, i); + const void * data = gguf_get_arr_data(ctx_gguf, i); + std::stringstream ss; + ss << "["; + for (int j = 0; j < arr_n; j++) { + if (arr_type == GGUF_TYPE_STRING) { + std::string val = gguf_get_arr_str(ctx_gguf, i, j); + // escape quotes + replace_all(val, "\\", "\\\\"); + replace_all(val, "\"", "\\\""); + ss << '"' << val << '"'; + } else if (arr_type == GGUF_TYPE_ARRAY) { + ss << "???"; + } else { + ss << gguf_data_to_str(arr_type, data, j); + } + if (j < arr_n - 1) { + ss << ", "; + } + } + ss << "]"; + return ss.str(); + } + default: + return gguf_data_to_str(type, gguf_get_val_data(ctx_gguf, i), 0); + } +} + +static void print_tensor_info(const ggml_tensor* tensor, const char* prefix = "") { + size_t tensor_size = ggml_nbytes(tensor); + printf("%s: n_dims = %d, name = %s, tensor_size=%zu, shape:[%d, %d, %d, %d], type: %d\n", + prefix, ggml_n_dims(tensor), tensor->name, tensor_size, + tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->ne[3], tensor->type); +} + +static projector_type clip_projector_type_from_string(const std::string & name) { + for (const auto & kv : PROJECTOR_TYPE_NAMES) { // NOLINT + if (kv.second == name) { + return kv.first; + } + } + return PROJECTOR_TYPE_UNKNOWN; +} + // // image data // @@ -205,6 +307,32 @@ struct clip_vision_model { struct ggml_tensor * mm_0_b; struct ggml_tensor * mm_2_w; struct ggml_tensor * mm_2_b; + + // MobileVLM projection + struct ggml_tensor * mm_model_mlp_1_w; + struct ggml_tensor * mm_model_mlp_1_b; + struct ggml_tensor * mm_model_mlp_3_w; + struct ggml_tensor * mm_model_mlp_3_b; + struct ggml_tensor * mm_model_block_1_block_0_0_w; + struct ggml_tensor * mm_model_block_1_block_0_1_w; + struct ggml_tensor * mm_model_block_1_block_0_1_b; + struct ggml_tensor * mm_model_block_1_block_1_fc1_w; + struct ggml_tensor * mm_model_block_1_block_1_fc1_b; + struct ggml_tensor * mm_model_block_1_block_1_fc2_w; + struct ggml_tensor * mm_model_block_1_block_1_fc2_b; + struct ggml_tensor * mm_model_block_1_block_2_0_w; + struct ggml_tensor * mm_model_block_1_block_2_1_w; + struct ggml_tensor * mm_model_block_1_block_2_1_b; + struct ggml_tensor * mm_model_block_2_block_0_0_w; + struct ggml_tensor * mm_model_block_2_block_0_1_w; + struct ggml_tensor * mm_model_block_2_block_0_1_b; + struct ggml_tensor * mm_model_block_2_block_1_fc1_w; + struct ggml_tensor * mm_model_block_2_block_1_fc1_b; + struct ggml_tensor * mm_model_block_2_block_1_fc2_w; + struct ggml_tensor * mm_model_block_2_block_1_fc2_b; + struct ggml_tensor * mm_model_block_2_block_2_0_w; + struct ggml_tensor * mm_model_block_2_block_2_1_w; + struct ggml_tensor * mm_model_block_2_block_2_1_b; }; struct clip_ctx { @@ -213,6 +341,7 @@ struct clip_ctx { bool has_llava_projector = false; struct clip_vision_model vision_model; + projector_type proj_type = PROJECTOR_TYPE_MLP; float image_mean[3]; float image_std[3]; @@ -430,16 +559,135 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32 free(patches_data); } + // shape [1, 576, 1024] + // ne is whcn, ne = [1024, 576, 1, 1] embeddings = ggml_get_rows(ctx0, embeddings, patches); - // mm projection 0 - embeddings = ggml_mul_mat(ctx0, model.mm_0_w, embeddings); - embeddings = ggml_add(ctx0, embeddings, model.mm_0_b); + // print_tensor_info(embeddings, "embeddings"); - embeddings = ggml_gelu(ctx0, embeddings); + // llava projector + if (ctx->proj_type == PROJECTOR_TYPE_MLP) { + embeddings = ggml_mul_mat(ctx0, model.mm_0_w, embeddings); + embeddings = ggml_add(ctx0, embeddings, model.mm_0_b); - embeddings = ggml_mul_mat(ctx0, model.mm_2_w, embeddings); - embeddings = ggml_add(ctx0, embeddings, model.mm_2_b); + embeddings = ggml_gelu(ctx0, embeddings); + + embeddings = ggml_mul_mat(ctx0, model.mm_2_w, embeddings); + embeddings = ggml_add(ctx0, embeddings, model.mm_2_b); + } + else if (ctx->proj_type == PROJECTOR_TYPE_LDP) { + // MobileVLM projector + int n_patch = 24; + struct ggml_tensor * mlp_1 = ggml_mul_mat(ctx0, model.mm_model_mlp_1_w, embeddings); + mlp_1 = ggml_add(ctx0, mlp_1, model.mm_model_mlp_1_b); + mlp_1 = ggml_gelu(ctx0, mlp_1); + struct ggml_tensor * mlp_3 = ggml_mul_mat(ctx0, model.mm_model_mlp_3_w, mlp_1); + mlp_3 = ggml_add(ctx0, mlp_3, model.mm_model_mlp_3_b); + // mlp_3 shape = [1, 576, 2048], ne = [2048, 576, 1, 1] + + // block 1 + struct ggml_tensor * block_1 = nullptr; + { + // transpose from [1, 576, 2048] --> [1, 2048, 576] --> [1, 2048, 24, 24] + mlp_3 = ggml_cont(ctx0, ggml_permute(ctx0, mlp_3, 1, 0, 2, 3)); + mlp_3 = ggml_reshape_4d(ctx0, mlp_3, n_patch, n_patch, mlp_3->ne[1], mlp_3->ne[2]); + // stride = 1, padding = 1, bias is nullptr + block_1 = ggml_conv_depthwise_2d(ctx0, model.mm_model_block_1_block_0_0_w, mlp_3, nullptr, 1, 1, 1, 1, 1, 1); + + // layer norm + // // block_1 shape = [1, 2048, 24, 24], ne = [24, 24, 2048, 1] + block_1 = ggml_cont(ctx0, ggml_permute(ctx0, block_1, 1, 2, 0, 3)); + // block_1 shape = [1, 24, 24, 2048], ne = [2048, 24, 24, 1] + block_1 = ggml_norm(ctx0, block_1, eps); + block_1 = ggml_add(ctx0, ggml_mul(ctx0, block_1, model.mm_model_block_1_block_0_1_w), model.mm_model_block_1_block_0_1_b); + block_1 = ggml_cont(ctx0, ggml_permute(ctx0, block_1, 2, 0, 1, 3)); + + // block_1 shape = [1, 2048, 24, 24], ne = [24, 24, 2048, 1] + // hardswish + struct ggml_tensor * block_1_hw = ggml_hardswish(ctx0, block_1); + + block_1 = ggml_pool_2d(ctx0, block_1_hw, GGML_OP_POOL_AVG, block_1_hw->ne[0], block_1_hw->ne[1], block_1_hw->ne[0], block_1_hw->ne[1], 0, 0); + // block_1 shape = [1, 2048, 1, 1], ne = [1, 1, 2048, 1] + // pointwise conv + block_1 = ggml_reshape_2d(ctx0, block_1, block_1->ne[0]*block_1->ne[1]*block_1->ne[2], block_1->ne[3]); + block_1 = ggml_mul_mat(ctx0, model.mm_model_block_1_block_1_fc1_w, block_1); + block_1 = ggml_add(ctx0, block_1, model.mm_model_block_1_block_1_fc1_b); + block_1 = ggml_relu(ctx0, block_1); + block_1 = ggml_mul_mat(ctx0, model.mm_model_block_1_block_1_fc2_w, block_1); + block_1 = ggml_add(ctx0, block_1, model.mm_model_block_1_block_1_fc2_b); + block_1 = ggml_hardsigmoid(ctx0, block_1); + // block_1_hw shape = [1, 2048, 24, 24], ne = [24, 24, 2048, 1], block_1 shape = [1, 2048], ne = [2048, 1, 1, 1] + block_1 = ggml_reshape_4d(ctx0, block_1, 1, 1, block_1->ne[0], block_1->ne[1]); + block_1 = ggml_mul(ctx0, block_1_hw, block_1); + + int w = block_1->ne[0], h = block_1->ne[1]; + block_1 = ggml_reshape_3d(ctx0, block_1, w*h, block_1->ne[2], block_1->ne[3]); + block_1 = ggml_cont(ctx0, ggml_permute(ctx0, block_1, 1, 0, 2, 3)); + + // block_1 shape = [1, 24*24, 2048], ne = [24*24, 2048, 1] + block_1 = ggml_mul_mat(ctx0, model.mm_model_block_1_block_2_0_w, block_1); + block_1 = ggml_reshape_4d(ctx0, block_1, block_1->ne[0], w, h, block_1->ne[3]); + + // block_1 shape = [1, 24, 24, 2048], ne = [2048, 24, 24, 1] + block_1 = ggml_norm(ctx0, block_1, eps); + block_1 = ggml_add(ctx0, ggml_mul(ctx0, block_1, model.mm_model_block_1_block_2_1_w), model.mm_model_block_1_block_2_1_b); + block_1 = ggml_cont(ctx0, ggml_permute(ctx0, block_1, 2, 0, 1, 3)); + // block1 shape = [1, 2048, 24, 24], ne = [24, 24, 2048, 1] + // residual + block_1 = ggml_add(ctx0, mlp_3, block_1); + } + + // block_2 + { + // stride = 2 + block_1 = ggml_conv_depthwise_2d(ctx0, model.mm_model_block_2_block_0_0_w, block_1, nullptr, 2, 2, 1, 1, 1, 1); + + // block_1 shape = [1, 2048, 12, 12], ne = [12, 12, 2048, 1] + // layer norm + block_1 = ggml_cont(ctx0, ggml_permute(ctx0, block_1, 1, 2, 0, 3)); + // block_1 shape = [1, 12, 12, 2048], ne = [2048, 12, 12, 1] + block_1 = ggml_norm(ctx0, block_1, eps); + block_1 = ggml_add(ctx0, ggml_mul(ctx0, block_1, model.mm_model_block_2_block_0_1_w), model.mm_model_block_2_block_0_1_b); + block_1 = ggml_cont(ctx0, ggml_permute(ctx0, block_1, 2, 0, 1, 3)); + // block_1 shape = [1, 2048, 12, 12], ne = [12, 12, 2048, 1] + // hardswish + struct ggml_tensor * block_1_hw = ggml_hardswish(ctx0, block_1); + + // not sure the parameters is right for globalAvgPooling + block_1 = ggml_pool_2d(ctx0, block_1_hw, GGML_OP_POOL_AVG, block_1_hw->ne[0], block_1_hw->ne[1], block_1_hw->ne[0], block_1_hw->ne[1], 0, 0); + // block_1 shape = [1, 2048, 1, 1], ne = [1, 1, 2048, 1] + // pointwise conv + block_1 = ggml_reshape_2d(ctx0, block_1, block_1->ne[0]*block_1->ne[1]*block_1->ne[2], block_1->ne[3]); + block_1 = ggml_mul_mat(ctx0, model.mm_model_block_2_block_1_fc1_w, block_1); + block_1 = ggml_add(ctx0, block_1, model.mm_model_block_2_block_1_fc1_b); + block_1 = ggml_relu(ctx0, block_1); + block_1 = ggml_mul_mat(ctx0, model.mm_model_block_2_block_1_fc2_w, block_1); + block_1 = ggml_add(ctx0, block_1, model.mm_model_block_2_block_1_fc2_b); + block_1 = ggml_hardsigmoid(ctx0, block_1); + + // block_1_hw shape = [1, 2048, 12, 12], ne = [12, 12, 2048, 1], block_1 shape = [1, 2048, 1, 1], ne = [1, 1, 2048, 1] + block_1 = ggml_reshape_4d(ctx0, block_1, 1, 1, block_1->ne[0], block_1->ne[1]); + block_1 = ggml_mul(ctx0, block_1_hw, block_1); + + int w = block_1->ne[0], h = block_1->ne[1]; + block_1 = ggml_reshape_3d(ctx0, block_1, w*h, block_1->ne[2], block_1->ne[3]); + block_1 = ggml_cont(ctx0, ggml_permute(ctx0, block_1, 1, 0, 2, 3)); + // block_1 shape = [1, 24*24, 2048], ne = [24*24, 2048, 1] + block_1 = ggml_mul_mat(ctx0, model.mm_model_block_2_block_2_0_w, block_1); + block_1 = ggml_reshape_4d(ctx0, block_1, block_1->ne[0], w, h, block_1->ne[3]); + + + // block_1 shape = [1, 12, 12, 2048], ne = [2048, 12, 12, 1] + block_1 = ggml_norm(ctx0, block_1, eps); + block_1 = ggml_add(ctx0, ggml_mul(ctx0, block_1, model.mm_model_block_2_block_2_1_w), model.mm_model_block_2_block_2_1_b); + block_1 = ggml_reshape_3d(ctx0, block_1, block_1->ne[0], block_1->ne[1] * block_1->ne[2], block_1->ne[3]); + // block_1 shape = [1, 144, 2048], ne = [2048, 144, 1] + } + embeddings = block_1; + } + else { + GGML_ASSERT(false); + } } // build the graph @@ -485,16 +733,55 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) { printf("\n"); } const int n_tensors = gguf_get_n_tensors(ctx); + // kv - if (verbosity >= 3) { - const int n_kv = gguf_get_n_kv(ctx); + const int n_kv = gguf_get_n_kv(ctx); + printf("%s: loaded meta data with %d key-value pairs and %d tensors from %s\n", + __func__, n_kv, n_tensors, fname); + { + std::map n_type; - for (int i = 0; i < n_kv; ++i) { - const char * key = gguf_get_key(ctx, i); + uint32_t n_type_max = 0; + enum ggml_type type_max = GGML_TYPE_F32; - printf("%s: kv[%d]: key = %s\n", __func__, i, key); + for (int i = 0; i < n_tensors; i++) { + enum ggml_type type = gguf_get_tensor_type(ctx, i); + + n_type[type]++; + + if (n_type_max < n_type[type]) { + n_type_max = n_type[type]; + type_max = type; + } + } + + printf("%s: Dumping metadata keys/values. Note: KV overrides do not apply in this output.\n", __func__); + for (int i = 0; i < n_kv; i++) { + const char * name = gguf_get_key(ctx, i); + const enum gguf_type type = gguf_get_kv_type(ctx, i); + const std::string type_name = + type == GGUF_TYPE_ARRAY + ? format("%s[%s,%d]", gguf_type_name(type), gguf_type_name(gguf_get_arr_type(ctx, i)), gguf_get_arr_n(ctx, i)) + : gguf_type_name(type); + + std::string value = gguf_kv_to_str(ctx, i); + const size_t MAX_VALUE_LEN = 40; + if (value.size() > MAX_VALUE_LEN) { + value = format("%s...", value.substr(0, MAX_VALUE_LEN - 3).c_str()); + } + replace_all(value, "\n", "\\n"); + + printf("%s: - kv %3d: %42s %-16s = %s\n", __func__, i, name, type_name.c_str(), value.c_str()); + } + + // print type counts + for (auto & kv : n_type) { + if (kv.second == 0) { + continue; + } + + printf("%s: - type %4s: %4d tensors\n", __func__, ggml_type_name(kv.first), kv.second); } - printf("\n"); } // data @@ -503,20 +790,35 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) { for (int i = 0; i < n_tensors; ++i) { const char * name = gguf_get_tensor_name(ctx, i); const size_t offset = gguf_get_tensor_offset(ctx, i); + enum ggml_type type = gguf_get_tensor_type(ctx, i); struct ggml_tensor * cur = ggml_get_tensor(meta, name); size_t tensor_size = ggml_nbytes(cur); buffer_size += tensor_size; if (verbosity >= 3) { - printf("%s: tensor[%d]: n_dims = %d, name = %s, tensor_size=%zu, offset=%zu\n", __func__, i, - ggml_n_dims(cur), cur->name, tensor_size, offset); + printf("%s: tensor[%d]: n_dims = %d, name = %s, tensor_size=%zu, offset=%zu, shape:[%d, %d, %d, %d], type: %d\n", __func__, i, + ggml_n_dims(cur), cur->name, tensor_size, offset, cur->ne[0], cur->ne[1], cur->ne[2], cur->ne[3], type); } } } + + buffer_size += n_tensors * 128 /* CLIP PADDING */; clip_ctx * new_clip = new clip_ctx; + // update projector type + { + int idx = gguf_find_key(ctx, KEY_PROJ_TYPE); + if (idx != -1) { + const std::string proj_type = gguf_get_val_str(ctx, idx); + new_clip->proj_type = clip_projector_type_from_string(proj_type); + } + else { + new_clip->proj_type = PROJECTOR_TYPE_MLP; + } + } + #ifdef GGML_USE_CUBLAS new_clip->backend = ggml_backend_cuda_init(0); printf("%s: CLIP using CUDA backend\n", __func__); @@ -661,10 +963,45 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) { vision_model.position_embeddings = get_tensor(new_clip->ctx_data, format(TN_POS_EMBD, "v")); vision_model.pre_ln_w = get_tensor(new_clip->ctx_data, format(TN_LN_PRE, "v", "weight")); vision_model.pre_ln_b = get_tensor(new_clip->ctx_data, format(TN_LN_PRE, "v", "bias")); - vision_model.mm_0_w = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 0, "weight")); - vision_model.mm_0_b = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 0, "bias")); - vision_model.mm_2_w = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 2, "weight")); - vision_model.mm_2_b = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 2, "bias")); + + // LLaVA projection + if (new_clip->proj_type == PROJECTOR_TYPE_MLP) { + vision_model.mm_0_w = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 0, "weight")); + vision_model.mm_0_b = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 0, "bias")); + vision_model.mm_2_w = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 2, "weight")); + vision_model.mm_2_b = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 2, "bias")); + } + else if (new_clip->proj_type == PROJECTOR_TYPE_LDP) { + // MobileVLM projection + vision_model.mm_model_mlp_1_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_MLP, 1, "weight")); + vision_model.mm_model_mlp_1_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_MLP, 1, "bias")); + vision_model.mm_model_mlp_3_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_MLP, 3, "weight")); + vision_model.mm_model_mlp_3_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_MLP, 3, "bias")); + vision_model.mm_model_block_1_block_0_0_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 0, "0.weight")); + vision_model.mm_model_block_1_block_0_1_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 0, "1.weight")); + vision_model.mm_model_block_1_block_0_1_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 0, "1.bias")); + vision_model.mm_model_block_1_block_1_fc1_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 1, "fc1.weight")); + vision_model.mm_model_block_1_block_1_fc1_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 1, "fc1.bias")); + vision_model.mm_model_block_1_block_1_fc2_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 1, "fc2.weight")); + vision_model.mm_model_block_1_block_1_fc2_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 1, "fc2.bias")); + vision_model.mm_model_block_1_block_2_0_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 2, "0.weight")); + vision_model.mm_model_block_1_block_2_1_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 2, "1.weight")); + vision_model.mm_model_block_1_block_2_1_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 2, "1.bias")); + vision_model.mm_model_block_2_block_0_0_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 0, "0.weight")); + vision_model.mm_model_block_2_block_0_1_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 0, "1.weight")); + vision_model.mm_model_block_2_block_0_1_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 0, "1.bias")); + vision_model.mm_model_block_2_block_1_fc1_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 1, "fc1.weight")); + vision_model.mm_model_block_2_block_1_fc1_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 1, "fc1.bias")); + vision_model.mm_model_block_2_block_1_fc2_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 1, "fc2.weight")); + vision_model.mm_model_block_2_block_1_fc2_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 1, "fc2.bias")); + vision_model.mm_model_block_2_block_2_0_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 2, "0.weight")); + vision_model.mm_model_block_2_block_2_1_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 2, "1.weight")); + vision_model.mm_model_block_2_block_2_1_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 2, "1.bias")); + } + else { + std::string proj_type = PROJECTOR_TYPE_NAMES[new_clip->proj_type]; + throw std::runtime_error(format("%s: don't support projector with: %s currently\n", __func__, proj_type.c_str())); + } vision_model.layers.resize(hparams.n_layer); for (int il = 0; il < hparams.n_layer; ++il) { @@ -1100,13 +1437,25 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i } int clip_n_mmproj_embd(const struct clip_ctx * ctx) { - return ctx->vision_model.mm_2_b->ne[0]; + if (ctx->proj_type == PROJECTOR_TYPE_LDP) { + return ctx->vision_model.mm_model_block_1_block_2_1_b->ne[0]; + } + else if (ctx->proj_type == PROJECTOR_TYPE_MLP) { + return ctx->vision_model.mm_2_b->ne[0]; + } + else { + std::string proj_type = PROJECTOR_TYPE_NAMES[ctx->proj_type]; + throw std::runtime_error(format("%s: don't support projector with: %s currently\n", __func__, proj_type.c_str())); + } } int clip_n_patches(const struct clip_ctx * ctx) { auto & params = ctx->vision_model.hparams; - - return (params.image_size / params.patch_size) * (params.image_size / params.patch_size); + int n_patches = (params.image_size / params.patch_size) * (params.image_size / params.patch_size); + if (ctx->proj_type == PROJECTOR_TYPE_LDP) { + n_patches /= 4; + } + return n_patches; } size_t clip_embd_nbytes(const struct clip_ctx * ctx) { diff --git a/examples/llava/convert-image-encoder-to-gguf.py b/examples/llava/convert-image-encoder-to-gguf.py index 03688e0ea..f5a3c9b46 100644 --- a/examples/llava/convert-image-encoder-to-gguf.py +++ b/examples/llava/convert-image-encoder-to-gguf.py @@ -81,6 +81,7 @@ ap.add_argument("--vision-only", action="store_true", required=False, ap.add_argument("--clip_model_is_vision", action="store_true", required=False, help="The clip model is a pure vision model (ShareGPT4V vision extract for example)") ap.add_argument("--llava-projector", help="Path to llava.projector file. If specified, save an image encoder for LLaVA models.") +ap.add_argument("--projector-type", help="Type of projector. Possible values: mlp, ldp", choices=["mlp", "ldp"], default="mlp") ap.add_argument("--image-mean", nargs=3, type=float, required=False, help="Override image mean values") ap.add_argument("--image-std", nargs=3, type=float, required=False, help="Override image std values") ap.add_argument("-o", "--output-dir", help="Directory to save GGUF files. Default is the original model directory", default=None) @@ -174,6 +175,8 @@ elif args.vision_only and not has_llava_projector: fout.add_description("vision-only CLIP model") elif has_llava_projector: fout.add_description("image encoder for LLaVA") + # add projector type + fout.add_string("clip.projector_type", args.projector_type) else: fout.add_description("two-tower CLIP model") @@ -218,7 +221,8 @@ if has_llava_projector: projector = torch.load(args.llava_projector) for name, data in projector.items(): name = get_tensor_name(name) - if data.ndim == 2: + # pw and dw conv ndim==4 + if data.ndim == 2 or data.ndim == 4: data = data.squeeze().numpy().astype(np.float16) else: data = data.squeeze().numpy().astype(np.float32) diff --git a/ggml.c b/ggml.c index cbf2d4bdd..a7a88e382 100644 --- a/ggml.c +++ b/ggml.c @@ -1418,6 +1418,9 @@ inline static void ggml_vec_tanh_f32 (const int n, float * y, const float * x) { inline static void ggml_vec_elu_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? x[i] : expf(x[i])-1; } inline static void ggml_vec_relu_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? x[i] : 0.f; } inline static void ggml_vec_leaky_relu_f32 (const int n, float * y, const float * x, const float ns) { for (int i = 0; i < n; ++i) y[i] = ((x[i] > 0.f) ? x[i] : 0.f) + ns * ((x[i] < 0.0f) ? x[i] : 0.f); } +// TODO: optimize performance +inline static void ggml_vec_hardswish_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = x[i] * fminf(1.0f, fmaxf(0.0f, (x[i] + 3.0f) / 6.0f)); } +inline static void ggml_vec_hardsigmoid_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = fminf(1.0f, fmaxf(0.0f, (x[i] + 3.0f) / 6.0f)); } static const float GELU_COEF_A = 0.044715f; static const float GELU_QUICK_COEF = -1.702f; @@ -1776,9 +1779,11 @@ static const char * GGML_UNARY_OP_NAME[GGML_UNARY_OP_COUNT] = { "GELU", "GELU_QUICK", "SILU", + "HARDSWISH", + "HARDSIGMOID", }; -static_assert(GGML_UNARY_OP_COUNT == 10, "GGML_UNARY_OP_COUNT != 10"); +static_assert(GGML_UNARY_OP_COUNT == 12, "GGML_UNARY_OP_COUNT != 12"); static_assert(sizeof(struct ggml_object)%GGML_MEM_ALIGN == 0, "ggml_object size must be a multiple of GGML_MEM_ALIGN"); @@ -3945,6 +3950,20 @@ struct ggml_tensor * ggml_silu_back( return result; } +// ggml hardswish +struct ggml_tensor * ggml_hardswish( + struct ggml_context * ctx, + struct ggml_tensor * a) { + return ggml_unary(ctx, a, GGML_UNARY_OP_HARDSWISH); +} + +// ggml hardsigmoid +struct ggml_tensor * ggml_hardsigmoid( + struct ggml_context * ctx, + struct ggml_tensor * a) { + return ggml_unary(ctx, a, GGML_UNARY_OP_HARDSIGMOID); +} + // ggml_norm static struct ggml_tensor * ggml_norm_impl( @@ -5344,6 +5363,33 @@ GGML_API struct ggml_tensor * ggml_conv_transpose_1d( return result; } +// ggml_conv_depthwise +struct ggml_tensor * ggml_conv_depthwise_2d( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + struct ggml_tensor * c, + int s0, + int s1, + int p0, + int p1, + int d0, + int d1) { + + struct ggml_tensor * new_a = ggml_reshape_4d(ctx, a, a->ne[0], a->ne[1], 1, a->ne[2] * a->ne[3]); + struct ggml_tensor * im2col = ggml_im2col(ctx, new_a, + ggml_reshape_4d(ctx, b, b->ne[0], b->ne[1], 1, b->ne[2] * b->ne[3]), + s0, s1, p0, p1, d0, d1, true); // [N * IC, OH, OW, KH * KW] + + struct ggml_tensor * result = + ggml_mul_mat(ctx, + ggml_reshape_4d(ctx, new_a, (new_a->ne[0] * new_a->ne[1]), new_a->ne[2], new_a->ne[3], 1), // [OC，1, KH, KW] => [1, OC, 1, KH * KW] + ggml_reshape_4d(ctx, im2col, im2col->ne[0], im2col->ne[2] * im2col->ne[1], b->ne[2], b->ne[3])); // [N * IC, OH, OW, KH * KW] => [N, IC, OH * OW, KH * KW] + + result = ggml_reshape_4d(ctx, result, im2col->ne[1], im2col->ne[2], b->ne[2], b->ne[3]); // [N, OC, OH, OW] + + return result; +} // ggml_conv_2d // im2col: [N, IC, IH, IW] => [N, OH, OW, IC*KH*KW] @@ -9333,6 +9379,87 @@ static void ggml_compute_forward_silu_back( } } + +static void ggml_compute_forward_hardswish_f32( + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + struct ggml_tensor * dst) { + assert(params->ith == 0); + assert(ggml_are_same_shape(src0, dst)); + + if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) { + return; + } + + const int n = ggml_nrows(src0); + const int nc = src0->ne[0]; + + assert(dst->nb[0] == sizeof(float)); + assert(src0->nb[0] == sizeof(float)); + + for (int i = 0; i < n; i++) { + ggml_vec_hardswish_f32(nc, + (float *) ((char *) dst->data + i*( dst->nb[1])), + (float *) ((char *) src0->data + i*(src0->nb[1]))); + } +} +static void ggml_compute_forward_hardswish( + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + struct ggml_tensor * dst) { + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_hardswish_f32(params, src0, dst); + } break; + default: + { + GGML_ASSERT(false); + } break; + } +} + +static void ggml_compute_forward_hardsigmoid_f32( + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + struct ggml_tensor * dst) { + assert(params->ith == 0); + assert(ggml_are_same_shape(src0, dst)); + + if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) { + return; + } + + const int n = ggml_nrows(src0); + const int nc = src0->ne[0]; + + assert(dst->nb[0] == sizeof(float)); + assert(src0->nb[0] == sizeof(float)); + + for (int i = 0; i < n; i++) { + ggml_vec_hardsigmoid_f32(nc, + (float *) ((char *) dst->data + i*( dst->nb[1])), + (float *) ((char *) src0->data + i*(src0->nb[1]))); + } +} + +static void ggml_compute_forward_hardsigmoid( + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + struct ggml_tensor * dst) { + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_hardsigmoid_f32(params, src0, dst); + } break; + default: + { + GGML_ASSERT(false); + } break; + } +} + + // ggml_compute_forward_norm static void ggml_compute_forward_norm_f32( @@ -12349,6 +12476,7 @@ static void ggml_compute_forward_im2col( } } + // ggml_compute_forward_conv_transpose_2d static void ggml_compute_forward_conv_transpose_2d( @@ -13917,6 +14045,14 @@ static void ggml_compute_forward_unary( { ggml_compute_forward_silu(params, src0, dst); } break; + case GGML_UNARY_OP_HARDSWISH: + { + ggml_compute_forward_hardswish(params, src0, dst); + } break; + case GGML_UNARY_OP_HARDSIGMOID: + { + ggml_compute_forward_hardsigmoid(params, src0, dst); + } break; default: { GGML_ASSERT(false); @@ -16330,6 +16466,8 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) { case GGML_UNARY_OP_TANH: case GGML_UNARY_OP_ELU: case GGML_UNARY_OP_RELU: + case GGML_UNARY_OP_HARDSWISH: // to opt for multiple threads + case GGML_UNARY_OP_HARDSIGMOID: // to opt for multiple threads { n_tasks = 1; } break; @@ -16562,7 +16700,6 @@ static thread_ret_t ggml_graph_compute_thread(void * data) { // distribute new work or execute it direct if 1T while (++node_n < cgraph->n_nodes) { GGML_PRINT_DEBUG_5("%s: %d/%d\n", __func__, node_n, cgraph->n_nodes); - struct ggml_tensor * node = cgraph->nodes[node_n]; const int n_tasks = ggml_get_n_tasks(node, n_threads); diff --git a/ggml.h b/ggml.h index de8162b81..dca7bd9ce 100644 --- a/ggml.h +++ b/ggml.h @@ -489,6 +489,8 @@ extern "C" { GGML_UNARY_OP_GELU, GGML_UNARY_OP_GELU_QUICK, GGML_UNARY_OP_SILU, + GGML_UNARY_OP_HARDSWISH, + GGML_UNARY_OP_HARDSIGMOID, GGML_UNARY_OP_COUNT, }; @@ -1032,6 +1034,16 @@ extern "C" { struct ggml_tensor * a, struct ggml_tensor * b); + // hardswish(x) = x * relu6(x + 3) / 6 + GGML_API struct ggml_tensor * ggml_hardswish( + struct ggml_context * ctx, + struct ggml_tensor * a); + + // hardsigmoid(x) = relu6(x + 3) / 6 + GGML_API struct ggml_tensor * ggml_hardsigmoid( + struct ggml_context * ctx, + struct ggml_tensor * a); + // normalize along rows GGML_API struct ggml_tensor * ggml_norm( struct ggml_context * ctx, @@ -1483,6 +1495,18 @@ extern "C" { int d1, bool is_2D); + GGML_API struct ggml_tensor * ggml_conv_depthwise_2d( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + struct ggml_tensor * c, + int s0, + int s1, + int p0, + int p1, + int d0, + int d1); + GGML_API struct ggml_tensor * ggml_conv_1d( struct ggml_context * ctx, struct ggml_tensor * a, From 780e24a22eb595b705cbe8284771e9ceff1c4dd2 Mon Sep 17 00:00:00 2001 From: Reinforce-II Date: Mon, 22 Jan 2024 21:15:08 +0800 Subject: [PATCH 26/66] ggml : parallelize FP32 conversion when using BLAS (#5045) * make GGML_TASK_INIT phase can be run in multithread * multithreaded dequantize in mul_mat when using blas library * minor fixes * update outdated comment * fix coding style * simplify code Co-authored-by: Georgi Gerganov --------- Co-authored-by: Georgi Gerganov --- ggml.c | 198 +++++++++++++++++++++++++++++++++++++++++++-------------- 1 file changed, 150 insertions(+), 48 deletions(-) diff --git a/ggml.c b/ggml.c index a7a88e382..f85045c9c 100644 --- a/ggml.c +++ b/ggml.c @@ -7810,6 +7810,9 @@ static void ggml_compute_forward_acc_f32( bool inplace = (bool) ((int32_t *) dst->op_params)[4]; if (!inplace && (params->type == GGML_TASK_INIT)) { + if (params->ith != 0) { + return; + } // memcpy needs to be synchronized across threads to avoid race conditions. // => do it in INIT phase memcpy( @@ -9952,11 +9955,30 @@ static void ggml_compute_forward_mul_mat( #if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) if (ggml_compute_forward_mul_mat_use_blas(dst)) { - if (params->ith != 0) { - return; - } + const int64_t ne_plane = ne01*ne00; + const int64_t desired_wsize = ne13*ne12*ne_plane*sizeof(float); + UNUSED(desired_wsize); if (params->type == GGML_TASK_INIT) { + if (type != GGML_TYPE_F32) { + assert(params->wsize >= desired_wsize); + // parallelize by src0 rows + for (int64_t i13 = 0; i13 < ne13; i13++) { + for (int64_t i12 = 0; i12 < ne12; i12++) { + // broadcast src0 into src1 across 2nd,3rd dimension + const int64_t i03 = i13/r3; + const int64_t i02 = i12/r2; + + const void * x = (char *) src0->data + i02*nb02 + i03*nb03; + float * const wdata = (float *) params->wdata + i13*ne12*ne_plane + i12*ne_plane; + ggml_to_float_t const to_float = type_traits[type].to_float; + + for (int64_t i01 = ith; i01 < ne01; i01 += nth) { + to_float((const char *) x + i01*nb01, wdata + i01*ne00, ne00); + } + } + } + } return; } @@ -9964,9 +9986,14 @@ static void ggml_compute_forward_mul_mat( return; } + // perform sgemm, parallelization controlled by blas lib + if (ith != 0) { + return; + } + + const int64_t tgemm0 = ggml_perf_time_us(); for (int64_t i13 = 0; i13 < ne13; i13++) { for (int64_t i12 = 0; i12 < ne12; i12++) { - // broadcast src0 into src1 across 2nd,3rd dimension const int64_t i03 = i13/r3; const int64_t i02 = i12/r2; @@ -9975,17 +10002,7 @@ static void ggml_compute_forward_mul_mat( float * d = (float *) ((char *) dst->data + i12*nb2 + i13*nb3); if (type != GGML_TYPE_F32) { - float * const wdata = params->wdata; - ggml_to_float_t const to_float = type_traits[type].to_float; - - size_t id = 0; - for (int64_t i01 = 0; i01 < ne01; ++i01) { - to_float((const char *) x + i01*nb01, wdata + id, ne00); - id += ne00; - } - - assert(id*sizeof(float) <= params->wsize); - x = wdata; + x = (float *) params->wdata + i13*ne12*ne_plane + i12*ne_plane; } cblas_sgemm(CblasRowMajor, CblasNoTrans, CblasTrans, @@ -9995,6 +10012,7 @@ static void ggml_compute_forward_mul_mat( 0.0f, d, ne01); } } + //printf("cblas_sgemm = %.3f ms, %lld flops\n", (ggml_perf_time_us() - tgemm0)/1000.0, ne13*ne12*ne1*ne01*ne10*2); //printf("CBLAS = %f ms, %d x %d x %d x %d\n", (ggml_perf_time_us() - t0)/1000.0, ne0, ne1, ne2, ne3); @@ -10003,6 +10021,9 @@ static void ggml_compute_forward_mul_mat( #endif if (params->type == GGML_TASK_INIT) { + if (ith != 0) { + return; + } if (src1->type != vec_dot_type) { char * wdata = params->wdata; const size_t row_size = ggml_row_size(vec_dot_type, ne10); @@ -10167,6 +10188,9 @@ static void ggml_compute_forward_mul_mat_id( #define MMID_MATRIX_ROW(row_id, i1) matrix_rows[(row_id)*ne11 + (i1)] if (params->type == GGML_TASK_INIT) { + if (ith != 0) { + return; + } char * wdata = params->wdata; if (src1->type != vec_dot_type) { const size_t row_size = ggml_row_size(vec_dot_type, ne10); @@ -10352,6 +10376,9 @@ static void ggml_compute_forward_out_prod_f32( return; } #endif + if (ith != 0) { + return; + } ggml_vec_set_f32(ne0*ne1*ne2*ne3, dst->data, 0); return; } @@ -10535,6 +10562,9 @@ static void ggml_compute_forward_out_prod_q_f32( // TODO: #if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) || defined(GGML_USE_CLBLAST) if (params->type == GGML_TASK_INIT) { + if (ith != 0) { + return; + } ggml_vec_set_f32(ne0*ne1*ne2*ne3, dst->data, 0); return; } @@ -10719,6 +10749,9 @@ static void ggml_compute_forward_set_f32( bool inplace = (bool) ((int32_t *) dst->op_params)[4]; if (!inplace && (params->type == GGML_TASK_INIT)) { + if (params->ith != 0) { + return; + } // memcpy needs to be synchronized across threads to avoid race conditions. // => do it in INIT phase memcpy( @@ -11043,6 +11076,9 @@ static void ggml_compute_forward_get_rows_back_f32_f16( // ggml_compute_forward_dup_same_cont(params, opt0, dst); if (params->type == GGML_TASK_INIT) { + if (params->ith != 0) { + return; + } memset(dst->data, 0, ggml_nbytes(dst)); } @@ -11077,6 +11113,9 @@ static void ggml_compute_forward_get_rows_back_f32( // ggml_compute_forward_dup_same_cont(params, opt0, dst); if (params->type == GGML_TASK_INIT) { + if (params->ith != 0) { + return; + } memset(dst->data, 0, ggml_nbytes(dst)); } @@ -11214,6 +11253,9 @@ static void ggml_compute_forward_diag_mask_f32( GGML_ASSERT(n_past >= 0); if (!inplace && (params->type == GGML_TASK_INIT)) { + if (ith != 0) { + return; + } // memcpy needs to be synchronized across threads to avoid race conditions. // => do it in INIT phase GGML_ASSERT(ggml_nelements(dst) == ggml_nelements(src0)); @@ -12184,6 +12226,9 @@ static void ggml_compute_forward_conv_transpose_1d_f16_f32( GGML_ASSERT(nb10 == sizeof(float)); if (params->type == GGML_TASK_INIT) { + if (ith != 0) { + return; + } memset(params->wdata, 0, params->wsize); // permute kernel data (src0) from (K x Cout x Cin) to (Cin x K x Cout) @@ -12278,6 +12323,9 @@ static void ggml_compute_forward_conv_transpose_1d_f32( GGML_ASSERT(nb10 == sizeof(float)); if (params->type == GGML_TASK_INIT) { + if (ith != 0) { + return; + } memset(params->wdata, 0, params->wsize); // prepare kernel data (src0) from (K x Cout x Cin) to (Cin x K x Cout) @@ -12502,6 +12550,9 @@ static void ggml_compute_forward_conv_transpose_2d( GGML_ASSERT(nb10 == sizeof(float)); if (params->type == GGML_TASK_INIT) { + if (ith != 0) { + return; + } memset(params->wdata, 0, params->wsize); // permute kernel data (src0) from (Kw x Kh x Cout x Cin) to (Cin x Kw x Kh x Cout) @@ -14116,6 +14167,9 @@ static void ggml_compute_forward_add_rel_pos_f32( const bool inplace = (bool) ((int32_t *) dst->op_params)[0]; if (!inplace && params->type == GGML_TASK_INIT) { + if (params->ith != 0) { + return; + } memcpy((char *) dst->data, (char *) src0->data, ggml_nbytes(dst)); return; } @@ -16409,8 +16463,9 @@ struct ggml_compute_state_shared { const int n_threads; // synchronization primitives - atomic_int n_active; // num active threads - atomic_int node_n; // active graph node + atomic_int n_active; // num active threads + atomic_int node_n; // active graph node + atomic_int node_task; // active graph node task phase bool (*abort_callback)(void * data); // abort ggml_graph_compute when true void * abort_callback_data; @@ -16658,6 +16713,34 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) { return n_tasks; } +static void ggml_graph_compute_thread_sync_node(int * node_n, struct ggml_compute_state * state, const bool do_yield) { + // wait for other threads to finish + const int last_node_n = * node_n; + + while (true) { + if (do_yield) { + sched_yield(); + } + + * node_n = atomic_load(&state->shared->node_n); + if (* node_n != last_node_n) break; + } +} + +static void ggml_graph_compute_thread_sync_task(int * task_phase, struct ggml_compute_state * state, const bool do_yield) { + // wait for other threads to finish + const int last_task_phase = * task_phase; + + while (true) { + if (do_yield) { + sched_yield(); + } + + * task_phase = atomic_load(&state->shared->node_task); + if (* task_phase != last_task_phase) break; + } +} + static thread_ret_t ggml_graph_compute_thread(void * data) { struct ggml_compute_state * state = (struct ggml_compute_state *) data; @@ -16668,7 +16751,8 @@ static thread_ret_t ggml_graph_compute_thread(void * data) { set_numa_thread_affinity(state->ith, n_threads); - int node_n = -1; + int node_n = -1; + int task_phase = GGML_TASK_FINALIZE; while (true) { if (cplan->abort_callback && cplan->abort_callback(cplan->abort_callback_data)) { @@ -16708,13 +16792,13 @@ static thread_ret_t ggml_graph_compute_thread(void * data) { params.nth = n_tasks; - /* INIT */ - if (GGML_OP_HAS_INIT[node->op]) { - params.type = GGML_TASK_INIT; - ggml_compute_forward(¶ms, node); - } - if (n_tasks == 1) { + /* INIT */ + if (GGML_OP_HAS_INIT[node->op]) { + params.type = GGML_TASK_INIT; + ggml_compute_forward(¶ms, node); + } + // TODO: maybe push node_n to the atomic but if other threads see n_tasks is 1, // they do something more efficient than spinning (?) params.type = GGML_TASK_COMPUTE; @@ -16735,38 +16819,24 @@ static thread_ret_t ggml_graph_compute_thread(void * data) { } } - atomic_store(&state->shared->n_active, n_threads); - atomic_store(&state->shared->node_n, node_n); + task_phase = GGML_TASK_INIT; + atomic_store(&state->shared->n_active, n_threads); + atomic_store(&state->shared->node_n, node_n); + atomic_store(&state->shared->node_task, task_phase); } else { - // wait for other threads to finish - const int last = node_n; - - const bool do_yield = last < 0 || cgraph->nodes[last]->op == GGML_OP_MUL_MAT; - - while (true) { - // TODO: this sched_yield can have significant impact on the performance - either positive or negative - // depending on the workload and the operating system. - // since it is not clear what is the best approach, it should potentially become user-configurable - // ref: https://github.com/ggerganov/ggml/issues/291 - // UPD: adding the do_yield flag seems to resolve the issue universally - if (do_yield) { - sched_yield(); - } - - node_n = atomic_load(&state->shared->node_n); - if (node_n != last) break; - }; + ggml_graph_compute_thread_sync_node(&node_n, state, false); + ggml_graph_compute_thread_sync_task(&task_phase, state, false); } // check if we should stop if (node_n >= cgraph->n_nodes) break; - /* COMPUTE */ + /* INIT & COMPUTE */ struct ggml_tensor * node = cgraph->nodes[node_n]; const int n_tasks = ggml_get_n_tasks(node, n_threads); struct ggml_compute_params params = { - /*.type =*/ GGML_TASK_COMPUTE, + /*.type =*/ GGML_TASK_INIT, /*.ith =*/ state->ith, /*.nth =*/ n_tasks, /*.wsize =*/ cplan->work_size, @@ -16774,8 +16844,39 @@ static thread_ret_t ggml_graph_compute_thread(void * data) { }; if (state->ith < n_tasks) { + if (GGML_OP_HAS_INIT[node->op]) { + ggml_compute_forward(¶ms, node); + } + } + + if (atomic_fetch_sub(&state->shared->n_active, 1) == 1) { + task_phase = GGML_TASK_COMPUTE; + atomic_store(&state->shared->n_active, n_threads); + atomic_store(&state->shared->node_task, task_phase); + } + else { + // TODO: this sched_yield can have significant impact on the performance - either positive or negative + // depending on the workload and the operating system. + // since it is not clear what is the best approach, it should potentially become user-configurable + // ref: https://github.com/ggerganov/ggml/issues/291 + // UPD: adding the do_yield flag seems to resolve the issue universally + const bool do_yield = node_n < 0 || cgraph->nodes[node_n]->op == GGML_OP_MUL_MAT; + ggml_graph_compute_thread_sync_task(&task_phase, state, do_yield); + } + + if (state->ith < n_tasks) { + params.type = GGML_TASK_COMPUTE; ggml_compute_forward(¶ms, node); } + + if (atomic_fetch_sub(&state->shared->n_active, 1) == 1) { + task_phase = GGML_TASK_FINALIZE; + atomic_store(&state->shared->n_active, n_threads); + atomic_store(&state->shared->node_task, task_phase); + } + else { + ggml_graph_compute_thread_sync_task(&task_phase, state, false); + } } return GGML_EXIT_SUCCESS; @@ -16832,8 +16933,8 @@ struct ggml_cplan ggml_graph_plan(const struct ggml_cgraph * cgraph, int n_threa #if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) if (ggml_compute_forward_mul_mat_use_blas(node)) { if (node->src[0]->type != GGML_TYPE_F32) { - // here we need memory just for single 2D matrix from src0 - cur = ggml_type_size(GGML_TYPE_F32)*(node->src[0]->ne[0]*node->src[0]->ne[1]); + // here we need memory for fully dequantized matrix from src0 + cur = ggml_type_size(GGML_TYPE_F32)*ggml_nelements(node->src[0]); } } else #endif @@ -16987,6 +17088,7 @@ int ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cplan * cplan) { /*.n_threads =*/ n_threads, /*.n_active =*/ n_threads, /*.node_n =*/ -1, + /*.node_task =*/ GGML_TASK_FINALIZE, /*.abort_callback =*/ NULL, /*.abort_callback_data =*/ NULL, }; From 6f9939d119b2d004c264952eb510bd106455531e Mon Sep 17 00:00:00 2001 From: Kawrakow <48489457+ikawrakow@users.noreply.github.com> Date: Mon, 22 Jan 2024 16:10:14 +0200 Subject: [PATCH 27/66] KL-divergence (#5076) * kl-divergence: be able to save all logits to a file * Add ability to compute KL-divergence --------- Co-authored-by: Iwan Kawrakow --- common/common.cpp | 9 + common/common.h | 3 + examples/perplexity/perplexity.cpp | 319 ++++++++++++++++++++++++++++- 3 files changed, 329 insertions(+), 2 deletions(-) diff --git a/common/common.cpp b/common/common.cpp index 0e4b8bab2..0a7096171 100644 --- a/common/common.cpp +++ b/common/common.cpp @@ -672,6 +672,12 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) { if (params.logdir.back() != DIRECTORY_SEPARATOR) { params.logdir += DIRECTORY_SEPARATOR; } + } else if (arg == "--save-all-logits" || arg == "--kl-divergence-base") { + if (++i >= argc) { + invalid_param = true; + break; + } + params.logits_file = argv[i]; } else if (arg == "--perplexity" || arg == "--all-logits") { params.logits_all = true; } else if (arg == "--ppl-stride") { @@ -716,6 +722,8 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) { break; } params.multiple_choice_tasks = std::stoi(argv[i]); + } else if (arg == "--kl-divergence") { + params.kl_divergence = true; } else if (arg == "--ignore-eos") { params.ignore_eos = true; } else if (arg == "--no-penalize-nl") { @@ -967,6 +975,7 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) { printf(" --winogrande-tasks N number of tasks to use when computing the Winogrande score (default: %zu)\n", params.winogrande_tasks); printf(" --multiple-choice compute multiple choice score over random tasks from datafile supplied with -f\n"); printf(" --multiple-choice-tasks N number of tasks to use when computing the multiple choice score (default: %zu)\n", params.winogrande_tasks); + printf(" --kl-divergence computes KL-divergence to logits provided via --kl-divergence-base"); printf(" --keep N number of tokens to keep from the initial prompt (default: %d, -1 = all)\n", params.n_keep); printf(" --draft N number of tokens to draft for speculative decoding (default: %d)\n", params.n_draft); printf(" --chunks N max number of chunks to process (default: %d, -1 = all)\n", params.n_chunks); diff --git a/common/common.h b/common/common.h index c69ad7e94..214a379b5 100644 --- a/common/common.h +++ b/common/common.h @@ -91,6 +91,7 @@ struct gpt_params { std::string input_suffix = ""; // string to suffix user inputs with std::vector antiprompt; // string upon seeing which more user input is prompted std::string logdir = ""; // directory in which to save YAML log files + std::string logits_file = ""; // file for saving *all* logits std::vector kv_overrides; @@ -111,6 +112,8 @@ struct gpt_params { bool multiple_choice = false; // compute TruthfulQA score over random tasks from datafile supplied in prompt size_t multiple_choice_tasks = 0; // number of tasks to use when computing the TruthfulQA score. If 0, all tasks will be computed + bool kl_divergence = false; // compute KL-divergence + bool mul_mat_q = true; // if true, use mul_mat_q kernels instead of cuBLAS bool random_prompt = false; // do not randomize prompt if none provided bool use_color = false; // use color to distinguish generations and inputs diff --git a/examples/perplexity/perplexity.cpp b/examples/perplexity/perplexity.cpp index b7ef9a084..1b7f85f49 100644 --- a/examples/perplexity/perplexity.cpp +++ b/examples/perplexity/perplexity.cpp @@ -112,6 +112,43 @@ static results_log_softmax log_softmax(int n_vocab, const float * logits, int to return {logits[tok] - max_logit - log(sum_exp), logits[tok], expf(logits[tok] - max_logit) / (float) sum_exp}; } +static inline int nearest_int(float fval) { + //assert(fval <= 4194303.f); + float val = fval + 12582912.f; + int i; memcpy(&i, &val, sizeof(int)); + return (i & 0x007fffff) - 0x00400000; +} + +static double log_softmax(int n_vocab, const float * logits, uint16_t * log_prob, int tok) { + float max_logit = logits[0]; + float min_logit = logits[0]; + for (int i = 1; i < n_vocab; ++i) { + max_logit = std::max(max_logit, logits[i]); + min_logit = std::min(min_logit, logits[i]); + } + min_logit = std::max(min_logit, max_logit - 16); + double sum_exp = 0.0; + for (int i = 0; i < n_vocab; ++i) { + sum_exp += expf(logits[i] - max_logit); + } + const float log_sum_exp = log(sum_exp); + const float min_log_prob = min_logit - max_logit - log_sum_exp; + const float scale = (max_logit - min_logit)/65535.f; + float * d = (float *)log_prob; + d[0] = scale; + d[1] = min_log_prob; + log_prob += 4; + if (scale) { + const float inv_scale = 1/scale; + for (int i = 0; i < n_vocab; ++i) { + log_prob[i] = logits[i] > min_logit ? nearest_int(inv_scale*(logits[i] - min_logit)) : 0; + } + } else { + std::memset(log_prob, 0, n_vocab*sizeof(uint16_t)); + } + return max_logit + log_sum_exp - logits[tok]; +} + static void process_logits( int n_vocab, const float * logits, const int * tokens, int n_token, std::vector & workers, double & nll, double & nll2, float * logit_history, float * prob_history @@ -147,6 +184,114 @@ static void process_logits( } } +static void process_logits(std::ostream& out, int n_vocab, const float * logits, const int * tokens, int n_token, + std::vector & workers, std::vector & log_probs, double & nll, double & nll2) { + std::mutex mutex; + const int nv = 2*((n_vocab + 1)/2) + 4; + int counter = 0; + auto compute = [&mutex, &counter, &log_probs, &nll, &nll2, n_vocab, logits, tokens, n_token, nv] () { + double local_nll = 0; + double local_nll2 = 0; + while (true) { + std::unique_lock lock(mutex); + int i = counter++; + if (i >= n_token) { + nll += local_nll; nll2 += local_nll2; + break; + } + lock.unlock(); + const double v = log_softmax(n_vocab, logits + i*n_vocab, log_probs.data() + i*nv, tokens[i+1]); + local_nll += v; + local_nll2 += v*v; + } + }; + for (auto & w : workers) { + w = std::thread(compute); + } + compute(); + for (auto & w : workers) { + w.join(); + } + out.write((const char *)log_probs.data(), n_token*nv*sizeof(uint16_t)); +} + +struct kl_divergence_result { + double sum_nll = 0; + double sum_nll2 = 0; + double sum_kld = 0; + double sum_kld2 = 0; + double sum_nll_diff = 0; + double sum_nll_diff2 = 0; + size_t count = 0; +}; + +static void log_softmax(int n_vocab, const float * logits, const uint16_t * base_log_prob, int tok, kl_divergence_result & kld) { + float max_logit = logits[0]; + for (int i = 1; i < n_vocab; ++i) { + max_logit = std::max(max_logit, logits[i]); + } + double sum_exp = 0.0; + for (int i = 0; i < n_vocab; ++i) { + sum_exp += expf(logits[i] - max_logit); + } + const float log_sum_exp = log(sum_exp); + const float * d = (const float *)base_log_prob; + const float scale = d[0]; + const float min_log_prob = d[1]; + base_log_prob += 4; + float nll = max_logit + log_sum_exp - logits[tok]; + kld.sum_nll += nll; + kld.sum_nll2 += nll*nll; + nll += (scale*base_log_prob[tok] + min_log_prob); + kld.sum_nll_diff += nll; + kld.sum_nll_diff2 += nll*nll; + max_logit += log_sum_exp; + double sum = 0; + for (int i = 0; i < n_vocab; ++i) { + const float p_log_base = scale*base_log_prob[i] + min_log_prob; + if (p_log_base > -16.f) { + const float p_base = expf(p_log_base); + sum += p_base * (p_log_base - logits[i] + max_logit); + } + } + kld.sum_kld += sum; + kld.sum_kld2 += sum*sum; + ++kld.count; +} + +static void process_logits(int n_vocab, const float * logits, const int * tokens, int n_token, + std::vector & workers, const std::vector & base_log_probs, kl_divergence_result & kld) { + std::mutex mutex; + const int nv = 2*((n_vocab + 1)/2) + 4; + int counter = 0; + auto compute = [&mutex, &counter, &base_log_probs, &kld, n_vocab, logits, tokens, n_token, nv] () { + kl_divergence_result local_kld; + while (true) { + std::unique_lock lock(mutex); + int i = counter++; + if (i >= n_token) { + kld.sum_nll += local_kld.sum_nll; + kld.sum_nll2 += local_kld.sum_nll2; + kld.sum_kld += local_kld.sum_kld; + kld.sum_kld2 += local_kld.sum_kld2; + kld.sum_nll_diff += local_kld.sum_nll_diff; + kld.sum_nll_diff2 += local_kld.sum_nll_diff2; + kld.count += local_kld.count; + break; + } + lock.unlock(); + log_softmax(n_vocab, logits + i*n_vocab, base_log_probs.data() + i*nv, tokens[i+1], local_kld); + } + }; + for (auto & w : workers) { + w = std::thread(compute); + } + compute(); + for (auto & w : workers) { + w.join(); + } +} + static results_perplexity perplexity_v2(llama_context * ctx, const gpt_params & params) { // Download: https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-2-raw-v1.zip?ref=salesforce-research // Run `./perplexity -m models/7B/ggml-model-q4_0.bin -f wiki.test.raw` @@ -294,6 +439,18 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx)); const int n_ctx = llama_n_ctx(ctx); + std::ofstream logits_stream; + if (!params.logits_file.empty()) { + logits_stream.open(params.logits_file.c_str()); + if (!logits_stream.is_open()) { + fprintf(stderr, "%s: failed to open %s for writing\n", __func__, params.logits_file.c_str()); + return {}; + } + fprintf(stderr, "%s: saving all logits to %s\n", __func__, params.logits_file.c_str()); + logits_stream.write("_logits_", 8); + logits_stream.write((const char *)&n_ctx, sizeof(n_ctx)); + } + auto tim1 = std::chrono::high_resolution_clock::now(); fprintf(stderr, "%s: tokenizing the input ..\n", __func__); @@ -336,6 +493,15 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par std::vector workers(std::thread::hardware_concurrency() - 1); + std::vector log_probs; + if (!params.logits_file.empty()) { + logits_stream.write((const char *)&n_vocab, sizeof(n_vocab)); + logits_stream.write((const char *)&n_chunk, sizeof(n_chunk)); + logits_stream.write((const char *)tokens.data(), n_chunk*n_ctx*sizeof(tokens[0])); + const int nv = 2*((n_vocab + 1)/2) + 4; + log_probs.resize(n_ctx * nv); + } + for (int i = 0; i < n_chunk; ++i) { const int start = i * n_ctx; const int end = start + n_ctx; @@ -398,8 +564,13 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par // process the entire prompt. const int first = n_ctx/2; const float * all_logits = num_batches > 1 ? logits.data() : llama_get_logits(ctx); - process_logits(n_vocab, all_logits + first*n_vocab, tokens.data() + start + first, n_ctx - 1 - first, - workers, nll, nll2, logit_history.data() + start + first, prob_history.data() + start + first); + if (!params.logits_file.empty()) { + process_logits(logits_stream, n_vocab, all_logits + first*n_vocab, tokens.data() + start + first, n_ctx - 1 - first, + workers, log_probs, nll, nll2); + } else { + process_logits(n_vocab, all_logits + first*n_vocab, tokens.data() + start + first, n_ctx - 1 - first, + workers, nll, nll2, logit_history.data() + start + first, prob_history.data() + start + first); + } count += n_ctx - first - 1; // perplexity is e^(average negative log-likelihood) @@ -1414,6 +1585,148 @@ static void multiple_choice_score(llama_context * ctx, const gpt_params & params printf("\n"); } +static void kl_divergence(llama_context * ctx, const gpt_params & params) { + if (params.logits_file.empty()) { + fprintf(stderr, "%s: you must provide a name of a file containing the log probabilities of the base model\n", __func__); + return; + } + std::ifstream in(params.logits_file.c_str(), std::ios::binary); + if (!in) { + fprintf(stderr, "%s: failed to open %s\n", __func__, params.logits_file.c_str()); + return; + } + { + char check[9]; check[8] = 0; + in.read(check, 8); + if (in.fail() || strncmp("_logits_", check, 8) != 0) { + fprintf(stderr, "%s: %s does not look like a file containing log-probabilities\n", __func__, params.logits_file.c_str()); + return; + } + } + + uint32_t n_ctx; + in.read((char *)&n_ctx, sizeof(n_ctx)); + if (n_ctx > llama_n_ctx(ctx)) { + fprintf(stderr, "%s: %s has been computed with %d, while the current context is %d. Increase it with -c and retry\n", + __func__, params.logits_file.c_str(), n_ctx, params.n_ctx); + } + + int n_vocab, n_chunk; + in.read((char *)&n_vocab, sizeof(n_vocab)); + in.read((char *)&n_chunk, sizeof(n_chunk)); + if (in.fail()) { + fprintf(stderr, "%s: failed rwading n_vocab, n_chunk from %s\n", __func__, params.logits_file.c_str()); + return; + } + if (n_vocab != llama_n_vocab(llama_get_model(ctx))) { + fprintf(stderr, "%s: inconsistent vocabulary (%d vs %d)\n", __func__, n_vocab, llama_n_vocab(llama_get_model(ctx))); + } + + std::vector tokens(n_ctx * n_chunk); + if (in.read((char *)tokens.data(), tokens.size()*sizeof(tokens[0])).fail()) { + fprintf(stderr, "%s: failed reading evaluation tokens from %s\n", __func__, params.logits_file.c_str()); + return; + } + + const int n_batch = params.n_batch; + const int num_batches = (n_ctx + n_batch - 1)/n_batch; + const int nv = 2*((n_vocab + 1)/2) + 4; + const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx)); + + std::vector log_probs_uint16(size_t(n_ctx - 1 - n_ctx/2) * nv); + std::vector logits; + if (num_batches > 1) { + logits.reserve(n_ctx * n_vocab); + } + + std::vector workers(std::thread::hardware_concurrency() - 1); + + auto mean_and_uncertainty = [] (double sum, double sum2, size_t count) { + if (count < 1) { + return std::make_pair(0., 0.); + } + double f = sum/count; + double df = sum2/count - f*f; + df = df > 0 && count > 10 ? sqrt(df/(count-1)) : 0.; + return std::make_pair(f, df); + }; + + kl_divergence_result kld; + + for (int i = 0; i < n_chunk; ++i) { + const int start = i * n_ctx; + const int end = start + n_ctx; + + const auto t_start = std::chrono::high_resolution_clock::now(); + + if (in.read((char *)log_probs_uint16.data(), log_probs_uint16.size()*sizeof(uint16_t)).fail()) { + fprintf(stderr, "%s: failed reading log-probs for chunk %d\n", __func__, i); + return; + } + + // clear the KV cache + llama_kv_cache_clear(ctx); + + for (int j = 0; j < num_batches; ++j) { + const int batch_start = start + j * n_batch; + const int batch_size = std::min(end - batch_start, n_batch); + + // save original token and restore it after eval + const auto token_org = tokens[batch_start]; + + // add BOS token for the first batch of each chunk + if (add_bos && j == 0) { + tokens[batch_start] = llama_token_bos(llama_get_model(ctx)); + } + + if (llama_decode(ctx, llama_batch_get_one(tokens.data() + batch_start, batch_size, j * n_batch, 0))) { + fprintf(stderr, "%s : failed to eval\n", __func__); + return; + } + + // restore the original token in case it was set to BOS + tokens[batch_start] = token_org; + + if (num_batches > 1) { + const auto * batch_logits = llama_get_logits(ctx); + logits.insert(logits.end(), batch_logits, batch_logits + batch_size * n_vocab); + } + } + + const auto t_end = std::chrono::high_resolution_clock::now(); + + if (i == 0) { + const float t_total = std::chrono::duration(t_end - t_start).count(); + fprintf(stderr, "%s: %.2f seconds per pass - ETA ", __func__, t_total); + int total_seconds = (int)(t_total * n_chunk); + if (total_seconds >= 60*60) { + fprintf(stderr, "%d hours ", total_seconds / (60*60)); + total_seconds = total_seconds % (60*60); + } + fprintf(stderr, "%.2f minutes\n", total_seconds / 60.0); + + printf("\nchunk PPL ln(PPL(Q)/PPL(base)) KL-Divergence\n"); + } + + const int first = n_ctx/2; + const float * all_logits = num_batches > 1 ? logits.data() : llama_get_logits(ctx); + process_logits(n_vocab, all_logits + first*n_vocab, tokens.data() + start + first, n_ctx - 1 - first, + workers, log_probs_uint16, kld); + + auto ppl = mean_and_uncertainty(kld.sum_nll, kld.sum_nll2, kld.count); + auto log_ppl_ratio = mean_and_uncertainty(kld.sum_nll_diff, kld.sum_nll_diff2, kld.count); + auto kl_div = mean_and_uncertainty(kld.sum_kld, kld.sum_kld2, kld.count); + + printf("%4d %10.4lf %10.5lf ± %10.5f %10.5f ± %10.5lf\n", i+1, exp(ppl.first), + log_ppl_ratio.first, log_ppl_ratio.second, kl_div.first, kl_div.second); + + fflush(stdout); + + logits.clear(); + } + printf("\n"); + +} int main(int argc, char ** argv) { gpt_params params; @@ -1476,6 +1789,8 @@ int main(int argc, char ** argv) { winogrande_score(ctx, params); } else if (params.multiple_choice) { multiple_choice_score(ctx, params); + } else if (params.kl_divergence) { + kl_divergence(ctx, params); } else { results = perplexity(ctx, params); } From 011e8ec577fd135cbc02993d3ea9840c516d6a1c Mon Sep 17 00:00:00 2001 From: slaren Date: Mon, 22 Jan 2024 23:42:41 +0100 Subject: [PATCH 28/66] llama : fix not enough space in buffer with Qwen (#5086) --- llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/llama.cpp b/llama.cpp index 8c906a22f..f6f1ec0f4 100644 --- a/llama.cpp +++ b/llama.cpp @@ -4440,9 +4440,9 @@ static struct ggml_tensor * llm_build_kv( // these nodes are added to the graph together so that they are not reordered // by doing so, the number of splits in the graph is reduced + ggml_build_forward_expand(graph, q_cur); ggml_build_forward_expand(graph, k_cur); ggml_build_forward_expand(graph, v_cur); - ggml_build_forward_expand(graph, q_cur); llm_build_kv_store(ctx, hparams, kv, graph, k_cur, v_cur, n_ctx, n_tokens, kv_head, cb, il); From 125d03a5036a02a983c8e98c2cdc126e061afb8e Mon Sep 17 00:00:00 2001 From: Michael Coppola Date: Tue, 23 Jan 2024 01:51:27 -0500 Subject: [PATCH 29/66] llama.vim : added api key support (#5090) Co-authored-by: Michael Coppola --- examples/llama.vim | 5 ++++- 1 file changed, 4 insertions(+), 1 deletion(-) diff --git a/examples/llama.vim b/examples/llama.vim index f03fadfb7..1b5ad6ba0 100644 --- a/examples/llama.vim +++ b/examples/llama.vim @@ -6,7 +6,7 @@ " Similarly, you could add an insert mode keybind with " inoremap call llama#doLlamaGen() " -" g:llama_api_url and g:llama_overrides can be configured in your .vimrc +" g:llama_api_url, g:llama_api_key and g:llama_overrides can be configured in your .vimrc " let g:llama_api_url = "192.168.1.10:8080" " llama_overrides can also be set through buffer/window scopes. For instance " autocmd filetype python let b:llama_overrides = {"temp": 0.2} @@ -82,6 +82,9 @@ func llama#doLlamaGen() endif let l:querydata.prompt = join(l:buflines, "\n") let l:curlcommand = copy(s:curlcommand) + if exists("g:llama_api_key") + call extend(l:curlcommand, ['--header', 'Authorization: Bearer ' .. g:llama_api_key]) + endif let l:curlcommand[2] = json_encode(l:querydata) let b:job = job_start(l:curlcommand, {"callback": function("s:callbackHandler", [l:cbuffer])}) endfunction From 2bed4aa3f37cb4e39e16e9ec7b595a7738fd5faf Mon Sep 17 00:00:00 2001 From: Xuan Son Nguyen Date: Tue, 23 Jan 2024 08:11:39 +0100 Subject: [PATCH 30/66] devops : add intel oneapi dockerfile (#5068) Co-authored-by: Xuan Son Nguyen --- .devops/main-intel.Dockerfile | 26 ++++++++++++++++++++++++++ .github/workflows/docker.yml | 1 + CMakeLists.txt | 5 +++++ 3 files changed, 32 insertions(+) create mode 100644 .devops/main-intel.Dockerfile diff --git a/.devops/main-intel.Dockerfile b/.devops/main-intel.Dockerfile new file mode 100644 index 000000000..e1e6acc24 --- /dev/null +++ b/.devops/main-intel.Dockerfile @@ -0,0 +1,26 @@ +ARG ONEAPI_VERSION=2024.0.1-devel-ubuntu22.04 +ARG UBUNTU_VERSION=22.04 + +FROM intel/hpckit:$ONEAPI_VERSION as build + +RUN apt-get update && \ + apt-get install -y git + +WORKDIR /app + +COPY . . + +# for some reasons, "-DLLAMA_BLAS=ON -DLLAMA_BLAS_VENDOR=Intel10_64lp -DLLAMA_NATIVE=ON" give worse performance +RUN mkdir build && \ + cd build && \ + cmake .. -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx && \ + cmake --build . --config Release --target main server + +FROM ubuntu:$UBUNTU_VERSION as runtime + +COPY --from=build /app/build/bin/main /main +COPY --from=build /app/build/bin/server /server + +ENV LC_ALL=C.utf8 + +ENTRYPOINT [ "/main" ] diff --git a/.github/workflows/docker.yml b/.github/workflows/docker.yml index 87904b75e..825b8f503 100644 --- a/.github/workflows/docker.yml +++ b/.github/workflows/docker.yml @@ -35,6 +35,7 @@ jobs: - { tag: "full-cuda", dockerfile: ".devops/full-cuda.Dockerfile", platforms: "linux/amd64" } - { tag: "light-rocm", dockerfile: ".devops/main-rocm.Dockerfile", platforms: "linux/amd64,linux/arm64" } - { tag: "full-rocm", dockerfile: ".devops/full-rocm.Dockerfile", platforms: "linux/amd64,linux/arm64" } + - { tag: "light-intel", dockerfile: ".devops/main-intel.Dockerfile", platforms: "linux/amd64" } steps: - name: Check out the repo uses: actions/checkout@v3 diff --git a/CMakeLists.txt b/CMakeLists.txt index 5a333ff52..af3665129 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -478,6 +478,11 @@ function(get_flags CCID CCVER) if (CCVER VERSION_GREATER_EQUAL 8.1.0) set(CXX_FLAGS ${CXX_FLAGS} -Wextra-semi) endif() + elseif (CCID MATCHES "Intel") + # enable max optimization level when using Intel compiler + set(C_FLAGS -ipo -O3 -static -fp-model=fast -flto -fno-stack-protector) + set(CXX_FLAGS -ipo -O3 -static -fp-model=fast -flto -fno-stack-protector) + add_link_options(-fuse-ld=lld -static-intel) endif() set(GF_C_FLAGS ${C_FLAGS} PARENT_SCOPE) From 89758723c75ba594e401f6513751beeba7ca1d28 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Tue, 23 Jan 2024 14:12:57 +0200 Subject: [PATCH 31/66] minor : clean-up some warnings and style (#5094) * minor : clean-up some warnings and style ggml-ci * ggml : add comment --- common/common.cpp | 10 +++---- examples/llava/clip.cpp | 47 ++++++++++++------------------ examples/perplexity/perplexity.cpp | 4 +-- ggml.c | 9 +++--- ggml.h | 1 - llama.cpp | 24 +++++++-------- 6 files changed, 42 insertions(+), 53 deletions(-) diff --git a/common/common.cpp b/common/common.cpp index 0a7096171..6b07f1197 100644 --- a/common/common.cpp +++ b/common/common.cpp @@ -216,12 +216,10 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) { } // store the external file name in params params.prompt_file = argv[i]; - file.seekg(0, std::ios::end); - size_t size = file.tellg(); - file.seekg(0, std::ios::beg); - params.prompt.resize(size); - file.read((char *)params.prompt.data(), size); - fprintf(stderr, "Read %zu bytes from binary file %s\n", size, argv[i]); + std::ostringstream ss; + ss << file.rdbuf(); + params.prompt = ss.str(); + fprintf(stderr, "Read %zu bytes from binary file %s\n", params.prompt.size(), argv[i]); } else if (arg == "-f" || arg == "--file") { if (++i >= argc) { invalid_param = true; diff --git a/examples/llava/clip.cpp b/examples/llava/clip.cpp index 6161fd858..4a0338a37 100644 --- a/examples/llava/clip.cpp +++ b/examples/llava/clip.cpp @@ -2,18 +2,6 @@ // so there might be still unnecessary artifacts hanging around // I'll gradually clean and extend it -#include -#include -#include -#include -#include -#include -#include -#include -#include -#include -#include - #include "clip.h" #include "ggml.h" #include "ggml-alloc.h" @@ -30,6 +18,19 @@ #define STB_IMAGE_IMPLEMENTATION #include "stb_image.h" +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include + static std::string format(const char * fmt, ...) { va_list ap; va_list ap2; @@ -217,9 +218,9 @@ static std::string gguf_kv_to_str(const struct gguf_context * ctx_gguf, int i) { static void print_tensor_info(const ggml_tensor* tensor, const char* prefix = "") { size_t tensor_size = ggml_nbytes(tensor); - printf("%s: n_dims = %d, name = %s, tensor_size=%zu, shape:[%d, %d, %d, %d], type: %d\n", + printf("%s: n_dims = %d, name = %s, tensor_size=%zu, shape:[%" PRId64 ", %" PRId64 ", %" PRId64 ", %" PRId64 "], type = %s\n", prefix, ggml_n_dims(tensor), tensor->name, tensor_size, - tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->ne[3], tensor->type); + tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->ne[3], ggml_type_name(tensor->type)); } static projector_type clip_projector_type_from_string(const std::string & name) { @@ -592,7 +593,7 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32 mlp_3 = ggml_cont(ctx0, ggml_permute(ctx0, mlp_3, 1, 0, 2, 3)); mlp_3 = ggml_reshape_4d(ctx0, mlp_3, n_patch, n_patch, mlp_3->ne[1], mlp_3->ne[2]); // stride = 1, padding = 1, bias is nullptr - block_1 = ggml_conv_depthwise_2d(ctx0, model.mm_model_block_1_block_0_0_w, mlp_3, nullptr, 1, 1, 1, 1, 1, 1); + block_1 = ggml_conv_depthwise_2d(ctx0, model.mm_model_block_1_block_0_0_w, mlp_3, 1, 1, 1, 1, 1, 1); // layer norm // // block_1 shape = [1, 2048, 24, 24], ne = [24, 24, 2048, 1] @@ -640,7 +641,7 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32 // block_2 { // stride = 2 - block_1 = ggml_conv_depthwise_2d(ctx0, model.mm_model_block_2_block_0_0_w, block_1, nullptr, 2, 2, 1, 1, 1, 1); + block_1 = ggml_conv_depthwise_2d(ctx0, model.mm_model_block_2_block_0_0_w, block_1, 2, 2, 1, 1, 1, 1); // block_1 shape = [1, 2048, 12, 12], ne = [12, 12, 2048, 1] // layer norm @@ -741,18 +742,10 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) { { std::map n_type; - uint32_t n_type_max = 0; - enum ggml_type type_max = GGML_TYPE_F32; - for (int i = 0; i < n_tensors; i++) { enum ggml_type type = gguf_get_tensor_type(ctx, i); n_type[type]++; - - if (n_type_max < n_type[type]) { - n_type_max = n_type[type]; - type_max = type; - } } printf("%s: Dumping metadata keys/values. Note: KV overrides do not apply in this output.\n", __func__); @@ -795,14 +788,12 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) { size_t tensor_size = ggml_nbytes(cur); buffer_size += tensor_size; if (verbosity >= 3) { - printf("%s: tensor[%d]: n_dims = %d, name = %s, tensor_size=%zu, offset=%zu, shape:[%d, %d, %d, %d], type: %d\n", __func__, i, - ggml_n_dims(cur), cur->name, tensor_size, offset, cur->ne[0], cur->ne[1], cur->ne[2], cur->ne[3], type); + printf("%s: tensor[%d]: n_dims = %d, name = %s, tensor_size=%zu, offset=%zu, shape:[%" PRIu64 ", %" PRIu64 ", %" PRIu64 ", %" PRIu64 "], type = %s\n", + __func__, i, ggml_n_dims(cur), cur->name, tensor_size, offset, cur->ne[0], cur->ne[1], cur->ne[2], cur->ne[3], ggml_type_name(type)); } } } - - buffer_size += n_tensors * 128 /* CLIP PADDING */; clip_ctx * new_clip = new clip_ctx; diff --git a/examples/perplexity/perplexity.cpp b/examples/perplexity/perplexity.cpp index 1b7f85f49..de6d3eb41 100644 --- a/examples/perplexity/perplexity.cpp +++ b/examples/perplexity/perplexity.cpp @@ -1202,11 +1202,11 @@ static void winogrande_score(llama_context * ctx, const gpt_params & params) { printf("Final Winogrande score(%d tasks): %.4lf +/- %.4lf\n", n_done, 100*p, sigma); } -static bool deserialize_string(std::istream& in, std::string& str) { +static bool deserialize_string(std::istream & in, std::string & str) { uint32_t size; if (!in.read((char *)&size, sizeof(size)).fail()) { str.resize(size); - if (!in.read((char *)str.data(), size).fail()) return true; + if (!in.read((char *)&str[0], size).fail()) return true; } return false; } diff --git a/ggml.c b/ggml.c index f85045c9c..ca98fde8a 100644 --- a/ggml.c +++ b/ggml.c @@ -5368,14 +5368,12 @@ struct ggml_tensor * ggml_conv_depthwise_2d( struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b, - struct ggml_tensor * c, int s0, int s1, int p0, int p1, int d0, int d1) { - struct ggml_tensor * new_a = ggml_reshape_4d(ctx, a, a->ne[0], a->ne[1], 1, a->ne[2] * a->ne[3]); struct ggml_tensor * im2col = ggml_im2col(ctx, new_a, ggml_reshape_4d(ctx, b, b->ne[0], b->ne[1], 1, b->ne[2] * b->ne[3]), @@ -9991,7 +9989,7 @@ static void ggml_compute_forward_mul_mat( return; } - const int64_t tgemm0 = ggml_perf_time_us(); + //const int64_t tgemm0 = ggml_perf_time_us(); for (int64_t i13 = 0; i13 < ne13; i13++) { for (int64_t i12 = 0; i12 < ne12; i12++) { const int64_t i03 = i13/r3; @@ -16934,7 +16932,10 @@ struct ggml_cplan ggml_graph_plan(const struct ggml_cgraph * cgraph, int n_threa if (ggml_compute_forward_mul_mat_use_blas(node)) { if (node->src[0]->type != GGML_TYPE_F32) { // here we need memory for fully dequantized matrix from src0 - cur = ggml_type_size(GGML_TYPE_F32)*ggml_nelements(node->src[0]); + // take into account that src0 can be broadcasted into src1[2,3] + cur = ggml_type_size(GGML_TYPE_F32) + * node->src[0]->ne[0]*node->src[0]->ne[1] + * node->src[1]->ne[2]*node->src[1]->ne[3]; } } else #endif diff --git a/ggml.h b/ggml.h index dca7bd9ce..1c4976271 100644 --- a/ggml.h +++ b/ggml.h @@ -1499,7 +1499,6 @@ extern "C" { struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b, - struct ggml_tensor * c, int s0, int s1, int p0, diff --git a/llama.cpp b/llama.cpp index f6f1ec0f4..582e82260 100644 --- a/llama.cpp +++ b/llama.cpp @@ -2300,18 +2300,18 @@ struct llama_model_loader { } switch (type_max) { - case GGML_TYPE_F32: ftype = LLAMA_FTYPE_ALL_F32; break; - case GGML_TYPE_F16: ftype = LLAMA_FTYPE_MOSTLY_F16; break; - case GGML_TYPE_Q4_0: ftype = LLAMA_FTYPE_MOSTLY_Q4_0; break; - case GGML_TYPE_Q4_1: ftype = LLAMA_FTYPE_MOSTLY_Q4_1; break; - case GGML_TYPE_Q5_0: ftype = LLAMA_FTYPE_MOSTLY_Q5_0; break; - case GGML_TYPE_Q5_1: ftype = LLAMA_FTYPE_MOSTLY_Q5_1; break; - case GGML_TYPE_Q8_0: ftype = LLAMA_FTYPE_MOSTLY_Q8_0; break; - case GGML_TYPE_Q2_K: ftype = LLAMA_FTYPE_MOSTLY_Q2_K; break; - case GGML_TYPE_Q3_K: ftype = LLAMA_FTYPE_MOSTLY_Q3_K_M; break; - case GGML_TYPE_Q4_K: ftype = LLAMA_FTYPE_MOSTLY_Q4_K_M; break; - case GGML_TYPE_Q5_K: ftype = LLAMA_FTYPE_MOSTLY_Q5_K_M; break; - case GGML_TYPE_Q6_K: ftype = LLAMA_FTYPE_MOSTLY_Q6_K; break; + case GGML_TYPE_F32: ftype = LLAMA_FTYPE_ALL_F32; break; + case GGML_TYPE_F16: ftype = LLAMA_FTYPE_MOSTLY_F16; break; + case GGML_TYPE_Q4_0: ftype = LLAMA_FTYPE_MOSTLY_Q4_0; break; + case GGML_TYPE_Q4_1: ftype = LLAMA_FTYPE_MOSTLY_Q4_1; break; + case GGML_TYPE_Q5_0: ftype = LLAMA_FTYPE_MOSTLY_Q5_0; break; + case GGML_TYPE_Q5_1: ftype = LLAMA_FTYPE_MOSTLY_Q5_1; break; + case GGML_TYPE_Q8_0: ftype = LLAMA_FTYPE_MOSTLY_Q8_0; break; + case GGML_TYPE_Q2_K: ftype = LLAMA_FTYPE_MOSTLY_Q2_K; break; + case GGML_TYPE_Q3_K: ftype = LLAMA_FTYPE_MOSTLY_Q3_K_M; break; + case GGML_TYPE_Q4_K: ftype = LLAMA_FTYPE_MOSTLY_Q4_K_M; break; + case GGML_TYPE_Q5_K: ftype = LLAMA_FTYPE_MOSTLY_Q5_K_M; break; + case GGML_TYPE_Q6_K: ftype = LLAMA_FTYPE_MOSTLY_Q6_K; break; case GGML_TYPE_IQ2_XXS: ftype = LLAMA_FTYPE_MOSTLY_IQ2_XXS; break; case GGML_TYPE_IQ2_XS: ftype = LLAMA_FTYPE_MOSTLY_IQ2_XS; break; default: From 9ecdd12e95aee20d6dfaf5f5a0f0ce5ac1fb2747 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Johannes=20G=C3=A4=C3=9Fler?= Date: Tue, 23 Jan 2024 13:31:56 +0100 Subject: [PATCH 32/66] CUDA: more info when no device code (#5088) --- ggml-cuda.cu | 89 +++++++++++++++++++++++++++++++--------------------- 1 file changed, 54 insertions(+), 35 deletions(-) diff --git a/ggml-cuda.cu b/ggml-cuda.cu index ec3837fb8..7f460449e 100644 --- a/ggml-cuda.cu +++ b/ggml-cuda.cu @@ -13,6 +13,10 @@ #include #include +// stringize macro for converting __CUDA_ARCH_LIST__ (list of integers) to string +#define STRINGIZE_IMPL(...) #__VA_ARGS__ +#define STRINGIZE(...) STRINGIZE_IMPL(__VA_ARGS__) + #if defined(GGML_USE_HIPBLAS) #include #include @@ -584,13 +588,28 @@ static cuda_device_capabilities g_device_caps[GGML_CUDA_MAX_DEVICES] = { {0, 0, static cublasHandle_t g_cublas_handles[GGML_CUDA_MAX_DEVICES] = {nullptr}; [[noreturn]] -static __device__ void bad_arch() { - printf("ERROR: ggml-cuda was compiled without support for the current GPU architecture.\n"); +static __device__ void no_device_code( + const char * file_name, const int line, const char * function_name, const int arch, const char * arch_list) { + +#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) + printf("%s:%d: ERROR: HIP kernel %s has no device code compatible with HIP arch %d.\n", + file_name, line, function_name, arch); + (void) arch_list; +#else + printf("%s:%d: ERROR: CUDA kernel %s has no device code compatible with CUDA arch %d. ggml-cuda.cu was compiled for: %s\n", + file_name, line, function_name, arch, arch_list); +#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) __trap(); - (void) bad_arch; // suppress unused function warning + (void) no_device_code; // suppress unused function warning } +#ifdef __CUDA_ARCH__ +#define NO_DEVICE_CODE no_device_code(__FILE__, __LINE__, __FUNCTION__, __CUDA_ARCH__, STRINGIZE(__CUDA_ARCH_LIST__)) +#else +#define NO_DEVICE_CODE GGML_ASSERT(false && "NO_DEVICE_CODE not valid in host code.") +#endif // __CUDA_ARCH__ + static __device__ __forceinline__ float warp_reduce_sum(float x) { #pragma unroll for (int mask = 16; mask > 0; mask >>= 1) { @@ -617,7 +636,7 @@ static __device__ __forceinline__ half2 warp_reduce_sum(half2 a) { return a; #else (void) a; - bad_arch(); + NO_DEVICE_CODE; #endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL } @@ -638,7 +657,7 @@ static __device__ __forceinline__ half2 warp_reduce_max(half2 x) { return x; #else (void) x; - bad_arch(); + NO_DEVICE_CODE; #endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL && CUDART_VERSION >= CUDART_HMAX } @@ -2421,7 +2440,7 @@ static __global__ void dequantize_block_q8_0_f16(const void * __restrict__ vx, h } #else (void) vx; (void) y; (void) k; - bad_arch(); + NO_DEVICE_CODE; #endif // __CUDA_ARCH__ >= CC_PASCAL } @@ -2452,7 +2471,7 @@ template static __device__ __forceinline__ float vec_dot_q4_0_q8_1_imp // second part effectively subtracts 8 from each quant value return d4 * (sumi * ds8f.x - (8*vdr/QI4_0) * ds8f.y); #else - bad_arch(); + NO_DEVICE_CODE; #endif // __CUDA_ARCH__ >= MIN_CC_DP4A } @@ -2489,7 +2508,7 @@ template static __device__ __forceinline__ float vec_dot_q4_1_q8_1_imp // scale second part of sum by QI8_1/(vdr * QR4_1) to compensate for multiple threads adding it return sumi * d4d8 + m4s8 / (QI8_1 / (vdr * QR4_1)); #else - bad_arch(); + NO_DEVICE_CODE; #endif // __CUDA_ARCH__ >= MIN_CC_DP4A } @@ -2524,7 +2543,7 @@ template static __device__ __forceinline__ float vec_dot_q5_0_q8_1_imp // second part effectively subtracts 16 from each quant value return d5 * (sumi * ds8f.x - (16*vdr/QI5_0) * ds8f.y); #else - bad_arch(); + NO_DEVICE_CODE; #endif // __CUDA_ARCH__ >= MIN_CC_DP4A } @@ -2569,7 +2588,7 @@ template static __device__ __forceinline__ float vec_dot_q5_1_q8_1_imp return sumi*d5d8 + m5s8 / (QI5_1 / vdr); #else - bad_arch(); + NO_DEVICE_CODE; #endif // __CUDA_ARCH__ >= MIN_CC_DP4A } @@ -2590,7 +2609,7 @@ template static __device__ __forceinline__ float vec_dot_q8_0_q8_1_imp return d8_0*d8_1 * sumi; #else - bad_arch(); + NO_DEVICE_CODE; #endif // __CUDA_ARCH__ >= MIN_CC_DP4A } @@ -2620,7 +2639,7 @@ template static __device__ __forceinline__ float vec_dot_q8_1_q8_1_imp // scale second part of sum by QI8_1/ vdr to compensate for multiple threads adding it return sumi*d8d8 + m8s8 / (QI8_1 / vdr); #else - bad_arch(); + NO_DEVICE_CODE; #endif // __CUDA_ARCH__ >= MIN_CC_DP4A } @@ -2655,7 +2674,7 @@ static __device__ __forceinline__ float vec_dot_q2_K_q8_1_impl_mmvq( return dm2f.x*sumf_d - dm2f.y*sumf_m; #else - bad_arch(); + NO_DEVICE_CODE; #endif // __CUDA_ARCH__ >= MIN_CC_DP4A } @@ -2692,7 +2711,7 @@ static __device__ __forceinline__ float vec_dot_q2_K_q8_1_impl_mmq( return d8 * (dm2f.x*sumi_d - dm2f.y*sumi_m); #else - bad_arch(); + NO_DEVICE_CODE; #endif // __CUDA_ARCH__ >= MIN_CC_DP4A } @@ -2732,7 +2751,7 @@ static __device__ __forceinline__ float vec_dot_q3_K_q8_1_impl_mmvq( return d3 * sumf; #else - bad_arch(); + NO_DEVICE_CODE; #endif // __CUDA_ARCH__ >= MIN_CC_DP4A } @@ -2757,7 +2776,7 @@ static __device__ __forceinline__ float vec_dot_q3_K_q8_1_impl_mmq( return d3*d8 * sumi; #else - bad_arch(); + NO_DEVICE_CODE; #endif // __CUDA_ARCH__ >= MIN_CC_DP4A } @@ -2790,7 +2809,7 @@ static __device__ __forceinline__ float vec_dot_q4_K_q8_1_impl_vmmq( return dm4f.x*sumf_d - dm4f.y*sumf_m; #else - bad_arch(); + NO_DEVICE_CODE; #endif // __CUDA_ARCH__ >= MIN_CC_DP4A } @@ -2823,7 +2842,7 @@ static __device__ __forceinline__ float vec_dot_q4_K_q8_1_impl_mmq( return dm4f.x*sumf_d - dm4f.y*sumf_m; #else - bad_arch(); + NO_DEVICE_CODE; #endif // __CUDA_ARCH__ >= MIN_CC_DP4A } @@ -2863,7 +2882,7 @@ static __device__ __forceinline__ float vec_dot_q5_K_q8_1_impl_vmmq( return dm5f.x*sumf_d - dm5f.y*sumf_m; #else - bad_arch(); + NO_DEVICE_CODE; #endif // __CUDA_ARCH__ >= MIN_CC_DP4A } @@ -2896,7 +2915,7 @@ static __device__ __forceinline__ float vec_dot_q5_K_q8_1_impl_mmq( return dm4f.x*sumf_d - dm4f.y*sumf_m; #else - bad_arch(); + NO_DEVICE_CODE; #endif // __CUDA_ARCH__ >= MIN_CC_DP4A } @@ -2926,7 +2945,7 @@ static __device__ __forceinline__ float vec_dot_q6_K_q8_1_impl_mmvq( return d*sumf; #else - bad_arch(); + NO_DEVICE_CODE; #endif // __CUDA_ARCH__ >= MIN_CC_DP4A } @@ -2957,7 +2976,7 @@ static __device__ __forceinline__ float vec_dot_q6_K_q8_1_impl_mmq( return d6 * sumf_d; #else - bad_arch(); + NO_DEVICE_CODE; #endif // __CUDA_ARCH__ >= MIN_CC_DP4A } @@ -3823,7 +3842,7 @@ static __device__ __forceinline__ float vec_dot_q4_K_q8_1( return dall * sumf_d - dmin * sumf_m; #else - bad_arch(); + NO_DEVICE_CODE; #endif // __CUDA_ARCH__ >= MIN_CC_DP4A #endif @@ -4006,7 +4025,7 @@ static __device__ __forceinline__ float vec_dot_q5_K_q8_1( return d * sumf_d; #else - bad_arch(); + NO_DEVICE_CODE; #endif // __CUDA_ARCH__ >= MIN_CC_DP4A #endif @@ -4501,7 +4520,7 @@ template static __global__ void (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); #else (void) vec_dot_q4_0_q8_1_mul_mat; - bad_arch(); + NO_DEVICE_CODE; #endif // __CUDA_ARCH__ >= CC_VOLTA } @@ -4570,7 +4589,7 @@ template static __global__ void (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); #else (void) vec_dot_q4_1_q8_1_mul_mat; - bad_arch(); + NO_DEVICE_CODE; #endif // __CUDA_ARCH__ >= CC_VOLTA } @@ -4637,7 +4656,7 @@ template static __global__ void (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); #else (void) vec_dot_q5_0_q8_1_mul_mat; - bad_arch(); + NO_DEVICE_CODE; #endif // __CUDA_ARCH__ >= CC_VOLTA } @@ -4704,7 +4723,7 @@ mul_mat_q5_1( (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); #else (void) vec_dot_q5_1_q8_1_mul_mat; - bad_arch(); + NO_DEVICE_CODE; #endif // __CUDA_ARCH__ >= CC_VOLTA } @@ -4771,7 +4790,7 @@ template static __global__ void (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); #else (void) vec_dot_q8_0_q8_1_mul_mat; - bad_arch(); + NO_DEVICE_CODE; #endif // __CUDA_ARCH__ >= CC_VOLTA } @@ -4838,7 +4857,7 @@ mul_mat_q2_K( (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); #else (void) vec_dot_q2_K_q8_1_mul_mat; - bad_arch(); + NO_DEVICE_CODE; #endif // __CUDA_ARCH__ >= CC_VOLTA } @@ -4907,7 +4926,7 @@ template static __global__ void (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); #else (void) vec_dot_q3_K_q8_1_mul_mat; - bad_arch(); + NO_DEVICE_CODE; #endif // __CUDA_ARCH__ >= CC_VOLTA } @@ -4976,7 +4995,7 @@ template static __global__ void (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); #else (void) vec_dot_q4_K_q8_1_mul_mat; - bad_arch(); + NO_DEVICE_CODE; #endif // __CUDA_ARCH__ >= CC_VOLTA } @@ -5043,7 +5062,7 @@ mul_mat_q5_K( (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); #else (void) vec_dot_q5_K_q8_1_mul_mat; - bad_arch(); + NO_DEVICE_CODE; #endif // __CUDA_ARCH__ >= CC_VOLTA } @@ -5112,7 +5131,7 @@ template static __global__ void (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); #else (void) vec_dot_q6_K_q8_1_mul_mat; - bad_arch(); + NO_DEVICE_CODE; #endif // __CUDA_ARCH__ >= CC_VOLTA } @@ -5835,7 +5854,7 @@ static __global__ void soft_max_f16(const float * x, const float * y, float * ds } #else (void) x; (void) y; (void) dst; (void) ncols_par; (void) nrows_y; (void) scale; - bad_arch(); + NO_DEVICE_CODE; #endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL && CUDART_VERSION >= CUDART_HMAX } From 44879ee885f48ecf4675dd216b373dce0a6f3690 Mon Sep 17 00:00:00 2001 From: Kawrakow <48489457+ikawrakow@users.noreply.github.com> Date: Tue, 23 Jan 2024 15:17:20 +0200 Subject: [PATCH 33/66] Additional KL-divergence statistics (#5081) * perplexity: add top-token probability * perplexity: add additional KL-divergence statistics * perplexity: a better organized KL-divergence statistics output --------- Co-authored-by: Iwan Kawrakow --- examples/perplexity/perplexity.cpp | 71 +++++++++++++++++++++++++----- 1 file changed, 61 insertions(+), 10 deletions(-) diff --git a/examples/perplexity/perplexity.cpp b/examples/perplexity/perplexity.cpp index de6d3eb41..8d2204969 100644 --- a/examples/perplexity/perplexity.cpp +++ b/examples/perplexity/perplexity.cpp @@ -222,13 +222,18 @@ struct kl_divergence_result { double sum_kld2 = 0; double sum_nll_diff = 0; double sum_nll_diff2 = 0; + size_t n_same_top = 0; size_t count = 0; }; -static void log_softmax(int n_vocab, const float * logits, const uint16_t * base_log_prob, int tok, kl_divergence_result & kld) { +static double log_softmax(int n_vocab, const float * logits, const uint16_t * base_log_prob, int tok, kl_divergence_result & kld) { float max_logit = logits[0]; + int imax = 0; for (int i = 1; i < n_vocab; ++i) { - max_logit = std::max(max_logit, logits[i]); + if (logits[i] > max_logit) { + max_logit = logits[i]; + imax = i; + } } double sum_exp = 0.0; for (int i = 0; i < n_vocab; ++i) { @@ -247,8 +252,14 @@ static void log_softmax(int n_vocab, const float * logits, const uint16_t * base kld.sum_nll_diff2 += nll*nll; max_logit += log_sum_exp; double sum = 0; + int imax_base = -1; + float p_log_base_max = 0; for (int i = 0; i < n_vocab; ++i) { const float p_log_base = scale*base_log_prob[i] + min_log_prob; + if (i == 0 || p_log_base > p_log_base_max) { + p_log_base_max = p_log_base; + imax_base = i; + } if (p_log_base > -16.f) { const float p_base = expf(p_log_base); sum += p_base * (p_log_base - logits[i] + max_logit); @@ -257,14 +268,17 @@ static void log_softmax(int n_vocab, const float * logits, const uint16_t * base kld.sum_kld += sum; kld.sum_kld2 += sum*sum; ++kld.count; + if (imax == imax_base) ++kld.n_same_top; + return sum; } static void process_logits(int n_vocab, const float * logits, const int * tokens, int n_token, - std::vector & workers, const std::vector & base_log_probs, kl_divergence_result & kld) { + std::vector & workers, const std::vector & base_log_probs, kl_divergence_result & kld, + float * kld_values) { std::mutex mutex; const int nv = 2*((n_vocab + 1)/2) + 4; int counter = 0; - auto compute = [&mutex, &counter, &base_log_probs, &kld, n_vocab, logits, tokens, n_token, nv] () { + auto compute = [&mutex, &counter, &base_log_probs, &kld, n_vocab, logits, tokens, n_token, nv, kld_values] () { kl_divergence_result local_kld; while (true) { std::unique_lock lock(mutex); @@ -276,11 +290,13 @@ static void process_logits(int n_vocab, const float * logits, const int * tokens kld.sum_kld2 += local_kld.sum_kld2; kld.sum_nll_diff += local_kld.sum_nll_diff; kld.sum_nll_diff2 += local_kld.sum_nll_diff2; + kld.n_same_top += local_kld.n_same_top; kld.count += local_kld.count; break; } lock.unlock(); - log_softmax(n_vocab, logits + i*n_vocab, base_log_probs.data() + i*nv, tokens[i+1], local_kld); + double v = log_softmax(n_vocab, logits + i*n_vocab, base_log_probs.data() + i*nv, tokens[i+1], local_kld); + kld_values[i] = (float)v; } }; for (auto & w : workers) { @@ -1615,7 +1631,7 @@ static void kl_divergence(llama_context * ctx, const gpt_params & params) { in.read((char *)&n_vocab, sizeof(n_vocab)); in.read((char *)&n_chunk, sizeof(n_chunk)); if (in.fail()) { - fprintf(stderr, "%s: failed rwading n_vocab, n_chunk from %s\n", __func__, params.logits_file.c_str()); + fprintf(stderr, "%s: failed reading n_vocab, n_chunk from %s\n", __func__, params.logits_file.c_str()); return; } if (n_vocab != llama_n_vocab(llama_get_model(ctx))) { @@ -1634,6 +1650,7 @@ static void kl_divergence(llama_context * ctx, const gpt_params & params) { const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx)); std::vector log_probs_uint16(size_t(n_ctx - 1 - n_ctx/2) * nv); + std::vector kld_values(size_t(n_ctx - 1 - n_ctx/2)*n_chunk); std::vector logits; if (num_batches > 1) { logits.reserve(n_ctx * n_vocab); @@ -1652,6 +1669,7 @@ static void kl_divergence(llama_context * ctx, const gpt_params & params) { }; kl_divergence_result kld; + auto kld_ptr = kld_values.data(); for (int i = 0; i < n_chunk; ++i) { const int start = i * n_ctx; @@ -1705,20 +1723,24 @@ static void kl_divergence(llama_context * ctx, const gpt_params & params) { } fprintf(stderr, "%.2f minutes\n", total_seconds / 60.0); - printf("\nchunk PPL ln(PPL(Q)/PPL(base)) KL-Divergence\n"); + printf("\nchunk PPL ln(PPL(Q)/PPL(base)) KL-Divergence Same top\n"); } const int first = n_ctx/2; const float * all_logits = num_batches > 1 ? logits.data() : llama_get_logits(ctx); process_logits(n_vocab, all_logits + first*n_vocab, tokens.data() + start + first, n_ctx - 1 - first, - workers, log_probs_uint16, kld); + workers, log_probs_uint16, kld, kld_ptr); + kld_ptr += n_ctx - 1 - first; auto ppl = mean_and_uncertainty(kld.sum_nll, kld.sum_nll2, kld.count); auto log_ppl_ratio = mean_and_uncertainty(kld.sum_nll_diff, kld.sum_nll_diff2, kld.count); auto kl_div = mean_and_uncertainty(kld.sum_kld, kld.sum_kld2, kld.count); + auto p_top = 1.*kld.n_same_top/kld.count; + auto d_p_top = sqrt(p_top*(1 - p_top)/(kld.count - 1)); - printf("%4d %10.4lf %10.5lf ± %10.5f %10.5f ± %10.5lf\n", i+1, exp(ppl.first), - log_ppl_ratio.first, log_ppl_ratio.second, kl_div.first, kl_div.second); + printf("%4d %10.4lf %10.5lf ± %10.5f %10.5f ± %10.5lf %.5f ± %.5f\n", i+1, exp(ppl.first), + log_ppl_ratio.first, log_ppl_ratio.second, kl_div.first, kl_div.second, + p_top, d_p_top); fflush(stdout); @@ -1726,6 +1748,35 @@ static void kl_divergence(llama_context * ctx, const gpt_params & params) { } printf("\n"); + if (kld.count < 100) return; // we do not wish to do statistics on so few values + + std::sort(kld_values.begin(), kld_values.end()); + + printf("===== KL-divergence statistics\n"); + auto kl_div = mean_and_uncertainty(kld.sum_kld, kld.sum_kld2, kld.count); + printf("Average: %10.6f ±%10.6lf\n", kl_div.first, kl_div.second); + auto kld_median = kld_values.size()%2 == 0 ? 0.5f*(kld_values[kld_values.size()/2] + kld_values[kld_values.size()/2-1]) + : kld_values[kld_values.size()/2]; + printf("Median : %10.6f\n", kld_median); + + auto percentile = [&kld_values] (float fraction) { + if (fraction <= 0) return kld_values.front(); + if (fraction >= 1) return kld_values.back(); + float p = fraction*(kld_values.size() - 1); + size_t ip = size_t(p); p -= ip; + return (1 - p)*kld_values[ip] + p*kld_values[std::min(ip+1, kld_values.size()-1)]; + }; + + printf("Maximum: %10.6f\n", kld_values.back()); + printf("KLD_99 : %10.6f\n", percentile(0.99f)); + printf("KLD_95 : %10.6f\n", percentile(0.95f)); + printf("KLD_90 : %10.6f\n", percentile(0.90f)); + + printf("Minimum: %10.6f\n", kld_values.front()); + printf("KLD_01 : %10.6f\n", percentile(0.01f)); + printf("KLD_05 : %10.6f\n", percentile(0.05f)); + printf("KLD_10 : %10.6f\n", percentile(0.10f)); + } int main(int argc, char ** argv) { From 26d607608d794efa56df3bdb6043a2f94c1d632c Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Tue, 23 Jan 2024 15:50:56 +0200 Subject: [PATCH 34/66] metal : disable support for MUL_MAT F32 x F16 --- ggml-metal.m | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/ggml-metal.m b/ggml-metal.m index 912ddc83f..4b3eb4914 100644 --- a/ggml-metal.m +++ b/ggml-metal.m @@ -668,7 +668,8 @@ static bool ggml_metal_supports_op(const struct ggml_metal_context * ctx, const return true; case GGML_OP_MUL_MAT: case GGML_OP_MUL_MAT_ID: - return ctx->support_simdgroup_reduction; + return ctx->support_simdgroup_reduction && + (op->src[0]->type != GGML_TYPE_F32 || op->src[1]->type == GGML_TYPE_F32); case GGML_OP_CPY: case GGML_OP_DUP: case GGML_OP_CONT: From 1387ea21178f9f154944013d4dd9764b54c69deb Mon Sep 17 00:00:00 2001 From: slaren Date: Wed, 24 Jan 2024 12:48:14 +0100 Subject: [PATCH 35/66] llama : pre-allocate input tensors in a separate buffer (#5100) --- ggml-alloc.c | 4 +- llama.cpp | 345 ++++++++++++++++++++++++--------------------------- 2 files changed, 167 insertions(+), 182 deletions(-) diff --git a/ggml-alloc.c b/ggml-alloc.c index 89b85d348..60141a34d 100644 --- a/ggml-alloc.c +++ b/ggml-alloc.c @@ -109,8 +109,8 @@ void ggml_tallocr_alloc(ggml_tallocr_t alloc, struct ggml_tensor * tensor) { if (block->size >= size) { best_fit_block = alloc->n_free_blocks - 1; } else { - fprintf(stderr, "%s: not enough space in the buffer (needed %zu, largest block available %zu)\n", - __func__, size, max_avail); + fprintf(stderr, "%s: not enough space in the buffer to allocate %s (needed %zu, largest block available %zu)\n", + __func__, tensor->name, size, max_avail); GGML_ASSERT(!"not enough space in the buffer"); return; } diff --git a/llama.cpp b/llama.cpp index 582e82260..114046db9 100644 --- a/llama.cpp +++ b/llama.cpp @@ -1669,6 +1669,9 @@ struct llama_context { for (ggml_backend_t backend : backends) { ggml_backend_free(backend); } + + ggml_backend_buffer_free(buf_input); + ggml_free(ctx_input); } llama_cparams cparams; @@ -1715,8 +1718,14 @@ struct llama_context { // allocator for the input tensors ggml_tallocr * alloc = nullptr; - // temporary buffer for copying data to/from the backend - std::vector> buf_copy; + // input tensors + ggml_backend_buffer_t buf_input = nullptr; + ggml_context * ctx_input = nullptr; + struct ggml_tensor * inp_tokens; // I32 [n_batch] + struct ggml_tensor * inp_embd; // F32 [n_embd, n_batch] + struct ggml_tensor * inp_pos; // I32 [n_batch] + struct ggml_tensor * inp_KQ_mask; // F32 [n_ctx, n_batch] + struct ggml_tensor * inp_K_shift; // I32 [n_ctx] #ifdef GGML_USE_MPI ggml_mpi_context * ctx_mpi = NULL; @@ -4089,22 +4098,24 @@ static struct ggml_tensor * llm_build_inp_embd( const llama_hparams & hparams, const llama_batch & batch, struct ggml_tensor * tok_embd, + struct ggml_tensor * inp_tokens, + struct ggml_tensor * inp_embd, const llm_build_cb & cb) { const int64_t n_embd = hparams.n_embd; struct ggml_tensor * inpL; if (batch.token) { - struct ggml_tensor * inp_tokens = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, batch.n_tokens); + struct ggml_tensor * inp_tokens_v = ggml_view_1d(ctx, inp_tokens, batch.n_tokens, 0); cb(inp_tokens, "inp_tokens", -1); - inpL = ggml_get_rows(ctx, tok_embd, inp_tokens); + inpL = ggml_get_rows(ctx, tok_embd, inp_tokens_v); } else { #ifdef GGML_USE_MPI GGML_ASSERT(false && "not implemented"); #endif - inpL = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, batch.n_tokens); + inpL = ggml_view_2d(ctx, inp_embd, n_embd, batch.n_tokens, inp_embd->nb[1], 0); } return inpL; @@ -4118,6 +4129,7 @@ static void llm_build_k_shift( const llama_cparams & cparams, const llama_kv_cache & kv, struct ggml_cgraph * graph, + struct ggml_tensor * K_shift, llm_rope_type type, int64_t n_ctx, float freq_base, @@ -4134,9 +4146,6 @@ static void llm_build_k_shift( const float beta_fast = cparams.yarn_beta_fast; const float beta_slow = cparams.yarn_beta_slow; - struct ggml_tensor * K_shift = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, n_ctx); - cb(K_shift, "K_shift", -1); - int rope_type = 0; switch (type) { @@ -4457,6 +4466,7 @@ static struct ggml_tensor * llm_build_kv( struct llm_build_context { const llama_model & model; + const llama_context & lctx; const llama_hparams & hparams; const llama_cparams & cparams; const llama_batch & batch; @@ -4503,6 +4513,7 @@ struct llm_build_context { const llm_build_cb & cb, bool worst_case) : model (lctx.model), + lctx (lctx), hparams (model.hparams), cparams (lctx.cparams), batch (batch), @@ -4563,20 +4574,20 @@ struct llm_build_context { struct ggml_tensor * cur; struct ggml_tensor * inpL; - inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, cb); + inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, lctx.inp_tokens, lctx.inp_embd, cb); cb(inpL, "inp_embd", -1); // inp_pos - contains the positions - struct ggml_tensor * inp_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens); + struct ggml_tensor * inp_pos = ggml_view_1d(ctx0, lctx.inp_pos, n_tokens, 0); cb(inp_pos, "inp_pos", -1); // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1); + struct ggml_tensor * KQ_mask = ggml_view_2d(ctx0, lctx.inp_KQ_mask, n_kv, n_tokens, n_kv*ggml_type_size(lctx.inp_KQ_mask->type), 0); cb(KQ_mask, "KQ_mask", -1); // shift the entire K-cache if needed if (do_rope_shift) { - llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE, n_ctx, freq_base, freq_scale, cb); + llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, lctx.inp_K_shift, LLM_ROPE, n_ctx, freq_base, freq_scale, cb); } for (int il = 0; il < n_layer; ++il) { @@ -4747,20 +4758,20 @@ struct llm_build_context { struct ggml_tensor * cur; struct ggml_tensor * inpL; - inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, cb); + inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, lctx.inp_tokens, lctx.inp_embd, cb); cb(inpL, "inp_embd", -1); // inp_pos - contains the positions - struct ggml_tensor * inp_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens); + struct ggml_tensor * inp_pos = ggml_view_1d(ctx0, lctx.inp_pos, n_tokens, 0); cb(inp_pos, "inp_pos", -1); // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1); + struct ggml_tensor * KQ_mask = ggml_view_2d(ctx0, lctx.inp_KQ_mask, n_kv, n_tokens, n_kv*ggml_type_size(lctx.inp_KQ_mask->type), 0); cb(KQ_mask, "KQ_mask", -1); // shift the entire K-cache if needed if (do_rope_shift) { - llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE, n_ctx, freq_base, freq_scale, cb); + llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, lctx.inp_K_shift, LLM_ROPE, n_ctx, freq_base, freq_scale, cb); } for (int il = 0; il < n_layer; ++il) { @@ -4868,20 +4879,20 @@ struct llm_build_context { struct ggml_tensor * cur; struct ggml_tensor * inpL; - inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, cb); + inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, lctx.inp_tokens, lctx.inp_embd, cb); cb(inpL, "inp_embd", -1); // inp_pos - contains the positions - struct ggml_tensor * inp_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens); + struct ggml_tensor * inp_pos = ggml_view_1d(ctx0, lctx.inp_pos, n_tokens, 0); cb(inp_pos, "inp_pos", -1); // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1); + struct ggml_tensor * KQ_mask = ggml_view_2d(ctx0, lctx.inp_KQ_mask, n_kv, n_tokens, n_kv*ggml_type_size(lctx.inp_KQ_mask->type), 0); cb(KQ_mask, "KQ_mask", -1); // shift the entire K-cache if needed if (do_rope_shift) { - llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE_NEOX, n_ctx, freq_base, freq_scale, cb); + llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, lctx.inp_K_shift, LLM_ROPE_NEOX, n_ctx, freq_base, freq_scale, cb); } for (int il = 0; il < n_layer; ++il) { @@ -4990,15 +5001,15 @@ struct llm_build_context { struct ggml_tensor * pos; struct ggml_tensor * inpL; - inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, cb); + inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, lctx.inp_tokens, lctx.inp_embd, cb); cb(inpL, "inp_embd", -1); // inp_pos - contains the positions - struct ggml_tensor * inp_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens); + struct ggml_tensor * inp_pos = ggml_view_1d(ctx0, lctx.inp_pos, n_tokens, 0); cb(inp_pos, "inp_pos", -1); // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1); + struct ggml_tensor * KQ_mask = ggml_view_2d(ctx0, lctx.inp_KQ_mask, n_kv, n_tokens, n_kv*ggml_type_size(lctx.inp_KQ_mask->type), 0); cb(KQ_mask, "KQ_mask", -1); pos = ggml_get_rows(ctx0, model.pos_embd, inp_pos); @@ -5087,19 +5098,19 @@ struct llm_build_context { struct ggml_tensor * cur; struct ggml_tensor * inpL; - inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, cb); + inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, lctx.inp_tokens, lctx.inp_embd, cb); cb(inpL, "inp_embd", -1); // inp_pos - contains the positions - struct ggml_tensor * inp_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens); + struct ggml_tensor * inp_pos = ggml_view_1d(ctx0, lctx.inp_pos, n_tokens, 0); cb(inp_pos, "inp_pos", -1); // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1); + struct ggml_tensor * KQ_mask = ggml_view_2d(ctx0, lctx.inp_KQ_mask, n_kv, n_tokens, n_kv*ggml_type_size(lctx.inp_KQ_mask->type), 0); cb(KQ_mask, "KQ_mask", -1); if (do_rope_shift) { - llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE_NEOX, n_ctx, freq_base, freq_scale, cb); + llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, lctx.inp_K_shift, LLM_ROPE_NEOX, n_ctx, freq_base, freq_scale, cb); } for (int il = 0; il < n_layer; ++il) { @@ -5294,11 +5305,11 @@ struct llm_build_context { struct ggml_tensor * cur; struct ggml_tensor * inpL; - inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, cb); + inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, lctx.inp_tokens, lctx.inp_embd, cb); cb(inpL, "inp_embd", -1); // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1); + struct ggml_tensor * KQ_mask = ggml_view_2d(ctx0, lctx.inp_KQ_mask, n_kv, n_tokens, n_kv*ggml_type_size(lctx.inp_KQ_mask->type), 0); cb(KQ_mask, "KQ_mask", -1); for (int il = 0; il < n_layer; ++il) { @@ -5384,11 +5395,11 @@ struct llm_build_context { struct ggml_tensor * cur; struct ggml_tensor * inpL; - inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, cb); + inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, lctx.inp_tokens, lctx.inp_embd, cb); cb(inpL, "inp_embd", -1); // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1); + struct ggml_tensor * KQ_mask = ggml_view_2d(ctx0, lctx.inp_KQ_mask, n_kv, n_tokens, n_kv*ggml_type_size(lctx.inp_KQ_mask->type), 0); cb(KQ_mask, "KQ_mask", -1); inpL = llm_build_norm(ctx0, inpL, hparams, @@ -5477,11 +5488,11 @@ struct llm_build_context { struct ggml_tensor * cur; struct ggml_tensor * inpL; - inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, cb); + inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, lctx.inp_tokens, lctx.inp_embd, cb); cb(inpL, "inp_embd", -1); // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1); + struct ggml_tensor * KQ_mask = ggml_view_2d(ctx0, lctx.inp_KQ_mask, n_kv, n_tokens, n_kv*ggml_type_size(lctx.inp_KQ_mask->type), 0); cb(KQ_mask, "KQ_mask", -1); for (int il = 0; il < n_layer; ++il) { @@ -5573,20 +5584,20 @@ struct llm_build_context { struct ggml_tensor * cur; struct ggml_tensor * inpL; - inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, cb); + inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, lctx.inp_tokens, lctx.inp_embd, cb); cb(inpL, "inp_embd", -1); // inp_pos - contains the positions - struct ggml_tensor * inp_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens); + struct ggml_tensor * inp_pos = ggml_view_1d(ctx0, lctx.inp_pos, n_tokens, 0); cb(inp_pos, "inp_pos", -1); // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1); + struct ggml_tensor * KQ_mask = ggml_view_2d(ctx0, lctx.inp_KQ_mask, n_kv, n_tokens, n_kv*ggml_type_size(lctx.inp_KQ_mask->type), 0); cb(KQ_mask, "KQ_mask", -1); // shift the entire K-cache if needed if (do_rope_shift) { - llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE_NEOX, n_ctx, freq_base, freq_scale, cb); + llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, lctx.inp_K_shift, LLM_ROPE_NEOX, n_ctx, freq_base, freq_scale, cb); } for (int il = 0; il < n_layer; ++il) { @@ -5696,20 +5707,20 @@ struct llm_build_context { struct ggml_tensor * cur; struct ggml_tensor * inpL; - inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, cb); + inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, lctx.inp_tokens, lctx.inp_embd, cb); cb(inpL, "inp_embd", -1); // inp_pos - contains the positions - struct ggml_tensor * inp_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens); + struct ggml_tensor * inp_pos = ggml_view_1d(ctx0, lctx.inp_pos, n_tokens, 0); cb(inp_pos, "inp_pos", -1); // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1); + struct ggml_tensor * KQ_mask = ggml_view_2d(ctx0, lctx.inp_KQ_mask, n_kv, n_tokens, n_kv*ggml_type_size(lctx.inp_KQ_mask->type), 0); cb(KQ_mask, "KQ_mask", -1); // shift the entire K-cache if needed if (do_rope_shift) { - llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE_NEOX, n_ctx, freq_base, freq_scale, cb); + llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, lctx.inp_K_shift, LLM_ROPE_NEOX, n_ctx, freq_base, freq_scale, cb); } for (int il = 0; il < n_layer; ++il) { @@ -5810,20 +5821,20 @@ struct llm_build_context { struct ggml_tensor * cur; struct ggml_tensor * inpL; - inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, cb); + inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, lctx.inp_tokens, lctx.inp_embd, cb); cb(inpL, "inp_embd", -1); // inp_pos - contains the positions - struct ggml_tensor * inp_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens); + struct ggml_tensor * inp_pos = ggml_view_1d(ctx0, lctx.inp_pos, n_tokens, 0); cb(inp_pos, "inp_pos", -1); // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1); + struct ggml_tensor * KQ_mask = ggml_view_2d(ctx0, lctx.inp_KQ_mask, n_kv, n_tokens, n_kv*ggml_type_size(lctx.inp_KQ_mask->type), 0); cb(KQ_mask, "KQ_mask", -1); // shift the entire K-cache if needed if (do_rope_shift) { - llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE_NEOX, n_ctx, freq_base, freq_scale, cb); + llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, lctx.inp_K_shift, LLM_ROPE_NEOX, n_ctx, freq_base, freq_scale, cb); } for (int il = 0; il < n_layer; ++il) { @@ -5931,20 +5942,20 @@ struct llm_build_context { struct ggml_tensor * ffn_output; struct ggml_tensor * inpL; - inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, cb); + inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, lctx.inp_tokens, lctx.inp_embd, cb); cb(inpL, "inp_embd", -1); // inp_pos - contains the positions - struct ggml_tensor * inp_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens); + struct ggml_tensor * inp_pos = ggml_view_1d(ctx0, lctx.inp_pos, n_tokens, 0); cb(inp_pos, "inp_pos", -1); // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1); + struct ggml_tensor * KQ_mask = ggml_view_2d(ctx0, lctx.inp_KQ_mask, n_kv, n_tokens, n_kv*ggml_type_size(lctx.inp_KQ_mask->type), 0); cb(KQ_mask, "KQ_mask", -1); // shift the entire K-cache if needed if (do_rope_shift) { - llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE_NEOX, n_ctx, freq_base, freq_scale, cb); + llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, lctx.inp_K_shift, LLM_ROPE_NEOX, n_ctx, freq_base, freq_scale, cb); } for (int il = 0; il < n_layer; ++il) { @@ -6053,20 +6064,20 @@ struct llm_build_context { struct ggml_tensor * cur; struct ggml_tensor * inpL; - inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, cb); + inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, lctx.inp_tokens, lctx.inp_embd, cb); cb(inpL, "inp_embd", -1); // inp_pos - contains the positions - struct ggml_tensor * inp_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens); + struct ggml_tensor * inp_pos = ggml_view_1d(ctx0, lctx.inp_pos, n_tokens, 0); cb(inp_pos, "inp_pos", -1); // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1); + struct ggml_tensor * KQ_mask = ggml_view_2d(ctx0, lctx.inp_KQ_mask, n_kv, n_tokens, n_kv*ggml_type_size(lctx.inp_KQ_mask->type), 0); cb(KQ_mask, "KQ_mask", -1); // shift the entire K-cache if needed if (do_rope_shift) { - llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE, n_ctx, freq_base, freq_scale, cb); + llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, lctx.inp_K_shift, LLM_ROPE, n_ctx, freq_base, freq_scale, cb); } for (int il = 0; il < n_layer; ++il) { @@ -6160,15 +6171,15 @@ struct llm_build_context { struct ggml_tensor * pos; struct ggml_tensor * inpL; - inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, cb); + inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, lctx.inp_tokens, lctx.inp_embd, cb); cb(inpL, "inp_embd", -1); // inp_pos - contains the positions - struct ggml_tensor * inp_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens); + struct ggml_tensor * inp_pos = ggml_view_1d(ctx0, lctx.inp_pos, n_tokens, 0); cb(inp_pos, "inp_pos", -1); // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1); + struct ggml_tensor * KQ_mask = ggml_view_2d(ctx0, lctx.inp_KQ_mask, n_kv, n_tokens, n_kv*ggml_type_size(lctx.inp_KQ_mask->type), 0); cb(KQ_mask, "KQ_mask", -1); pos = ggml_get_rows(ctx0, model.pos_embd, inp_pos); @@ -6258,20 +6269,20 @@ struct llm_build_context { struct ggml_tensor * cur; struct ggml_tensor * inpL; - inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, cb); + inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, lctx.inp_tokens, lctx.inp_embd, cb); cb(inpL, "inp_embd", -1); // inp_pos - contains the positions - struct ggml_tensor * inp_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens); + struct ggml_tensor * inp_pos = ggml_view_1d(ctx0, lctx.inp_pos, n_tokens, 0); cb(inp_pos, "inp_pos", -1); // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1); + struct ggml_tensor * KQ_mask = ggml_view_2d(ctx0, lctx.inp_KQ_mask, n_kv, n_tokens, n_kv*ggml_type_size(lctx.inp_KQ_mask->type), 0); cb(KQ_mask, "KQ_mask", -1); // shift the entire K-cache if needed if (do_rope_shift) { - llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE, n_ctx, freq_base, freq_scale, cb); + llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, lctx.inp_K_shift, LLM_ROPE, n_ctx, freq_base, freq_scale, cb); } for (int il = 0; il < n_layer; ++il) { @@ -6365,15 +6376,7 @@ static struct ggml_cgraph * llama_build_graph( // check if we should build the worst-case graph (for memory measurement) const bool worst_case = ggml_tallocr_is_measure(lctx.alloc); - // keep track of the input that has already been allocated - bool alloc_inp_tokens = false; - bool alloc_inp_embd = false; - bool alloc_inp_pos = false; - bool alloc_inp_KQ_mask = false; - bool alloc_inp_K_shift = false; - // this callback allows us to apply custom logic to each tensor (e.g. ggml-alloc, offloading, etc.) - // TODO: improve handling of input and output tensors, then replace this with ggml_set_name llm_build_cb cb = [&](struct ggml_tensor * cur, const char * name, int il) { if (il >= 0) { ggml_format_name(cur, "%s-%d", name, il); @@ -6381,127 +6384,79 @@ static struct ggml_cgraph * llama_build_graph( ggml_set_name(cur, name); } - if (!lctx.cparams.offload_kqv) { if (strcmp(name, "kqv_merged_cont") == 0) { // all nodes between the KV store and the attention output are run on the CPU ggml_backend_sched_set_node_backend(lctx.sched, cur, lctx.backend_cpu); } } - - // - // allocate input tensors and set input data - // - - if (!alloc_inp_tokens && strcmp(name, "inp_tokens") == 0) { - ggml_tallocr_alloc(lctx.alloc, cur); - - if (!ggml_tallocr_is_measure(lctx.alloc) && batch.token) { - const int64_t n_tokens = cur->ne[0]; - - ggml_backend_tensor_set(cur, batch.token, 0, n_tokens*ggml_element_size(cur)); - } - - alloc_inp_tokens = true; - } - - if (!alloc_inp_embd && strcmp(name, "inp_embd") == 0 && batch.embd) { - ggml_tallocr_alloc(lctx.alloc, cur); - - if (!ggml_tallocr_is_measure(lctx.alloc) && batch.embd) { - const int64_t n_embd = cur->ne[0]; - const int64_t n_tokens = cur->ne[1]; - - ggml_backend_tensor_set(cur, batch.embd, 0, n_tokens*n_embd*ggml_element_size(cur)); - } - - alloc_inp_embd = true; - } - - if (!alloc_inp_pos && strcmp(name, "inp_pos") == 0) { - ggml_tallocr_alloc(lctx.alloc, cur); - - if (!ggml_tallocr_is_measure(lctx.alloc) && batch.pos) { - const int64_t n_tokens = cur->ne[0]; - - static_assert(std::is_same::value, "llama_pos must be int32_t"); - ggml_backend_tensor_set(cur, batch.pos, 0, n_tokens*ggml_element_size(cur)); - } - - alloc_inp_pos = true; - } - - if (!alloc_inp_KQ_mask && strcmp(name, "KQ_mask") == 0) { - ggml_tallocr_alloc(lctx.alloc, cur); - - if (!ggml_tallocr_is_measure(lctx.alloc)) { - const int64_t n_kv = cur->ne[0]; - const int64_t n_tokens = cur->ne[1]; - - float * data; - if (ggml_backend_buffer_is_host(cur->buffer)) { - data = (float *) cur->data; - } else { - lctx.buf_copy.resize(ggml_nbytes(cur)); - data = (float *) lctx.buf_copy.data(); - } - - for (int h = 0; h < 1; ++h) { - for (int j = 0; j < n_tokens; ++j) { - const llama_pos pos = batch.pos[j]; - const llama_seq_id seq_id = batch.seq_id[j][0]; - - for (int i = 0; i < n_kv; ++i) { - float f; - if (!lctx.kv_self.cells[i].has_seq_id(seq_id) || lctx.kv_self.cells[i].pos > pos) { - f = -INFINITY; - } else { - f = 0; - } - data[h*(n_kv*n_tokens) + j*n_kv + i] = f; - } - } - } - - if (data != cur->data) { - ggml_backend_tensor_set(cur, data, 0, ggml_nbytes(cur)); - } - } - - alloc_inp_KQ_mask = true; - } - - if (!alloc_inp_K_shift && strcmp(name, "K_shift") == 0) { - ggml_tallocr_alloc(lctx.alloc, cur); - - if (!ggml_tallocr_is_measure(lctx.alloc)) { - const int64_t n_ctx = cur->ne[0]; - - int32_t * data; - if (ggml_backend_buffer_is_host(cur->buffer)) { - data = (int32_t *) cur->data; - } else { - lctx.buf_copy.resize(ggml_nbytes(cur)); - data = (int32_t *) lctx.buf_copy.data(); - } - - for (int i = 0; i < n_ctx; ++i) { - data[i] = lctx.kv_self.cells[i].delta; - } - - if (data != cur->data) { - ggml_backend_tensor_set(cur, data, 0, ggml_nbytes(cur)); - } - } - - alloc_inp_K_shift = true; - } }; struct ggml_cgraph * result = NULL; struct llm_build_context llm(lctx, batch, cb, worst_case); + // + // set input data + // + + if (!ggml_tallocr_is_measure(lctx.alloc)) { + if (batch.token) { + const int64_t n_tokens = batch.n_tokens; + + ggml_backend_tensor_set(lctx.inp_tokens, batch.token, 0, n_tokens*ggml_element_size(lctx.inp_tokens)); + } + + if (batch.embd) { + const int64_t n_embd = llm.n_embd; + const int64_t n_tokens = batch.n_tokens; + + ggml_backend_tensor_set(lctx.inp_embd, batch.embd, 0, n_tokens*n_embd*ggml_element_size(lctx.inp_embd)); + } + + if (batch.pos) { + const int64_t n_tokens = batch.n_tokens; + + ggml_backend_tensor_set(lctx.inp_pos, batch.pos, 0, n_tokens*ggml_element_size(lctx.inp_pos)); + } + + { + const int64_t n_kv = llm.n_kv; + const int64_t n_tokens = batch.n_tokens; + + GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_KQ_mask->buffer)); + float * data = (float *) lctx.inp_KQ_mask->data; + + for (int h = 0; h < 1; ++h) { + for (int j = 0; j < n_tokens; ++j) { + const llama_pos pos = batch.pos[j]; + const llama_seq_id seq_id = batch.seq_id[j][0]; + + for (int i = 0; i < n_kv; ++i) { + float f; + if (!lctx.kv_self.cells[i].has_seq_id(seq_id) || lctx.kv_self.cells[i].pos > pos) { + f = -INFINITY; + } else { + f = 0; + } + data[h*(n_kv*n_tokens) + j*n_kv + i] = f; + } + } + } + } + + if (llm.do_rope_shift) { + const int64_t n_ctx = llm.n_ctx; + + GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_K_shift->buffer)); + int32_t * data = (int32_t *) lctx.inp_K_shift->data; + + for (int i = 0; i < n_ctx; ++i) { + data[i] = lctx.kv_self.cells[i].delta; + } + } + } + llm.init(); switch (model.arch) { @@ -9964,6 +9919,35 @@ struct llama_context * llama_new_context_with_model( ctx->embedding.resize(hparams.n_embd); } + // graph inputs + { + ggml_init_params init_params = { + /* .mem_size */ ggml_tensor_overhead()*5, + /* .mem_buffer */ nullptr, + /* .no_alloc */ true, + }; + ctx->ctx_input = ggml_init(init_params); + + ctx->inp_tokens = ggml_new_tensor_1d(ctx->ctx_input, GGML_TYPE_I32, cparams.n_batch); + ctx->inp_embd = ggml_new_tensor_2d(ctx->ctx_input, GGML_TYPE_F32, hparams.n_embd, cparams.n_batch); + ctx->inp_pos = ggml_new_tensor_1d(ctx->ctx_input, GGML_TYPE_I32, cparams.n_batch); + ctx->inp_KQ_mask = ggml_new_tensor_2d(ctx->ctx_input, GGML_TYPE_F32, cparams.n_ctx, cparams.n_batch); + ctx->inp_K_shift = ggml_new_tensor_1d(ctx->ctx_input, GGML_TYPE_I32, cparams.n_ctx); + + ggml_set_name(ctx->inp_tokens, "inp_tokens"); + ggml_set_name(ctx->inp_embd, "inp_embd"); + ggml_set_name(ctx->inp_pos, "inp_pos"); + ggml_set_name(ctx->inp_KQ_mask, "inp_KQ_mask"); + ggml_set_name(ctx->inp_K_shift, "inp_K_shift"); + + ctx->buf_input = ggml_backend_alloc_ctx_tensors_from_buft(ctx->ctx_input, llama_default_buffer_type_cpu(true)); + + LLAMA_LOG_INFO("%s: %10s input buffer size = %8.2f MiB\n", __func__, + ggml_backend_buffer_name(ctx->buf_input), + ggml_backend_buffer_get_size(ctx->buf_input) / 1024.0 / 1024.0); + } + + // scheduler and compute buffers { // buffer types used for the compute buffer of each backend std::vector backend_buft; @@ -9990,9 +9974,6 @@ struct llama_context * llama_new_context_with_model( // initialize scheduler with the worst-case graph ggml_backend_sched_init_measure(ctx->sched, gf); - // note: the number of splits during measure is higher than during inference due to the kv shift - int n_splits = ggml_backend_sched_get_n_splits(ctx->sched); - LLAMA_LOG_INFO("%s: graph splits (measure): %d\n", __func__, n_splits); ctx->alloc = ggml_backend_sched_get_tallocr(ctx->sched, ctx->backend_cpu); for (ggml_backend_t backend : ctx->backends) { @@ -10001,6 +9982,10 @@ struct llama_context * llama_new_context_with_model( ggml_backend_buffer_name(buf), ggml_backend_buffer_get_size(buf) / 1024.0 / 1024.0); } + + // note: the number of splits during measure is higher than during inference due to the kv shift + int n_splits = ggml_backend_sched_get_n_splits(ctx->sched); + LLAMA_LOG_INFO("%s: graph splits (measure): %d\n", __func__, n_splits); } } From bf63d695b804b1c995c7ae4427a8a86936ea6d25 Mon Sep 17 00:00:00 2001 From: Michael Hueschen Date: Mon, 22 Jan 2024 03:17:05 -0700 Subject: [PATCH 36/66] nix: add cc to devShell LD_LIBRARY_PATH this fixes the error I encountered when trying to run the convert.py script in a venv: ``` $ nix develop [...]$ source .venv/bin/activate (.venv) [...]$ pip3 install -r requirements.txt <... clipped ...> [...]$ python3 ./convert.py Traceback (most recent call last): File "/home/mhueschen/projects-reference/llama.cpp/./convert.py", line 40, in from sentencepiece import SentencePieceProcessor File "/home/mhueschen/projects-reference/llama.cpp/.venv/lib/python3.11/site-packages/sentencepiece/__init__.py", line 13, in from . import _sentencepiece ImportError: libstdc++.so.6: cannot open shared object file: No such file or directory ``` however, I am not sure this is the cleanest way to address this linker issue... --- .devops/nix/package.nix | 6 ++++++ 1 file changed, 6 insertions(+) diff --git a/.devops/nix/package.nix b/.devops/nix/package.nix index c25d99f01..91ddb8890 100644 --- a/.devops/nix/package.nix +++ b/.devops/nix/package.nix @@ -221,10 +221,16 @@ effectiveStdenv.mkDerivation ( ; shell = mkShell { + NIX_LD_LIBRARY_PATH = lib.makeLibraryPath [ + effectiveStdenv.cc.cc + ]; name = "shell-${finalAttrs.finalPackage.name}"; description = "contains numpy and sentencepiece"; buildInputs = [ llama-python ]; inputsFrom = [ finalAttrs.finalPackage ]; + shellHook = '' + export LD_LIBRARY_PATH=$NIX_LD_LIBRARY_PATH + ''; }; shell-extra = mkShell { From c9b316c78fba31e65879a2ec91cbafd341b88cce Mon Sep 17 00:00:00 2001 From: Michael Hueschen Date: Mon, 22 Jan 2024 16:44:10 -0700 Subject: [PATCH 37/66] nix-shell: use addToSearchPath thx to @SomeoneSerge for the suggestion! --- .devops/nix/package.nix | 5 +---- 1 file changed, 1 insertion(+), 4 deletions(-) diff --git a/.devops/nix/package.nix b/.devops/nix/package.nix index 91ddb8890..a868a9a61 100644 --- a/.devops/nix/package.nix +++ b/.devops/nix/package.nix @@ -221,15 +221,12 @@ effectiveStdenv.mkDerivation ( ; shell = mkShell { - NIX_LD_LIBRARY_PATH = lib.makeLibraryPath [ - effectiveStdenv.cc.cc - ]; name = "shell-${finalAttrs.finalPackage.name}"; description = "contains numpy and sentencepiece"; buildInputs = [ llama-python ]; inputsFrom = [ finalAttrs.finalPackage ]; shellHook = '' - export LD_LIBRARY_PATH=$NIX_LD_LIBRARY_PATH + addToSearchPath "LD_LIBRARY_PATH" "${lib.getLib effectiveStdenv.cc.cc}/lib" ''; }; From cd4fddb29f81d6a1f6d51a0c016bc6b486d68def Mon Sep 17 00:00:00 2001 From: Engininja2 <139037756+Engininja2@users.noreply.github.com> Date: Wed, 24 Jan 2024 16:18:15 -0600 Subject: [PATCH 38/66] cuda : fix 2-bit quants on amd hip (#5105) * cuda : fix 2-bit quants on amd hip * use __low2float intrinsic function for new quants --- ggml-cuda.cu | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) diff --git a/ggml-cuda.cu b/ggml-cuda.cu index 7f460449e..05e5d18ab 100644 --- a/ggml-cuda.cu +++ b/ggml-cuda.cu @@ -4283,7 +4283,7 @@ static __device__ __forceinline__ float vec_dot_iq2_xxs_q8_1( q8 += 8; aux32 >>= 7; } - const float d = (float)bq2->d * (0.5f + aux32) * (float)bq8_1[ib32].ds.x * 0.25f; + const float d = (float)bq2->d * (0.5f + aux32) * __low2float(bq8_1[ib32].ds) * 0.25f; return d * sumi; #else // iqs is 0...15 @@ -4294,7 +4294,7 @@ static __device__ __forceinline__ float vec_dot_iq2_xxs_q8_1( const uint8_t * grid1 = (const uint8_t *)(iq2xxs_grid + aux8[2*il+0]); const uint8_t * grid2 = (const uint8_t *)(iq2xxs_grid + aux8[2*il+1]); const uint32_t aux32 = q2[2] | (q2[3] << 16); - const float d = (float)bq2->d * (0.5f + (aux32 >> 28)) * (float)bq8_1[ib32].ds.x * 0.25f; + const float d = (float)bq2->d * (0.5f + (aux32 >> 28)) * __low2float(bq8_1[ib32].ds) * 0.25f; const uint8_t signs1 = ksigns_iq2xs[(aux32 >> 14*il) & 127]; const uint8_t signs2 = ksigns_iq2xs[(aux32 >> (14*il + 7)) & 127]; const int8_t * q8 = bq8_1[ib32].qs + 16*il; @@ -4339,7 +4339,7 @@ static __device__ __forceinline__ float vec_dot_iq2_xs_q8_1( } q8 += 8; } - const float d = (float)bq2->d * (float)bq8_1[ib32].ds.x * 0.25f; + const float d = (float)bq2->d * __low2float(bq8_1[ib32].ds) * 0.25f; return d * ((0.5f + ls1) * sumi1 + (0.5f + ls2) * sumi2); #else assert(false); From ddc5a5033f948dc7ab0a3a6ec2d914d13c274077 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Thu, 25 Jan 2024 11:26:17 +0200 Subject: [PATCH 39/66] metal : show compile log messages --- ggml-metal.m | 13 ++++++++----- 1 file changed, 8 insertions(+), 5 deletions(-) diff --git a/ggml-metal.m b/ggml-metal.m index 4b3eb4914..60fef1a19 100644 --- a/ggml-metal.m +++ b/ggml-metal.m @@ -277,6 +277,10 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) { NSURL * libURL = [NSURL fileURLWithPath:libPath]; GGML_METAL_LOG_INFO("%s: loading '%s'\n", __func__, [libPath UTF8String]); ctx->library = [ctx->device newLibraryWithURL:libURL error:&error]; + if (error) { + GGML_METAL_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]); + return NULL; + } } else { GGML_METAL_LOG_INFO("%s: default.metallib not found, loading from source\n", __func__); @@ -315,13 +319,12 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) { //[options setFastMathEnabled:false]; ctx->library = [ctx->device newLibraryWithSource:src options:options error:&error]; + if (error) { + GGML_METAL_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]); + return NULL; + } } } - - if (error) { - GGML_METAL_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]); - return NULL; - } } // print MTL GPU family: From faa3526a1eba458120987ed8269e5616385a76f4 Mon Sep 17 00:00:00 2001 From: Kawrakow <48489457+ikawrakow@users.noreply.github.com> Date: Thu, 25 Jan 2024 17:58:53 +0200 Subject: [PATCH 40/66] Fix Q3_K_XS for MoE models (#5113) Co-authored-by: Iwan Kawrakow --- llama.cpp | 45 +++++++++++++++++++++++++-------------------- 1 file changed, 25 insertions(+), 20 deletions(-) diff --git a/llama.cpp b/llama.cpp index 114046db9..6a7506e85 100644 --- a/llama.cpp +++ b/llama.cpp @@ -8829,6 +8829,23 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty auto use_more_bits = [](int i_layer, int num_layers) -> bool { return i_layer < num_layers/8 || i_layer >= 7*num_layers/8 || (i_layer - num_layers/8)%3 == 2; }; + const int n_expert = std::max(1, (int)qs.model.hparams.n_expert); + auto layer_info = [n_expert] (int i_layer, int n_layer, const char * name) { + if (n_expert > 1) { + // Believe it or not, "experts" in the FFN of Mixtral-8x7B are not consecutive, but iccasionally randomly + // sprinkled in the model. Hence, simply dividing i_ffn_down by n_expert does not work + // for getting the current layer as I initially thought, and we need to resort to parsing the + // tensor name. + n_layer /= n_expert; + if (sscanf(name, "blk.%d.", &i_layer) != 1) { + throw std::runtime_error(format("Failed to determine layer for tensor %s", name)); + } + if (i_layer < 0 || i_layer >= n_layer) { + throw std::runtime_error(format("Bad layer %d for tensor %s. Must be in [0, %d)", i_layer, name, n_layer)); + } + } + return std::make_pair(i_layer, n_layer); + }; if (name == tn(LLM_TENSOR_OUTPUT, "weight")) { int nx = tensor->ne[0]; @@ -8890,24 +8907,8 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty new_type = GGML_TYPE_Q2_K; } } else if (name.find("ffn_down") != std::string::npos) { - const int n_expert = std::max(1, (int)qs.model.hparams.n_expert); - int i_layer, n_layer; - if (n_expert == 1) { - i_layer = qs.i_ffn_down; - n_layer = qs.n_ffn_down; - } else { - // Believe it or not, "experts" in the FFN of Mixtral-8x7B are not consecutive, but iccasionally randomly - // sprinkled in the model. Hence, simply dividing i_ffn_down by n_expert does not work - // for getting the current layer as I initially thought, and we need to resort to parsing the - // tensor name. - n_layer = qs.n_ffn_down / n_expert; - if (sscanf(name.c_str(), "blk.%d.ffn_down", &i_layer) != 1) { - throw std::runtime_error(format("Failed to determine layer for tensor %s", name.c_str())); - } - if (i_layer < 0 || i_layer >= n_layer) { - throw std::runtime_error(format("Bad layer %d for tensor %s. Must be in [0, %d)", i_layer, name.c_str(), n_layer)); - } - } + auto info = layer_info(qs.i_ffn_down, qs.n_ffn_down, name.c_str()); + int i_layer = info.first, n_layer = info.second; if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q3_K; else if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K_S || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_XS) { if (i_layer < n_layer/8) new_type = GGML_TYPE_Q4_K; @@ -8963,13 +8964,17 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty else if (ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M) new_type = GGML_TYPE_Q6_K; } else if (name.find("ffn_gate") != std::string::npos) { - if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_XS && !use_more_bits(qs.i_ffn_gate, qs.n_ffn_gate)) { + auto info = layer_info(qs.i_ffn_gate, qs.n_ffn_gate, name.c_str()); + int i_layer = info.first, n_layer = info.second; + if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_XS && !use_more_bits(i_layer, n_layer)) { new_type = GGML_TYPE_Q2_K; } ++qs.i_ffn_gate; } else if (name.find("ffn_up") != std::string::npos) { - if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_XS && !use_more_bits(qs.i_ffn_up, qs.n_ffn_up)) { + auto info = layer_info(qs.i_ffn_up, qs.n_ffn_up, name.c_str()); + int i_layer = info.first, n_layer = info.second; + if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_XS && !use_more_bits(i_layer, n_layer)) { new_type = GGML_TYPE_Q2_K; } ++qs.i_ffn_up; From 256d1bb0ddce6a0a21f5a7503019bdd5c1933cba Mon Sep 17 00:00:00 2001 From: Valentin Konovalov Date: Thu, 25 Jan 2024 12:05:51 -0500 Subject: [PATCH 41/66] android : use release cmake build type by default (#5123) --- examples/llama.android/app/build.gradle.kts | 1 + 1 file changed, 1 insertion(+) diff --git a/examples/llama.android/app/build.gradle.kts b/examples/llama.android/app/build.gradle.kts index 7815a8025..aadbe22c9 100644 --- a/examples/llama.android/app/build.gradle.kts +++ b/examples/llama.android/app/build.gradle.kts @@ -30,6 +30,7 @@ android { } externalNativeBuild { cmake { + arguments += "-DCMAKE_BUILD_TYPE=Release" cppFlags += listOf() arguments += listOf() } From d292f4f2047963f558dd516f1baaa71793e9acf2 Mon Sep 17 00:00:00 2001 From: Jared Van Bortel Date: Thu, 25 Jan 2024 14:51:24 -0500 Subject: [PATCH 42/66] examples : make pydantic scripts pass mypy and support py3.8 (#5099) --- .../pydantic-models-to-grammar-examples.py | 43 +---- examples/pydantic_models_to_grammar.py | 166 +++++++++--------- 2 files changed, 88 insertions(+), 121 deletions(-) diff --git a/examples/pydantic-models-to-grammar-examples.py b/examples/pydantic-models-to-grammar-examples.py index cbf376652..160966649 100644 --- a/examples/pydantic-models-to-grammar-examples.py +++ b/examples/pydantic-models-to-grammar-examples.py @@ -1,14 +1,14 @@ # Function calling example using pydantic models. import datetime +import importlib import json from enum import Enum -from typing import Union, Optional +from typing import Optional, Union import requests from pydantic import BaseModel, Field - -import importlib -from pydantic_models_to_grammar import generate_gbnf_grammar_and_documentation, convert_dictionary_to_pydantic_model, add_run_method_to_dynamic_model, create_dynamic_model_from_function +from pydantic_models_to_grammar import (add_run_method_to_dynamic_model, convert_dictionary_to_pydantic_model, + create_dynamic_model_from_function, generate_gbnf_grammar_and_documentation) # Function to get completion on the llama.cpp server with grammar. @@ -35,7 +35,7 @@ class SendMessageToUser(BaseModel): print(self.message) -# Enum for the calculator function. +# Enum for the calculator tool. class MathOperation(Enum): ADD = "add" SUBTRACT = "subtract" @@ -43,7 +43,7 @@ class MathOperation(Enum): DIVIDE = "divide" -# Very simple calculator tool for the agent. +# Simple pydantic calculator tool for the agent that can add, subtract, multiply, and divide. Docstring and description of fields will be used in system prompt. class Calculator(BaseModel): """ Perform a math operation on two numbers. @@ -148,37 +148,6 @@ def get_current_datetime(output_format: Optional[str] = None): return datetime.datetime.now().strftime(output_format) -# Enum for the calculator tool. -class MathOperation(Enum): - ADD = "add" - SUBTRACT = "subtract" - MULTIPLY = "multiply" - DIVIDE = "divide" - - - -# Simple pydantic calculator tool for the agent that can add, subtract, multiply, and divide. Docstring and description of fields will be used in system prompt. -class Calculator(BaseModel): - """ - Perform a math operation on two numbers. - """ - number_one: Union[int, float] = Field(..., description="First number.") - operation: MathOperation = Field(..., description="Math operation to perform.") - number_two: Union[int, float] = Field(..., description="Second number.") - - def run(self): - if self.operation == MathOperation.ADD: - return self.number_one + self.number_two - elif self.operation == MathOperation.SUBTRACT: - return self.number_one - self.number_two - elif self.operation == MathOperation.MULTIPLY: - return self.number_one * self.number_two - elif self.operation == MathOperation.DIVIDE: - return self.number_one / self.number_two - else: - raise ValueError("Unknown operation.") - - # Example function to get the weather def get_current_weather(location, unit): """Get the current weather in a given location""" diff --git a/examples/pydantic_models_to_grammar.py b/examples/pydantic_models_to_grammar.py index 848c1c367..9acc7cc6d 100644 --- a/examples/pydantic_models_to_grammar.py +++ b/examples/pydantic_models_to_grammar.py @@ -1,15 +1,21 @@ +from __future__ import annotations + import inspect import json +import re from copy import copy -from inspect import isclass, getdoc -from types import NoneType +from enum import Enum +from inspect import getdoc, isclass +from typing import TYPE_CHECKING, Any, Callable, List, Optional, Union, get_args, get_origin, get_type_hints from docstring_parser import parse -from pydantic import BaseModel, create_model, Field -from typing import Any, Type, List, get_args, get_origin, Tuple, Union, Optional, _GenericAlias -from enum import Enum -from typing import get_type_hints, Callable -import re +from pydantic import BaseModel, Field, create_model + +if TYPE_CHECKING: + from types import GenericAlias +else: + # python 3.8 compat + from typing import _GenericAlias as GenericAlias class PydanticDataType(Enum): @@ -43,7 +49,7 @@ class PydanticDataType(Enum): SET = "set" -def map_pydantic_type_to_gbnf(pydantic_type: Type[Any]) -> str: +def map_pydantic_type_to_gbnf(pydantic_type: type[Any]) -> str: if isclass(pydantic_type) and issubclass(pydantic_type, str): return PydanticDataType.STRING.value elif isclass(pydantic_type) and issubclass(pydantic_type, bool): @@ -57,22 +63,22 @@ def map_pydantic_type_to_gbnf(pydantic_type: Type[Any]) -> str: elif isclass(pydantic_type) and issubclass(pydantic_type, BaseModel): return format_model_and_field_name(pydantic_type.__name__) - elif get_origin(pydantic_type) == list: + elif get_origin(pydantic_type) is list: element_type = get_args(pydantic_type)[0] return f"{map_pydantic_type_to_gbnf(element_type)}-list" - elif get_origin(pydantic_type) == set: + elif get_origin(pydantic_type) is set: element_type = get_args(pydantic_type)[0] return f"{map_pydantic_type_to_gbnf(element_type)}-set" - elif get_origin(pydantic_type) == Union: + elif get_origin(pydantic_type) is Union: union_types = get_args(pydantic_type) union_rules = [map_pydantic_type_to_gbnf(ut) for ut in union_types] return f"union-{'-or-'.join(union_rules)}" - elif get_origin(pydantic_type) == Optional: + elif get_origin(pydantic_type) is Optional: element_type = get_args(pydantic_type)[0] return f"optional-{map_pydantic_type_to_gbnf(element_type)}" elif isclass(pydantic_type): return f"{PydanticDataType.CUSTOM_CLASS.value}-{format_model_and_field_name(pydantic_type.__name__)}" - elif get_origin(pydantic_type) == dict: + elif get_origin(pydantic_type) is dict: key_type, value_type = get_args(pydantic_type) return f"custom-dict-key-type-{format_model_and_field_name(map_pydantic_type_to_gbnf(key_type))}-value-type-{format_model_and_field_name(map_pydantic_type_to_gbnf(value_type))}" else: @@ -106,7 +112,6 @@ def get_members_structure(cls, rule_name): return f"{cls.__name__.lower()} ::= " + " | ".join(members) if cls.__annotations__ and cls.__annotations__ != {}: result = f'{rule_name} ::= "{{"' - type_list_rules = [] # Modify this comprehension members = [ f' "\\"{name}\\"" ":" {map_pydantic_type_to_gbnf(param_type)}' @@ -116,27 +121,25 @@ def get_members_structure(cls, rule_name): result += '"," '.join(members) result += ' "}"' - return result, type_list_rules - elif rule_name == "custom-class-any": + return result + if rule_name == "custom-class-any": result = f"{rule_name} ::= " result += "value" - type_list_rules = [] - return result, type_list_rules - else: - init_signature = inspect.signature(cls.__init__) - parameters = init_signature.parameters - result = f'{rule_name} ::= "{{"' - type_list_rules = [] - # Modify this comprehension too - members = [ - f' "\\"{name}\\"" ":" {map_pydantic_type_to_gbnf(param.annotation)}' - for name, param in parameters.items() - if name != "self" and param.annotation != inspect.Parameter.empty - ] + return result - result += '", "'.join(members) - result += ' "}"' - return result, type_list_rules + init_signature = inspect.signature(cls.__init__) + parameters = init_signature.parameters + result = f'{rule_name} ::= "{{"' + # Modify this comprehension too + members = [ + f' "\\"{name}\\"" ":" {map_pydantic_type_to_gbnf(param.annotation)}' + for name, param in parameters.items() + if name != "self" and param.annotation != inspect.Parameter.empty + ] + + result += '", "'.join(members) + result += ' "}"' + return result def regex_to_gbnf(regex_pattern: str) -> str: @@ -269,7 +272,7 @@ def generate_gbnf_float_rules(max_digit=None, min_digit=None, max_precision=None def generate_gbnf_rule_for_type( model_name, field_name, field_type, is_optional, processed_models, created_rules, field_info=None -) -> Tuple[str, list]: +) -> tuple[str, list[str]]: """ Generate GBNF rule for a given field type. @@ -283,7 +286,7 @@ def generate_gbnf_rule_for_type( :param field_info: Additional information about the field (optional). :return: Tuple containing the GBNF type and a list of additional rules. - :rtype: Tuple[str, list] + :rtype: tuple[str, list] """ rules = [] @@ -321,8 +324,7 @@ def generate_gbnf_rule_for_type( gbnf_type, rules = model_name + "-" + field_name, rules elif gbnf_type.startswith("custom-class-"): - nested_model_rules, field_types = get_members_structure(field_type, gbnf_type) - rules.append(nested_model_rules) + rules.append(get_members_structure(field_type, gbnf_type)) elif gbnf_type.startswith("custom-dict-"): key_type, value_type = get_args(field_type) @@ -341,14 +343,14 @@ def generate_gbnf_rule_for_type( union_rules = [] for union_type in union_types: - if isinstance(union_type, _GenericAlias): + if isinstance(union_type, GenericAlias): union_gbnf_type, union_rules_list = generate_gbnf_rule_for_type( model_name, field_name, union_type, False, processed_models, created_rules ) union_rules.append(union_gbnf_type) rules.extend(union_rules_list) - elif not issubclass(union_type, NoneType): + elif not issubclass(union_type, type(None)): union_gbnf_type, union_rules_list = generate_gbnf_rule_for_type( model_name, field_name, union_type, False, processed_models, created_rules ) @@ -424,14 +426,10 @@ def generate_gbnf_rule_for_type( else: gbnf_type, rules = gbnf_type, [] - if gbnf_type not in created_rules: - return gbnf_type, rules - else: - if gbnf_type in created_rules: - return gbnf_type, rules + return gbnf_type, rules -def generate_gbnf_grammar(model: Type[BaseModel], processed_models: set, created_rules: dict) -> (list, bool, bool): +def generate_gbnf_grammar(model: type[BaseModel], processed_models: set[type[BaseModel]], created_rules: dict[str, list[str]]) -> tuple[list[str], bool]: """ Generate GBnF Grammar @@ -452,7 +450,7 @@ def generate_gbnf_grammar(model: Type[BaseModel], processed_models: set, created ``` """ if model in processed_models: - return [] + return [], False processed_models.add(model) model_name = format_model_and_field_name(model.__name__) @@ -518,7 +516,7 @@ def generate_gbnf_grammar(model: Type[BaseModel], processed_models: set, created def generate_gbnf_grammar_from_pydantic_models( - models: List[Type[BaseModel]], outer_object_name: str = None, outer_object_content: str = None, + models: list[type[BaseModel]], outer_object_name: str | None = None, outer_object_content: str | None = None, list_of_outputs: bool = False ) -> str: """ @@ -528,7 +526,7 @@ def generate_gbnf_grammar_from_pydantic_models( * grammar. Args: - models (List[Type[BaseModel]]): A list of Pydantic models to generate the grammar from. + models (list[type[BaseModel]]): A list of Pydantic models to generate the grammar from. outer_object_name (str): Outer object name for the GBNF grammar. If None, no outer object will be generated. Eg. "function" for function calling. outer_object_content (str): Content for the outer rule in the GBNF grammar. Eg. "function_parameters" or "params" for function calling. list_of_outputs (str, optional): Allows a list of output objects @@ -543,9 +541,9 @@ def generate_gbnf_grammar_from_pydantic_models( # root ::= UserModel | PostModel # ... """ - processed_models = set() + processed_models: set[type[BaseModel]] = set() all_rules = [] - created_rules = {} + created_rules: dict[str, list[str]] = {} if outer_object_name is None: for model in models: model_rules, _ = generate_gbnf_grammar(model, processed_models, created_rules) @@ -608,7 +606,7 @@ def get_primitive_grammar(grammar): Returns: str: GBNF primitive grammar string. """ - type_list = [] + type_list: list[type[object]] = [] if "string-list" in grammar: type_list.append(str) if "boolean-list" in grammar: @@ -666,14 +664,14 @@ triple-quotes ::= "'''" """ def generate_markdown_documentation( - pydantic_models: List[Type[BaseModel]], model_prefix="Model", fields_prefix="Fields", + pydantic_models: list[type[BaseModel]], model_prefix="Model", fields_prefix="Fields", documentation_with_field_description=True ) -> str: """ Generate markdown documentation for a list of Pydantic models. Args: - pydantic_models (List[Type[BaseModel]]): List of Pydantic model classes. + pydantic_models (list[type[BaseModel]]): list of Pydantic model classes. model_prefix (str): Prefix for the model section. fields_prefix (str): Prefix for the fields section. documentation_with_field_description (bool): Include field descriptions in the documentation. @@ -731,7 +729,7 @@ def generate_markdown_documentation( def generate_field_markdown( - field_name: str, field_type: Type[Any], model: Type[BaseModel], depth=1, + field_name: str, field_type: type[Any], model: type[BaseModel], depth=1, documentation_with_field_description=True ) -> str: """ @@ -739,8 +737,8 @@ def generate_field_markdown( Args: field_name (str): Name of the field. - field_type (Type[Any]): Type of the field. - model (Type[BaseModel]): Pydantic model class. + field_type (type[Any]): Type of the field. + model (type[BaseModel]): Pydantic model class. depth (int): Indentation depth in the documentation. documentation_with_field_description (bool): Include field descriptions in the documentation. @@ -798,7 +796,7 @@ def generate_field_markdown( return field_text -def format_json_example(example: dict, depth: int) -> str: +def format_json_example(example: dict[str, Any], depth: int) -> str: """ Format a JSON example into a readable string with indentation. @@ -819,14 +817,14 @@ def format_json_example(example: dict, depth: int) -> str: def generate_text_documentation( - pydantic_models: List[Type[BaseModel]], model_prefix="Model", fields_prefix="Fields", + pydantic_models: list[type[BaseModel]], model_prefix="Model", fields_prefix="Fields", documentation_with_field_description=True ) -> str: """ Generate text documentation for a list of Pydantic models. Args: - pydantic_models (List[Type[BaseModel]]): List of Pydantic model classes. + pydantic_models (list[type[BaseModel]]): List of Pydantic model classes. model_prefix (str): Prefix for the model section. fields_prefix (str): Prefix for the fields section. documentation_with_field_description (bool): Include field descriptions in the documentation. @@ -885,7 +883,7 @@ def generate_text_documentation( def generate_field_text( - field_name: str, field_type: Type[Any], model: Type[BaseModel], depth=1, + field_name: str, field_type: type[Any], model: type[BaseModel], depth=1, documentation_with_field_description=True ) -> str: """ @@ -893,8 +891,8 @@ def generate_field_text( Args: field_name (str): Name of the field. - field_type (Type[Any]): Type of the field. - model (Type[BaseModel]): Pydantic model class. + field_type (type[Any]): Type of the field. + model (type[BaseModel]): Pydantic model class. depth (int): Indentation depth in the documentation. documentation_with_field_description (bool): Include field descriptions in the documentation. @@ -1017,8 +1015,8 @@ def generate_and_save_gbnf_grammar_and_documentation( pydantic_model_list, grammar_file_path="./generated_grammar.gbnf", documentation_file_path="./generated_grammar_documentation.md", - outer_object_name: str = None, - outer_object_content: str = None, + outer_object_name: str | None = None, + outer_object_content: str | None = None, model_prefix: str = "Output Model", fields_prefix: str = "Output Fields", list_of_outputs: bool = False, @@ -1053,8 +1051,8 @@ def generate_and_save_gbnf_grammar_and_documentation( def generate_gbnf_grammar_and_documentation( pydantic_model_list, - outer_object_name: str = None, - outer_object_content: str = None, + outer_object_name: str | None = None, + outer_object_content: str | None = None, model_prefix: str = "Output Model", fields_prefix: str = "Output Fields", list_of_outputs: bool = False, @@ -1086,9 +1084,9 @@ def generate_gbnf_grammar_and_documentation( def generate_gbnf_grammar_and_documentation_from_dictionaries( - dictionaries: List[dict], - outer_object_name: str = None, - outer_object_content: str = None, + dictionaries: list[dict[str, Any]], + outer_object_name: str | None = None, + outer_object_content: str | None = None, model_prefix: str = "Output Model", fields_prefix: str = "Output Fields", list_of_outputs: bool = False, @@ -1098,7 +1096,7 @@ def generate_gbnf_grammar_and_documentation_from_dictionaries( Generate GBNF grammar and documentation from a list of dictionaries. Args: - dictionaries (List[dict]): List of dictionaries representing Pydantic models. + dictionaries (list[dict]): List of dictionaries representing Pydantic models. outer_object_name (str): Outer object name for the GBNF grammar. If None, no outer object will be generated. Eg. "function" for function calling. outer_object_content (str): Content for the outer rule in the GBNF grammar. Eg. "function_parameters" or "params" for function calling. model_prefix (str): Prefix for the model section in the documentation. @@ -1120,7 +1118,7 @@ def generate_gbnf_grammar_and_documentation_from_dictionaries( return grammar, documentation -def create_dynamic_model_from_function(func: Callable): +def create_dynamic_model_from_function(func: Callable[..., Any]): """ Creates a dynamic Pydantic model from a given function's type hints and adds the function as a 'run' method. @@ -1135,6 +1133,7 @@ def create_dynamic_model_from_function(func: Callable): sig = inspect.signature(func) # Parse the docstring + assert func.__doc__ is not None docstring = parse(func.__doc__) dynamic_fields = {} @@ -1157,7 +1156,6 @@ def create_dynamic_model_from_function(func: Callable): f"Parameter '{param.name}' in function '{func.__name__}' lacks a description in the docstring") # Add parameter details to the schema - param_doc = next((d for d in docstring.params if d.arg_name == param.name), None) param_docs.append((param.name, param_doc)) if param.default == inspect.Parameter.empty: default_value = ... @@ -1166,10 +1164,10 @@ def create_dynamic_model_from_function(func: Callable): dynamic_fields[param.name] = ( param.annotation if param.annotation != inspect.Parameter.empty else str, default_value) # Creating the dynamic model - dynamic_model = create_model(f"{func.__name__}", **dynamic_fields) + dynamic_model = create_model(f"{func.__name__}", **dynamic_fields) # type: ignore[call-overload] - for param_doc in param_docs: - dynamic_model.model_fields[param_doc[0]].description = param_doc[1].description + for name, param_doc in param_docs: + dynamic_model.model_fields[name].description = param_doc.description dynamic_model.__doc__ = docstring.short_description @@ -1182,16 +1180,16 @@ def create_dynamic_model_from_function(func: Callable): return dynamic_model -def add_run_method_to_dynamic_model(model: Type[BaseModel], func: Callable): +def add_run_method_to_dynamic_model(model: type[BaseModel], func: Callable[..., Any]): """ Add a 'run' method to a dynamic Pydantic model, using the provided function. Args: - model (Type[BaseModel]): Dynamic Pydantic model class. + model (type[BaseModel]): Dynamic Pydantic model class. func (Callable): Function to be added as a 'run' method to the model. Returns: - Type[BaseModel]: Pydantic model class with the added 'run' method. + type[BaseModel]: Pydantic model class with the added 'run' method. """ def run_method_wrapper(self): @@ -1204,15 +1202,15 @@ def add_run_method_to_dynamic_model(model: Type[BaseModel], func: Callable): return model -def create_dynamic_models_from_dictionaries(dictionaries: List[dict]): +def create_dynamic_models_from_dictionaries(dictionaries: list[dict[str, Any]]): """ Create a list of dynamic Pydantic model classes from a list of dictionaries. Args: - dictionaries (List[dict]): List of dictionaries representing model structures. + dictionaries (list[dict]): List of dictionaries representing model structures. Returns: - List[Type[BaseModel]]: List of generated dynamic Pydantic model classes. + list[type[BaseModel]]: List of generated dynamic Pydantic model classes. """ dynamic_models = [] for func in dictionaries: @@ -1249,7 +1247,7 @@ def list_to_enum(enum_name, values): return Enum(enum_name, {value: value for value in values}) -def convert_dictionary_to_pydantic_model(dictionary: dict, model_name: str = "CustomModel") -> Type[BaseModel]: +def convert_dictionary_to_pydantic_model(dictionary: dict[str, Any], model_name: str = "CustomModel") -> type[Any]: """ Convert a dictionary to a Pydantic model class. @@ -1258,9 +1256,9 @@ def convert_dictionary_to_pydantic_model(dictionary: dict, model_name: str = "Cu model_name (str): Name of the generated Pydantic model. Returns: - Type[BaseModel]: Generated Pydantic model class. + type[BaseModel]: Generated Pydantic model class. """ - fields = {} + fields: dict[str, Any] = {} if "properties" in dictionary: for field_name, field_data in dictionary.get("properties", {}).items(): @@ -1277,7 +1275,7 @@ def convert_dictionary_to_pydantic_model(dictionary: dict, model_name: str = "Cu if items != {}: array = {"properties": items} array_type = convert_dictionary_to_pydantic_model(array, f"{model_name}_{field_name}_items") - fields[field_name] = (List[array_type], ...) + fields[field_name] = (List[array_type], ...) # type: ignore[valid-type] else: fields[field_name] = (list, ...) elif field_type == "object": From 5eaf9964fc797d4585c214db32a463d557f3ed33 Mon Sep 17 00:00:00 2001 From: l3utterfly Date: Fri, 26 Jan 2024 05:06:22 +0900 Subject: [PATCH 43/66] llama : dynamic temperature sampling (#4972) * implemented dynamic temperature sampling from koboldcpp * removed trailing whitespace * removed unused temp parameter in llama_sample_entropy * exposed exponent_val in dynamic temp sampler * added debug check for printf statements * use nullptr in llama_sample_softmax call during llama_sample_entropy this avoids counting the time taken stats twice Co-authored-by: Georgi Gerganov * return earlier if there is only 1 candiate (i.e. max_entropy == 0) * reformat 't' case in llama_sample_queue Co-authored-by: Jared Van Bortel * check for one or zero candidates case in llama_sample_entropy --------- Co-authored-by: Georgi Gerganov Co-authored-by: Jared Van Bortel --- common/sampling.cpp | 12 +++++++- common/sampling.h | 2 ++ llama.cpp | 67 +++++++++++++++++++++++++++++++++++++++++++++ llama.h | 8 ++++++ 4 files changed, 88 insertions(+), 1 deletion(-) diff --git a/common/sampling.cpp b/common/sampling.cpp index dd1ffeb1b..efd7eab6e 100644 --- a/common/sampling.cpp +++ b/common/sampling.cpp @@ -129,6 +129,8 @@ static void sampler_queue( const int n_vocab = llama_n_vocab(llama_get_model(ctx_main)); const float temp = params.temp; + const float dynatemp_range = params.dynatemp_range; + const float dynatemp_exponent = params.dynatemp_exponent; const int32_t top_k = params.top_k <= 0 ? n_vocab : params.top_k; const float top_p = params.top_p; const float min_p = params.min_p; @@ -143,7 +145,15 @@ static void sampler_queue( case 'y': llama_sample_typical (ctx_main, &cur_p, typical_p, min_keep); break; case 'p': llama_sample_top_p (ctx_main, &cur_p, top_p, min_keep); break; case 'm': llama_sample_min_p (ctx_main, &cur_p, min_p, min_keep); break; - case 't': llama_sample_temp (ctx_main, &cur_p, temp); break; + case 't': + if (dynatemp_range > 0) { + float dynatemp_min = std::max(0.0f, temp - dynatemp_range); + float dynatemp_max = std::max(0.0f, temp + dynatemp_range); + llama_sample_entropy(ctx_main, &cur_p, dynatemp_min, dynatemp_max, dynatemp_exponent); + } else { + llama_sample_temp(ctx_main, &cur_p, temp); + } + break; default : break; } } diff --git a/common/sampling.h b/common/sampling.h index 2ee180376..88899c094 100644 --- a/common/sampling.h +++ b/common/sampling.h @@ -18,6 +18,8 @@ typedef struct llama_sampling_params { float tfs_z = 1.00f; // 1.0 = disabled float typical_p = 1.00f; // 1.0 = disabled float temp = 0.80f; // <= 0.0 to sample greedily, 0.0 to not output probabilities + float dynatemp_range = 0.00f; // 0.0 = disabled + float dynatemp_exponent = 1.00f; // controls how entropy maps to temperature in dynamic temperature sampler int32_t penalty_last_n = 64; // last n tokens to penalize (0 = disable penalty, -1 = context size) float penalty_repeat = 1.10f; // 1.0 = disabled float penalty_freq = 0.00f; // 0.0 = disabled diff --git a/llama.cpp b/llama.cpp index 6a7506e85..823d42d7f 100644 --- a/llama.cpp +++ b/llama.cpp @@ -8151,6 +8151,73 @@ void llama_sample_typical(struct llama_context * ctx, llama_token_data_array * c } } +void llama_sample_entropy(struct llama_context * ctx, llama_token_data_array * candidates_p, float min_temp, float max_temp, float exponent_val) { + const int64_t t_start_sample_us = ggml_time_us(); + + // no need to do anything if there is only one (or zero) candidates + if(candidates_p->size <= 1) { + return; + } + + // Calculate maximum possible entropy + float max_entropy = -logf(1.0f / candidates_p->size); + + llama_sample_softmax(nullptr, candidates_p); + + // Calculate entropy of the softmax probabilities + float entropy = 0.0f; + for (size_t i = 0; i < candidates_p->size; ++i) { + float prob = candidates_p->data[i].p; + if (prob > 0.0f) { // Ensure no log(0) + entropy -= prob * logf(prob); + } + } + + // Normalize the entropy (max_entropy cannot be 0 here because we checked candidates_p->size != 1 above) + float normalized_entropy = entropy / max_entropy; + + // Map the normalized entropy to the desired temperature range using the power function + float dyn_temp = min_temp + (max_temp - min_temp) * powf(normalized_entropy, exponent_val); + +#ifdef DEBUG + LLAMA_LOG_INFO("Your text maxtemp value is: %f\n", max_temp); + LLAMA_LOG_INFO("Entropy: %f\n", entropy); + LLAMA_LOG_INFO("Max Possible Entropy: %f\n", max_entropy); + LLAMA_LOG_INFO("Normalized Entropy: %f\n", normalized_entropy); + LLAMA_LOG_INFO("Exponent: %f\n", exponent_val); + LLAMA_LOG_INFO("Dynamic Temperature (dyn_temp): %f\n", dyn_temp); +#endif + + // Apply the dynamically calculated temperature scaling + for (size_t i = 0; i < candidates_p->size; ++i) { + candidates_p->data[i].logit /= dyn_temp; + } + + // Re-compute softmax probabilities after scaling logits with dynamic temperature + double max_l_double = candidates_p->data[0].logit; + double cum_sum_double = 0.0; + for (size_t i = 0; i < candidates_p->size; ++i) { + double p = exp(candidates_p->data[i].logit - max_l_double); + candidates_p->data[i].p = p; // Store the scaled probability + cum_sum_double += p; + } + for (size_t i = 0; i < candidates_p->size; ++i) { + candidates_p->data[i].p /= cum_sum_double; // Re-normalize the probabilities + } + +#ifdef DEBUG + // Print the updated top 25 probabilities after temperature scaling + LLAMA_LOG_INFO("\nUpdated Top 25 Probabilities After Dynamic Temperature Scaling (in percentages):\n"); + for (size_t i = 0; i < 25 && i < candidates_p->size; ++i) { + LLAMA_LOG_INFO("Token %zu: %f%%\n", i + 1, candidates_p->data[i].p * 100.0f); + } +#endif + + if (ctx) { + ctx->t_sample_us += ggml_time_us() - t_start_sample_us; + } +} + void llama_sample_temp(struct llama_context * ctx, llama_token_data_array * candidates_p, float temp) { const int64_t t_start_sample_us = ggml_time_us(); diff --git a/llama.h b/llama.h index bb6054557..7b3634aa6 100644 --- a/llama.h +++ b/llama.h @@ -775,6 +775,14 @@ extern "C" { float p, size_t min_keep); + /// @details Dynamic temperature implementation described in the paper https://arxiv.org/abs/2309.02772. + LLAMA_API void llama_sample_entropy( + struct llama_context * ctx, + llama_token_data_array * candidates_p, + float min_temp, + float max_temp, + float exponent_val); + LLAMA_API void llama_sample_temp( struct llama_context * ctx, llama_token_data_array * candidates, From fe54033b69b83164cabb5f3ed92dc0ff7ea47605 Mon Sep 17 00:00:00 2001 From: XiaotaoChen Date: Fri, 26 Jan 2024 04:14:32 +0800 Subject: [PATCH 44/66] readme : add MobileVLM 1.7B/3B to the supported models list (#5107) Co-authored-by: Chenxiaotao03 --- README.md | 1 + 1 file changed, 1 insertion(+) diff --git a/README.md b/README.md index cbfba01bc..c2a33022c 100644 --- a/README.md +++ b/README.md @@ -112,6 +112,7 @@ as the main playground for developing new features for the [ggml](https://github - [x] [Bakllava](https://huggingface.co/models?search=SkunkworksAI/Bakllava) - [x] [Obsidian](https://huggingface.co/NousResearch/Obsidian-3B-V0.5) - [x] [ShareGPT4V](https://huggingface.co/models?search=Lin-Chen/ShareGPT4V) +- [x] [MobileVLM 1.7B/3B models](https://huggingface.co/models?search=mobileVLM) **Bindings:** From 1182cf4d4f6ee383b92695c2e3fe438086dcdba7 Mon Sep 17 00:00:00 2001 From: Kawrakow <48489457+ikawrakow@users.noreply.github.com> Date: Fri, 26 Jan 2024 09:14:39 +0200 Subject: [PATCH 45/66] Another bucket sort (#5109) * Initial bucket sort * Bucket sort: slightly better version * Bucket sort: another minor improvement --------- Co-authored-by: Iwan Kawrakow --- llama.cpp | 53 ++++++++++++++++++++++++++++++++++++++++++++++++++--- 1 file changed, 50 insertions(+), 3 deletions(-) diff --git a/llama.cpp b/llama.cpp index 823d42d7f..b03b67e16 100644 --- a/llama.cpp +++ b/llama.cpp @@ -7956,10 +7956,57 @@ void llama_sample_top_k(struct llama_context * ctx, llama_token_data_array * can auto comp = [](const llama_token_data & a, const llama_token_data & b) { return a.logit > b.logit; }; - if (k == (int) candidates->size) { - std::sort(candidates->data, candidates->data + candidates->size, comp); - } else { + if (k <= 128) { std::partial_sort(candidates->data, candidates->data + k, candidates->data + candidates->size, comp); + } else { + constexpr int nbuckets = 128; + constexpr float bucket_low = -10.0f; + constexpr float bucket_high = 10.0f; + constexpr float bucket_scale = nbuckets/(bucket_high - bucket_low); + constexpr float bucker_inter = -bucket_low * bucket_scale; + + std::vector bucket_idx(candidates->size); + std::vector histo(nbuckets, 0); + + for (int i = 0; i < (int)candidates->size; ++i) { + const float val = candidates->data[i].logit; + int ib = int(bucket_scale * val + bucker_inter); //nbuckets * (val - bucket_low) / (bucket_high - bucket_low); + ib = std::max(0, std::min(nbuckets-1, ib)); + bucket_idx[i] = ib; + ++histo[ib]; + } + int nhave = 0; + int ib = nbuckets - 1; + for ( ; ib >= 0; --ib) { + nhave += histo[ib]; + if (nhave >= k) break; + } + std::vector tmp_tokens(nhave); + auto ptr = tmp_tokens.data(); + std::vector bucket_ptrs; + bucket_ptrs.reserve(nbuckets - ib); + for (int j = nbuckets - 1; j >= ib; --j) { + bucket_ptrs.push_back(ptr); + ptr += histo[j]; + } + for (int i = 0; i < (int)candidates->size; ++i) { + int j = bucket_idx[i]; + if (j >= ib) { + *bucket_ptrs[nbuckets-1-j]++ = candidates->data[i]; + } + } + + ptr = tmp_tokens.data(); + int ndone = 0; + for (int j = nbuckets-1; j > ib; --j) { + std::sort(ptr, ptr + histo[j], comp); + ptr += histo[j]; + ndone += histo[j]; + } + std::partial_sort(ptr, ptr + k - ndone, ptr + histo[ib], comp); + + std::memcpy(candidates->data, tmp_tokens.data(), k*sizeof(llama_token_data)); + } candidates->sorted = true; } From aad0b01d7380a7cdfe0dd42307b18c7b6bac9575 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Fri, 26 Jan 2024 10:52:33 +0200 Subject: [PATCH 46/66] readme : update hot topics --- README.md | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/README.md b/README.md index c2a33022c..76e48ce8a 100644 --- a/README.md +++ b/README.md @@ -10,11 +10,11 @@ Inference of [LLaMA](https://arxiv.org/abs/2302.13971) model in pure C/C++ ### Hot topics +- ⚠️ Incoming backends: https://github.com/ggerganov/llama.cpp/discussions/5138 - New SOTA quantized models, including pure 2-bits: https://huggingface.co/ikawrakow - Collecting Apple Silicon performance stats: - M-series: https://github.com/ggerganov/llama.cpp/discussions/4167 - A-series: https://github.com/ggerganov/llama.cpp/discussions/4508 -- Added Mixtral support: https://github.com/ggerganov/llama.cpp/pull/4406 - Looking for contributions to improve and maintain the `server` example: https://github.com/ggerganov/llama.cpp/issues/4216 ---- From 38b431de232d1b736b5af19b8c7d72f7075a70bc Mon Sep 17 00:00:00 2001 From: Riceball LEE Date: Fri, 26 Jan 2024 17:10:28 +0800 Subject: [PATCH 47/66] gguf : fix "general.alignment" type in gguf_reader.py (#5136) --- gguf-py/gguf/gguf_reader.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/gguf-py/gguf/gguf_reader.py b/gguf-py/gguf/gguf_reader.py index 8682765ed..5b6d4ba6b 100644 --- a/gguf-py/gguf/gguf_reader.py +++ b/gguf-py/gguf/gguf_reader.py @@ -107,7 +107,7 @@ class GGUFReader: offs, tensors_fields = self._build_tensors_fields(offs, tensor_count) new_align = self.fields.get('general.alignment') if new_align is not None: - if new_align.types != [GGUFValueType.UINT64]: + if new_align.types != [GGUFValueType.UINT32]: raise ValueError('Bad type for general.alignment field') self.alignment = new_align.parts[-1][0] padding = offs % self.alignment From 6dd3c28c9cd1ef74b49d79f47d668759346a3c6c Mon Sep 17 00:00:00 2001 From: Paul Tsochantaris Date: Fri, 26 Jan 2024 12:16:07 +0000 Subject: [PATCH 48/66] metal : remove unused `n_buffers` and `buffers` (#5129) --- ggml-metal.m | 73 ++++++++++++---------------------------------------- 1 file changed, 16 insertions(+), 57 deletions(-) diff --git a/ggml-metal.m b/ggml-metal.m index 60fef1a19..ab3c84f7f 100644 --- a/ggml-metal.m +++ b/ggml-metal.m @@ -26,15 +26,6 @@ #define GGML_METAL_MAX_KERNELS 256 -struct ggml_metal_buffer { - const char * name; - - void * data; - size_t size; - - id metal; -}; - struct ggml_metal_kernel { id function; id pipeline; @@ -172,9 +163,6 @@ struct ggml_metal_context { dispatch_queue_t d_queue; - int n_buffers; - struct ggml_metal_buffer buffers[GGML_METAL_MAX_BUFFERS]; - struct ggml_metal_kernel kernels[GGML_METAL_MAX_KERNELS]; bool support_simdgroup_reduction; @@ -242,24 +230,20 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) { // Show all the Metal device instances in the system NSArray * devices = MTLCopyAllDevices(); for (id device in devices) { - NSString * s = [device name]; - GGML_METAL_LOG_INFO("%s: found device: %s\n", __func__, [s UTF8String]); + 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 // Pick and show default Metal device id device = MTLCreateSystemDefaultDevice(); - NSString * s = [device name]; - GGML_METAL_LOG_INFO("%s: picking default device: %s\n", __func__, [s UTF8String]); + 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); ctx->queue = [ctx->device newCommandQueue]; - ctx->n_buffers = 0; - ctx->d_queue = dispatch_queue_create("ggml-metal", DISPATCH_QUEUE_CONCURRENT); // load library @@ -534,10 +518,6 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) { static void ggml_metal_free(struct ggml_metal_context * ctx) { GGML_METAL_LOG_INFO("%s: deallocating\n", __func__); - for (int i = 0; i < ctx->n_buffers; ++i) { - [ctx->buffers[i].metal release]; - } - for (int i = 0; i < GGML_METAL_MAX_KERNELS; ++i) { if (ctx->kernels[i].pipeline) { [ctx->kernels[i].pipeline release]; @@ -580,51 +560,30 @@ struct ggml_backend_metal_buffer_context { // 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 ggml_metal_get_buffer(struct ggml_metal_context * ctx, struct ggml_tensor * t, size_t * offs) { +static id 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); const int64_t tsize = ggml_nbytes(t); ggml_backend_buffer_t buffer = t->view_src ? t->view_src->buffer : t->buffer; - // compatibility with ggml-backend - if (buffer && buffer->buft == ggml_backend_metal_buffer_type()) { - 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; - - //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) { - *offs = (size_t) ioffs; - - //GGML_METAL_LOG_INFO("%s: tensor '%16s', offs = %8ld\n", __func__, t->name, *offs); - - return buf_ctx->buffers[i].metal; - } - } - - GGML_METAL_LOG_ERROR("%s: error: tensor '%s' buffer is nil\n", __func__, t->name); - - return nil; - } + 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 < ctx->n_buffers; ++i) { - const int64_t ioffs = (int64_t) t->data - (int64_t) ctx->buffers[i].data; + 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; - //GGML_METAL_LOG_INFO("ioffs = %10ld, tsize = %10ld, sum = %10ld, ctx->buffers[%d].size = %10ld, name = %s\n", ioffs, tsize, ioffs + tsize, i, ctx->buffers[i].size, ctx->buffers[i].name); - if (ioffs >= 0 && ioffs + tsize <= (int64_t) ctx->buffers[i].size) { + //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) { *offs = (size_t) ioffs; - //GGML_METAL_LOG_INFO("%s: '%s' tensor '%16s', offs = %8ld\n", __func__, ctx->buffers[i].name, t->name, *offs); + //GGML_METAL_LOG_INFO("%s: tensor '%16s', offs = %8ld\n", __func__, t->name, *offs); - return ctx->buffers[i].metal; + return buf_ctx->buffers[i].metal; } } - GGML_METAL_LOG_ERROR("%s: error: buffer is nil\n", __func__); + GGML_METAL_LOG_ERROR("%s: error: tensor '%s' buffer is nil\n", __func__, t->name); return nil; } @@ -817,9 +776,9 @@ static bool ggml_metal_graph_compute( const enum ggml_type src1t = src1 ? src1->type : GGML_TYPE_COUNT; const enum ggml_type dstt = dst ? dst->type : GGML_TYPE_COUNT; - id id_src0 = src0 ? ggml_metal_get_buffer(ctx, src0, &offs_src0) : nil; - id id_src1 = src1 ? ggml_metal_get_buffer(ctx, src1, &offs_src1) : nil; - id id_dst = dst ? ggml_metal_get_buffer(ctx, dst, &offs_dst) : nil; + id id_src0 = src0 ? ggml_metal_get_buffer(src0, &offs_src0) : nil; + id id_src1 = src1 ? ggml_metal_get_buffer(src1, &offs_src1) : nil; + id 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) { @@ -1601,7 +1560,7 @@ static bool ggml_metal_graph_compute( struct ggml_tensor * src_cur = dst->src[2 + (j % n_as)]; size_t offs_src_cur = 0; - id id_src_cur = ggml_metal_get_buffer(ctx, src_cur, &offs_src_cur); + id id_src_cur = ggml_metal_get_buffer(src_cur, &offs_src_cur); [encoder setBuffer:id_src_cur offset:offs_src_cur atIndex:19 + j]; } @@ -1746,7 +1705,7 @@ static bool ggml_metal_graph_compute( struct ggml_tensor * src_cur = dst->src[2 + (j % n_as)]; size_t offs_src_cur = 0; - id id_src_cur = ggml_metal_get_buffer(ctx, src_cur, &offs_src_cur); + id id_src_cur = ggml_metal_get_buffer(src_cur, &offs_src_cur); [encoder setBuffer:id_src_cur offset:offs_src_cur atIndex:23 + j]; } From 413e7b0559f922bd4de5e9eec548829d111651b1 Mon Sep 17 00:00:00 2001 From: crasm Date: Fri, 26 Jan 2024 07:18:00 -0500 Subject: [PATCH 49/66] ci : add model tests + script wrapper (#4586) * scripts : add lib.sh and lib_test.sh * scripts : stub out new ci-run.sh script * scripts : switch to PascalCase for functions This looks a little odd at first, but I find it very useful as a convention to know if a command is part of our code vs a builtin. * scripts : add some fancy conversion from snake_case to PascalCase * Add venv to ci/run.sh * Revert scripts work * scripts : add wrapper script for local use of ci/run.sh * Simplify .gitignore for tests, clang-tidy fixes * Label all ctest tests * ci : ctest uses -L main * Attempt at writing ctest_with_model * Update test-model-load-cancel * ci : add ctest_with_model for debug and release ggml-ci * Fix gg_get_model function ggml-ci * got stuck on CMake * Add get_model.cpp to tests/CMakeLists.txt ggml-ci * Fix README.md output for ctest_with_model ggml-ci * workflows : use `-L main` for all ctest ggml-ci * Fixes * GG_RUN_CTEST_MODELFILE => LLAMACPP_TESTMODELFILE * Always show warning rather than failing if model file variable is not set * scripts : update usage text for ci-run.sh --- .github/workflows/build.yml | 12 ++--- .gitignore | 19 +------- Makefile | 7 ++- ci/run.sh | 81 +++++++++++++++++++++++++++----- scripts/ci-run.sh | 50 ++++++++++++++++++++ tests/.gitignore | 2 + tests/CMakeLists.txt | 14 ++++-- tests/get-model.cpp | 21 +++++++++ tests/get-model.h | 2 + tests/test-autorelease.cpp | 12 ++--- tests/test-model-load-cancel.cpp | 27 +++++++++++ 11 files changed, 199 insertions(+), 48 deletions(-) create mode 100755 scripts/ci-run.sh create mode 100644 tests/.gitignore create mode 100644 tests/get-model.cpp create mode 100644 tests/get-model.h create mode 100644 tests/test-model-load-cancel.cpp diff --git a/.github/workflows/build.yml b/.github/workflows/build.yml index c3aa6f992..d22a041a6 100644 --- a/.github/workflows/build.yml +++ b/.github/workflows/build.yml @@ -72,7 +72,7 @@ jobs: id: cmake_test run: | cd build - ctest --verbose --timeout 900 + ctest -L main --verbose --timeout 900 ubuntu-latest-cmake-sanitizer: runs-on: ubuntu-latest @@ -107,7 +107,7 @@ jobs: id: cmake_test run: | cd build - ctest --verbose --timeout 900 + ctest -L main --verbose --timeout 900 ubuntu-latest-cmake-mpi: runs-on: ubuntu-latest @@ -141,7 +141,7 @@ jobs: id: cmake_test run: | cd build - ctest --verbose + ctest -L main --verbose # TODO: build with LLAMA_NO_METAL because test-backend-ops fail on "Apple Paravirtual device" and I don't know # how to debug it. @@ -202,7 +202,7 @@ jobs: id: cmake_test run: | cd build - ctest --verbose --timeout 900 + ctest -L main --verbose --timeout 900 macOS-latest-cmake-ios: runs-on: macos-latest @@ -394,7 +394,7 @@ jobs: if: ${{ matrix.build != 'clblast' && (matrix.build != 'avx512' || env.HAS_AVX512F == '1') }} # not all machines have native AVX-512 run: | cd build - ctest -C Release --verbose --timeout 900 + ctest -L main -C Release --verbose --timeout 900 - name: Test (Intel SDE) id: cmake_test_sde @@ -406,7 +406,7 @@ jobs: 7z x "-o${env:RUNNER_TEMP}" $env:RUNNER_TEMP/sde.tar $sde = $(join-path $env:RUNNER_TEMP sde-external-${env:SDE_VERSION}-win/sde.exe) cd build - & $sde -future -- ctest -C Release --verbose --timeout 900 + & $sde -future -- ctest -L main -C Release --verbose --timeout 900 - name: Determine tag name id: tag diff --git a/.gitignore b/.gitignore index 5ab81445d..cb0069bfb 100644 --- a/.gitignore +++ b/.gitignore @@ -27,7 +27,7 @@ lcov-report/ gcovr-report/ -build*/ +build* out/ tmp/ @@ -89,20 +89,3 @@ examples/jeopardy/results.txt poetry.lock poetry.toml - -# Test binaries -/tests/test-grammar-parser -/tests/test-llama-grammar -/tests/test-double-float -/tests/test-grad0 -/tests/test-opt -/tests/test-quantize-fns -/tests/test-quantize-perf -/tests/test-sampling -/tests/test-tokenizer-0-llama -/tests/test-tokenizer-0-falcon -/tests/test-tokenizer-1-llama -/tests/test-tokenizer-1-bpe -/tests/test-rope -/tests/test-backend-ops -/tests/test-autorelease diff --git a/Makefile b/Makefile index a8658a596..82c89e87a 100644 --- a/Makefile +++ b/Makefile @@ -9,7 +9,7 @@ TEST_TARGETS = \ tests/test-llama-grammar tests/test-grammar-parser tests/test-double-float tests/test-grad0 tests/test-opt \ tests/test-quantize-fns tests/test-quantize-perf tests/test-sampling tests/test-tokenizer-0-llama \ tests/test-tokenizer-0-falcon tests/test-tokenizer-1-llama tests/test-tokenizer-1-bpe tests/test-rope \ - tests/test-backend-ops tests/test-autorelease + tests/test-backend-ops tests/test-model-load-cancel tests/test-autorelease # Code coverage output files COV_TARGETS = *.gcno tests/*.gcno *.gcda tests/*.gcda *.gcov tests/*.gcov lcov-report gcovr-report @@ -748,5 +748,8 @@ tests/test-c.o: tests/test-c.c llama.h tests/test-backend-ops: tests/test-backend-ops.cpp ggml.o $(OBJS) $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) -tests/test-autorelease: tests/test-autorelease.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS) +tests/test-model-load-cancel: tests/test-model-load-cancel.cpp ggml.o llama.o tests/get-model.cpp $(COMMON_DEPS) $(OBJS) + $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) + +tests/test-autorelease: tests/test-autorelease.cpp ggml.o llama.o tests/get-model.cpp $(COMMON_DEPS) $(OBJS) $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) diff --git a/ci/run.sh b/ci/run.sh index 791b17a19..2427e55a2 100755 --- a/ci/run.sh +++ b/ci/run.sh @@ -22,9 +22,9 @@ mkdir -p "$2" OUT=$(realpath "$1") MNT=$(realpath "$2") -rm -v $OUT/*.log -rm -v $OUT/*.exit -rm -v $OUT/*.md +rm -f "$OUT/*.log" +rm -f "$OUT/*.exit" +rm -f "$OUT/*.md" sd=`dirname $0` cd $sd/../ @@ -94,7 +94,7 @@ function gg_run_ctest_debug { (time cmake -DCMAKE_BUILD_TYPE=Debug ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log (time make -j ) 2>&1 | tee -a $OUT/${ci}-make.log - (time ctest --output-on-failure -E test-opt ) 2>&1 | tee -a $OUT/${ci}-ctest.log + (time ctest --output-on-failure -L main -E test-opt ) 2>&1 | tee -a $OUT/${ci}-ctest.log set +e } @@ -123,9 +123,9 @@ function gg_run_ctest_release { (time make -j ) 2>&1 | tee -a $OUT/${ci}-make.log if [ -z ${GG_BUILD_LOW_PERF} ]; then - (time ctest --output-on-failure ) 2>&1 | tee -a $OUT/${ci}-ctest.log + (time ctest --output-on-failure -L main ) 2>&1 | tee -a $OUT/${ci}-ctest.log else - (time ctest --output-on-failure -E test-opt ) 2>&1 | tee -a $OUT/${ci}-ctest.log + (time ctest --output-on-failure -L main -E test-opt ) 2>&1 | tee -a $OUT/${ci}-ctest.log fi set +e @@ -141,6 +141,61 @@ function gg_sum_ctest_release { gg_printf '```\n' } +function gg_get_model { + local gguf_3b="$MNT/models/open-llama/3B-v2/ggml-model-f16.gguf" + local gguf_7b="$MNT/models/open-llama/7B-v2/ggml-model-f16.gguf" + if [[ -s $gguf_3b ]]; then + echo -n "$gguf_3b" + elif [[ -s $gguf_7b ]]; then + echo -n "$gguf_7b" + else + echo >&2 "No model found. Can't run gg_run_ctest_with_model." + exit 1 + fi +} + +function gg_run_ctest_with_model_debug { + cd ${SRC} + + local model; model=$(gg_get_model) + cd build-ci-debug + set -e + (LLAMACPP_TEST_MODELFILE="$model" time ctest --output-on-failure -L model) 2>&1 | tee -a $OUT/${ci}-ctest.log + set +e + cd .. +} + +function gg_run_ctest_with_model_release { + cd ${SRC} + + local model; model=$(gg_get_model) + cd build-ci-release + set -e + (LLAMACPP_TEST_MODELFILE="$model" time ctest --output-on-failure -L model) 2>&1 | tee -a $OUT/${ci}-ctest.log + set +e + cd .. +} + +function gg_sum_ctest_with_model_debug { + gg_printf '### %s\n\n' "${ci}" + + gg_printf 'Runs ctest with model files in debug mode\n' + gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)" + gg_printf '```\n' + gg_printf '%s\n' "$(cat $OUT/${ci}-ctest.log)" + gg_printf '```\n' +} + +function gg_sum_ctest_with_model_release { + gg_printf '### %s\n\n' "${ci}" + + gg_printf 'Runs ctest with model files in release mode\n' + gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)" + gg_printf '```\n' + gg_printf '%s\n' "$(cat $OUT/${ci}-ctest.log)" + gg_printf '```\n' +} + # open_llama_3b_v2 function gg_run_open_llama_3b_v2 { @@ -183,8 +238,6 @@ function gg_run_open_llama_3b_v2 { wiki_test_60="${path_wiki}/wiki.test-60.raw" - ./bin/test-autorelease ${model_f16} - ./bin/quantize ${model_f16} ${model_q8_0} q8_0 ./bin/quantize ${model_f16} ${model_q4_0} q4_0 ./bin/quantize ${model_f16} ${model_q4_1} q4_1 @@ -507,14 +560,18 @@ function gg_sum_open_llama_7b_v2 { ## main if [ -z ${GG_BUILD_LOW_PERF} ]; then + # Create symlink: ./llama.cpp/models-mnt -> $MNT/models/models-mnt rm -rf ${SRC}/models-mnt - mnt_models=${MNT}/models mkdir -p ${mnt_models} ln -sfn ${mnt_models} ${SRC}/models-mnt - python3 -m pip install -r ${SRC}/requirements.txt - python3 -m pip install --editable gguf-py + # Create a fresh python3 venv and enter it + python3 -m venv "$MNT/venv" + source "$MNT/venv/bin/activate" + + pip install -r ${SRC}/requirements.txt --disable-pip-version-check + pip install --editable gguf-py --disable-pip-version-check fi ret=0 @@ -529,6 +586,8 @@ if [ -z ${GG_BUILD_LOW_PERF} ]; then else test $ret -eq 0 && gg_run open_llama_7b_v2 fi + test $ret -eq 0 && gg_run ctest_with_model_debug + test $ret -eq 0 && gg_run ctest_with_model_release fi fi diff --git a/scripts/ci-run.sh b/scripts/ci-run.sh new file mode 100755 index 000000000..06b5d9c6e --- /dev/null +++ b/scripts/ci-run.sh @@ -0,0 +1,50 @@ +#!/bin/bash +set -euo pipefail +this=$(realpath "$0"); readonly this +cd "$(dirname "$this")" +shellcheck "$this" + +if (( $# != 1 && $# != 2 )); then + cat >&2 <<'EOF' +usage: + ci-run.sh [] + +This script wraps ci/run.sh: +* If is a ramdisk, you can reduce writes to your SSD. If is not a ramdisk, keep in mind that total writes will increase by the size of . + (openllama_3b_v2: quantized models are about 30GB) +* Persistent model and data files are synced to and from , + excluding generated .gguf files. + (openllama_3b_v2: persistent files are about 6.6GB) +* defaults to ~/.cache/llama.cpp +EOF + exit 1 +fi + +cd .. # => llama.cpp repo root + +tmp="$1" +mkdir -p "$tmp" +tmp=$(realpath "$tmp") +echo >&2 "Using tmp=$tmp" + +cache="${2-$HOME/.cache/llama.cpp}" +mkdir -p "$cache" +cache=$(realpath "$cache") +echo >&2 "Using cache=$cache" + +_sync() { + local from="$1"; shift + local to="$1"; shift + + echo >&2 "Syncing from $from to $to" + mkdir -p "$from" "$to" + rsync -a "$from" "$to" --delete-during "$@" +} + +_sync "$(realpath .)/" "$tmp/llama.cpp" +_sync "$cache/ci-mnt/models/" "$tmp/llama.cpp/ci-mnt/models/" + +cd "$tmp/llama.cpp" +bash ci/run.sh ci-out ci-mnt + +_sync 'ci-mnt/models/' "$cache/ci-mnt/models/" --exclude='*.gguf' -P diff --git a/tests/.gitignore b/tests/.gitignore new file mode 100644 index 000000000..59be43b99 --- /dev/null +++ b/tests/.gitignore @@ -0,0 +1,2 @@ +* +!*.* diff --git a/tests/CMakeLists.txt b/tests/CMakeLists.txt index d7aaab843..3e40a78cd 100644 --- a/tests/CMakeLists.txt +++ b/tests/CMakeLists.txt @@ -1,6 +1,6 @@ function(llama_build_executable source) get_filename_component(TEST_TARGET ${source} NAME_WE) - add_executable(${TEST_TARGET} ${source}) + add_executable(${TEST_TARGET} ${source} get-model.cpp) install(TARGETS ${TEST_TARGET} RUNTIME) target_link_libraries(${TEST_TARGET} PRIVATE common) endfunction() @@ -8,14 +8,20 @@ endfunction() function(llama_test_executable name source) get_filename_component(TEST_TARGET ${source} NAME_WE) add_test(NAME ${name} COMMAND $ ${ARGN}) + set_property(TEST ${name} PROPERTY LABELS "main") endfunction() function(llama_build_and_test_executable source) + llama_build_and_test_executable_with_label(${source} "main") +endfunction() + +function(llama_build_and_test_executable_with_label source label) get_filename_component(TEST_TARGET ${source} NAME_WE) - add_executable(${TEST_TARGET} ${source}) + add_executable(${TEST_TARGET} ${source} get-model.cpp) install(TARGETS ${TEST_TARGET} RUNTIME) target_link_libraries(${TEST_TARGET} PRIVATE common) add_test(NAME ${TEST_TARGET} COMMAND $ ${ARGN}) + set_property(TEST ${TEST_TARGET} PROPERTY LABELS ${label}) endfunction() # llama_build_and_test_executable(test-double-float.cpp) # SLOW @@ -49,10 +55,12 @@ llama_build_and_test_executable(test-llama-grammar.cpp) llama_build_and_test_executable(test-grad0.cpp) # llama_build_and_test_executable(test-opt.cpp) # SLOW llama_build_and_test_executable(test-backend-ops.cpp) -llama_build_and_test_executable(test-autorelease.cpp) llama_build_and_test_executable(test-rope.cpp) +llama_build_and_test_executable_with_label(test-model-load-cancel.cpp "model") +llama_build_and_test_executable_with_label(test-autorelease.cpp "model") + # dummy executable - not installed get_filename_component(TEST_TARGET test-c.c NAME_WE) add_executable(${TEST_TARGET} test-c.c) diff --git a/tests/get-model.cpp b/tests/get-model.cpp new file mode 100644 index 000000000..4edb685f0 --- /dev/null +++ b/tests/get-model.cpp @@ -0,0 +1,21 @@ +#include +#include +#include + +#include "get-model.h" + +char * get_model_or_exit(int argc, char *argv[]) { + char * model_path; + if (argc > 1) { + model_path = argv[1]; + + } else { + model_path = getenv("LLAMACPP_TEST_MODELFILE"); + if (!model_path || strlen(model_path) == 0) { + fprintf(stderr, "\033[33mWARNING: No model file provided. Skipping this test. Set LLAMACPP_TEST_MODELFILE= to silence this warning and run this test.\n\033[0m"); + exit(EXIT_SUCCESS); + } + } + + return model_path; +} diff --git a/tests/get-model.h b/tests/get-model.h new file mode 100644 index 000000000..81a3a0fef --- /dev/null +++ b/tests/get-model.h @@ -0,0 +1,2 @@ +#pragma once +char * get_model_or_exit(int, char*[]); diff --git a/tests/test-autorelease.cpp b/tests/test-autorelease.cpp index 289c6ba6c..36a23c0bb 100644 --- a/tests/test-autorelease.cpp +++ b/tests/test-autorelease.cpp @@ -5,19 +5,15 @@ #include #include "llama.h" +#include "get-model.h" // This creates a new context inside a pthread and then tries to exit cleanly. int main(int argc, char ** argv) { - if (argc < 2) { - printf("Usage: %s model.gguf\n", argv[0]); - return 0; // intentionally return success - } + auto * model_path = get_model_or_exit(argc, argv); - const std::string fname = argv[1]; - - std::thread([&fname]() { + std::thread([&model_path]() { llama_backend_init(false); - auto * model = llama_load_model_from_file(fname.c_str(), llama_model_default_params()); + auto * model = llama_load_model_from_file(model_path, llama_model_default_params()); auto * ctx = llama_new_context_with_model(model, llama_context_default_params()); llama_free(ctx); llama_free_model(model); diff --git a/tests/test-model-load-cancel.cpp b/tests/test-model-load-cancel.cpp new file mode 100644 index 000000000..7ea4bbacc --- /dev/null +++ b/tests/test-model-load-cancel.cpp @@ -0,0 +1,27 @@ +#include "llama.h" +#include "get-model.h" + +#include + +int main(int argc, char *argv[] ) { + auto * model_path = get_model_or_exit(argc, argv); + auto * file = fopen(model_path, "r"); + if (file == nullptr) { + fprintf(stderr, "no model at '%s' found\n", model_path); + return EXIT_FAILURE; + } + + fprintf(stderr, "using '%s'\n", model_path); + fclose(file); + + llama_backend_init(false); + auto params = llama_model_params{}; + params.use_mmap = false; + params.progress_callback = [](float progress, void * ctx){ + (void) ctx; + return progress > 0.50; + }; + auto * model = llama_load_model_from_file(model_path, params); + llama_backend_free(); + return model == nullptr ? EXIT_SUCCESS : EXIT_FAILURE; +} From 48c857aa10aea73210a4a72da3f1a6f99269e75d Mon Sep 17 00:00:00 2001 From: Xuan Son Nguyen Date: Fri, 26 Jan 2024 13:42:20 +0100 Subject: [PATCH 50/66] server : refactored the task processing logic (#5065) * server: add llama_server_queue struct * server: add llama_server_response_event * server: add comments * server: move all mutexes away from server.cpp * server: correct multitask response * server: only add back deferred tasks when one slot is available * server: fix a race condition cause by "request_completion" --- Makefile | 2 +- examples/server/CMakeLists.txt | 2 +- examples/server/oai.hpp | 208 ++++++++ examples/server/server.cpp | 849 ++++++--------------------------- examples/server/utils.hpp | 507 ++++++++++++++++++++ 5 files changed, 876 insertions(+), 692 deletions(-) create mode 100644 examples/server/oai.hpp create mode 100644 examples/server/utils.hpp diff --git a/Makefile b/Makefile index 82c89e87a..b8858b412 100644 --- a/Makefile +++ b/Makefile @@ -619,7 +619,7 @@ embedding: examples/embedding/embedding.cpp ggml.o llama.o $(C save-load-state: examples/save-load-state/save-load-state.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS) $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) -server: examples/server/server.cpp examples/server/httplib.h examples/server/json.hpp examples/server/index.html.hpp examples/server/index.js.hpp examples/server/completion.js.hpp examples/llava/clip.cpp examples/llava/clip.h common/stb_image.h ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS) +server: examples/server/server.cpp examples/server/oai.hpp examples/server/utils.hpp examples/server/httplib.h examples/server/json.hpp examples/server/index.html.hpp examples/server/index.js.hpp examples/server/completion.js.hpp examples/llava/clip.cpp examples/llava/clip.h common/stb_image.h ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS) $(CXX) $(CXXFLAGS) -Iexamples/server $(filter-out %.h,$(filter-out %.hpp,$^)) -o $@ $(LDFLAGS) $(LWINSOCK2) -Wno-cast-qual gguf: examples/gguf/gguf.cpp ggml.o $(OBJS) diff --git a/examples/server/CMakeLists.txt b/examples/server/CMakeLists.txt index 81709e448..cc13b2d63 100644 --- a/examples/server/CMakeLists.txt +++ b/examples/server/CMakeLists.txt @@ -1,7 +1,7 @@ set(TARGET server) option(LLAMA_SERVER_VERBOSE "Build verbose logging option for Server" ON) include_directories(${CMAKE_CURRENT_SOURCE_DIR}) -add_executable(${TARGET} server.cpp json.hpp httplib.h) +add_executable(${TARGET} server.cpp oai.hpp utils.hpp json.hpp httplib.h) install(TARGETS ${TARGET} RUNTIME) target_compile_definitions(${TARGET} PRIVATE SERVER_VERBOSE=$ diff --git a/examples/server/oai.hpp b/examples/server/oai.hpp new file mode 100644 index 000000000..bc5db6eef --- /dev/null +++ b/examples/server/oai.hpp @@ -0,0 +1,208 @@ +#pragma once + +#include +#include +#include +#include +#include +#include + +#include "json.hpp" +#include "utils.hpp" + +#define DEFAULT_OAICOMPAT_MODEL "gpt-3.5-turbo-0613" + +using json = nlohmann::json; + +inline static json oaicompat_completion_params_parse( + const json &body /* openai api json semantics */) +{ + json llama_params; + + llama_params["__oaicompat"] = true; + + // Map OpenAI parameters to llama.cpp parameters + // + // For parameters that are defined by the OpenAI documentation (e.g. + // temperature), we explicitly specify OpenAI's intended default; we + // need to do that because sometimes OpenAI disagrees with llama.cpp + // + // https://platform.openai.com/docs/api-reference/chat/create + llama_sampling_params default_sparams; + llama_params["model"] = json_value(body, "model", std::string("unknown")); + llama_params["prompt"] = format_chatml(body["messages"]); // OpenAI 'messages' to llama.cpp 'prompt' + llama_params["cache_prompt"] = json_value(body, "cache_prompt", false); + llama_params["temperature"] = json_value(body, "temperature", 0.0); + llama_params["top_k"] = json_value(body, "top_k", default_sparams.top_k); + llama_params["top_p"] = json_value(body, "top_p", 1.0); + llama_params["n_predict"] = json_value(body, "max_tokens", -1); + llama_params["logit_bias"] = json_value(body, "logit_bias",json::object()); + llama_params["frequency_penalty"] = json_value(body, "frequency_penalty", 0.0); + llama_params["presence_penalty"] = json_value(body, "presence_penalty", 0.0); + llama_params["seed"] = json_value(body, "seed", LLAMA_DEFAULT_SEED); + llama_params["stream"] = json_value(body, "stream", false); + llama_params["mirostat"] = json_value(body, "mirostat", default_sparams.mirostat); + llama_params["mirostat_tau"] = json_value(body, "mirostat_tau", default_sparams.mirostat_tau); + llama_params["mirostat_eta"] = json_value(body, "mirostat_eta", default_sparams.mirostat_eta); + llama_params["penalize_nl"] = json_value(body, "penalize_nl", default_sparams.penalize_nl); + llama_params["typical_p"] = json_value(body, "typical_p", default_sparams.typical_p); + llama_params["repeat_last_n"] = json_value(body, "repeat_last_n", default_sparams.penalty_last_n); + llama_params["ignore_eos"] = json_value(body, "ignore_eos", false); + llama_params["tfs_z"] = json_value(body, "tfs_z", default_sparams.tfs_z); + + if (body.count("grammar") != 0) { + llama_params["grammar"] = json_value(body, "grammar", json::object()); + } + + // Handle 'stop' field + if (body.contains("stop") && body["stop"].is_string()) { + llama_params["stop"] = json::array({body["stop"].get()}); + } else { + llama_params["stop"] = json_value(body, "stop", json::array()); + } + + // Ensure there is ChatML-specific end sequence among stop words + llama_params["stop"].push_back("<|im_end|>"); + + return llama_params; +} + +inline static json format_final_response_oaicompat(const json &request, const task_result &response, bool streaming = false) +{ + json result = response.result_json; + + bool stopped_word = result.count("stopped_word") != 0; + bool stopped_eos = json_value(result, "stopped_eos", false); + int num_tokens_predicted = json_value(result, "tokens_predicted", 0); + int num_prompt_tokens = json_value(result, "tokens_evaluated", 0); + std::string content = json_value(result, "content", std::string("")); + + std::string finish_reason = "length"; + if (stopped_word || stopped_eos) { + finish_reason = "stop"; + } + + json choices = + streaming ? json::array({json{{"finish_reason", finish_reason}, + {"index", 0}, + {"delta", json::object()}}}) + : json::array({json{{"finish_reason", finish_reason}, + {"index", 0}, + {"message", json{{"content", content}, + {"role", "assistant"}}}}}); + + std::time_t t = std::time(0); + + json res = + json{{"choices", choices}, + {"created", t}, + {"model", + json_value(request, "model", std::string(DEFAULT_OAICOMPAT_MODEL))}, + {"object", streaming ? "chat.completion.chunk" : "chat.completion"}, + {"usage", + json{{"completion_tokens", num_tokens_predicted}, + {"prompt_tokens", num_prompt_tokens}, + {"total_tokens", num_tokens_predicted + num_prompt_tokens}}}, + {"id", gen_chatcmplid()}}; + + if (server_verbose) { + res["__verbose"] = result; + } + + if (result.contains("completion_probabilities")) { + res["completion_probabilities"] = json_value(result, "completion_probabilities", json::array()); + } + + return res; +} + +// return value is vector as there is one case where we might need to generate two responses +inline static std::vector format_partial_response_oaicompat(const task_result &response) { + json result = response.result_json; + + if (!result.contains("model") || !result.contains("oaicompat_token_ctr")) { + return std::vector({response.result_json}); + } + + bool first = json_value(result, "oaicompat_token_ctr", 0) == 0; + std::string modelname = json_value(result, "model", std::string(DEFAULT_OAICOMPAT_MODEL)); + + bool stopped_word = json_value(result, "stopped_word", false); + bool stopped_eos = json_value(result, "stopped_eos", false); + bool stopped_limit = json_value(result, "stopped_limit", false); + std::string content = json_value(result, "content", std::string("")); + + std::string finish_reason; + if (stopped_word || stopped_eos) { + finish_reason = "stop"; + } + if (stopped_limit) { + finish_reason = "length"; + } + + std::time_t t = std::time(0); + + json choices; + + if (!finish_reason.empty()) { + choices = json::array({json{{"finish_reason", finish_reason}, + {"index", 0}, + {"delta", json::object()}}}); + } else { + if (first) { + if (content.empty()) { + choices = json::array({json{{"finish_reason", nullptr}, + {"index", 0}, + {"delta", json{{"role", "assistant"}}}}}); + } else { + // We have to send this as two updates to conform to openai behavior + json initial_ret = json{{"choices", json::array({json{ + {"finish_reason", nullptr}, + {"index", 0}, + {"delta", json{ + {"role", "assistant"} + }}}})}, + {"created", t}, + {"id", gen_chatcmplid()}, + {"model", modelname}, + {"object", "chat.completion.chunk"}}; + + json second_ret = json{ + {"choices", json::array({json{{"finish_reason", nullptr}, + {"index", 0}, + {"delta", json{ + {"content", content}}} + }})}, + {"created", t}, + {"id", gen_chatcmplid()}, + {"model", modelname}, + {"object", "chat.completion.chunk"}}; + + return std::vector({initial_ret, second_ret}); + } + } else { + // Some idiosyncrasy in task processing logic makes several trailing calls + // with empty content, we ignore these at the calee site. + if (content.empty()) { + return std::vector({json::object()}); + } + + choices = json::array({json{ + {"finish_reason", nullptr}, + {"index", 0}, + {"delta", + json{ + {"content", content}, + }}, + }}); + } + } + + json ret = json{{"choices", choices}, + {"created", t}, + {"id", gen_chatcmplid()}, + {"model", modelname}, + {"object", "chat.completion.chunk"}}; + + return std::vector({ret}); +} diff --git a/examples/server/server.cpp b/examples/server/server.cpp index 0462fbd24..392836132 100644 --- a/examples/server/server.cpp +++ b/examples/server/server.cpp @@ -1,6 +1,8 @@ #include "common.h" #include "llama.h" #include "grammar-parser.h" +#include "utils.hpp" +#include "oai.hpp" #include "../llava/clip.h" @@ -23,17 +25,10 @@ #include #include -#include #include #include #include -#ifndef SERVER_VERBOSE -#define SERVER_VERBOSE 1 -#endif - -#define DEFAULT_OAICOMPAT_MODEL "gpt-3.5-turbo-0613" - using json = nlohmann::json; struct server_params @@ -46,197 +41,7 @@ struct server_params int32_t write_timeout = 600; }; -static bool server_verbose = false; - -#if SERVER_VERBOSE != 1 -#define LOG_VERBOSE(MSG, ...) -#else -#define LOG_VERBOSE(MSG, ...) \ - do \ - { \ - if (server_verbose) \ - { \ - server_log("VERBOSE", __func__, __LINE__, MSG, __VA_ARGS__); \ - } \ - } while (0) -#endif - -#define LOG_ERROR( MSG, ...) server_log("ERROR", __func__, __LINE__, MSG, __VA_ARGS__) -#define LOG_WARNING(MSG, ...) server_log("WARNING", __func__, __LINE__, MSG, __VA_ARGS__) -#define LOG_INFO( MSG, ...) server_log("INFO", __func__, __LINE__, MSG, __VA_ARGS__) - -json oaicompat_completion_params_parse(const json &body); -std::string format_chatml(std::vector messages); - - -// -// base64 utils (TODO: move to common in the future) -// - -static const std::string base64_chars = - "ABCDEFGHIJKLMNOPQRSTUVWXYZ" - "abcdefghijklmnopqrstuvwxyz" - "0123456789+/"; - -static inline bool is_base64(uint8_t c) -{ - return (isalnum(c) || (c == '+') || (c == '/')); -} - -static std::vector base64_decode(const std::string & encoded_string) -{ - int i = 0; - int j = 0; - int in_ = 0; - - int in_len = encoded_string.size(); - - uint8_t char_array_4[4]; - uint8_t char_array_3[3]; - - std::vector ret; - - while (in_len-- && (encoded_string[in_] != '=') && is_base64(encoded_string[in_])) - { - char_array_4[i++] = encoded_string[in_]; in_++; - if (i == 4) - { - for (i = 0; i <4; i++) - { - char_array_4[i] = base64_chars.find(char_array_4[i]); - } - - char_array_3[0] = ((char_array_4[0] ) << 2) + ((char_array_4[1] & 0x30) >> 4); - char_array_3[1] = ((char_array_4[1] & 0xf) << 4) + ((char_array_4[2] & 0x3c) >> 2); - char_array_3[2] = ((char_array_4[2] & 0x3) << 6) + char_array_4[3]; - - for (i = 0; (i < 3); i++) - { - ret.push_back(char_array_3[i]); - } - i = 0; - } - } - - if (i) - { - for (j = i; j <4; j++) - { - char_array_4[j] = 0; - } - - for (j = 0; j <4; j++) - { - char_array_4[j] = base64_chars.find(char_array_4[j]); - } - - char_array_3[0] = ((char_array_4[0] ) << 2) + ((char_array_4[1] & 0x30) >> 4); - char_array_3[1] = ((char_array_4[1] & 0xf) << 4) + ((char_array_4[2] & 0x3c) >> 2); - char_array_3[2] = ((char_array_4[2] & 0x3) << 6) + char_array_4[3]; - - for (j = 0; (j < i - 1); j++) - { - ret.push_back(char_array_3[j]); - } - } - - return ret; -} - -// -// parallel -// - -enum server_state { - SERVER_STATE_LOADING_MODEL, // Server is starting up, model not fully loaded yet - SERVER_STATE_READY, // Server is ready and model is loaded - SERVER_STATE_ERROR // An error occurred, load_model failed -}; - -enum task_type { - TASK_TYPE_COMPLETION, - TASK_TYPE_CANCEL, -}; - -struct task_server { - int id; - int target_id; - task_type type; - json data; - bool infill_mode = false; - bool embedding_mode = false; - int multitask_id = -1; -}; - -struct task_result { - int id; - int multitask_id = -1; - bool stop; - bool error; - json result_json; -}; - -struct task_multi { - int id; - std::set subtasks_remaining{}; - std::vector results{}; -}; - -// TODO: can become bool if we can't find use of more states -enum slot_state -{ - IDLE, - PROCESSING, -}; - -enum slot_command -{ - NONE, - LOAD_PROMPT, - RELEASE, -}; - -struct slot_params -{ - bool stream = true; - bool cache_prompt = false; // remember the prompt to avoid reprocessing all prompt - - uint32_t seed = -1; // RNG seed - int32_t n_keep = 0; // number of tokens to keep from initial prompt - int32_t n_predict = -1; // new tokens to predict - - std::vector antiprompt; - - json input_prefix; - json input_suffix; -}; - -struct slot_image -{ - int32_t id; - - bool request_encode_image = false; - float * image_embedding = nullptr; - int32_t image_tokens = 0; - - clip_image_u8 * img_data; - - std::string prefix_prompt; // before of this image -}; - -// completion token output with probabilities -struct completion_token_output -{ - struct token_prob - { - llama_token tok; - float prob; - }; - - std::vector probs; - llama_token tok; - std::string text_to_send; -}; +bool server_verbose = false; static size_t common_part(const std::vector &a, const std::vector &b) { @@ -292,28 +97,6 @@ static std::string tokens_to_str(llama_context *ctx, Iter begin, Iter end) return ret; } -static void server_log(const char *level, const char *function, int line, - const char *message, const nlohmann::ordered_json &extra) -{ - nlohmann::ordered_json log - { - {"timestamp", time(nullptr)}, - {"level", level}, - {"function", function}, - {"line", line}, - {"message", message}, - }; - - if (!extra.empty()) - { - log.merge_patch(extra); - } - - const std::string str = log.dump(-1, ' ', false, json::error_handler_t::replace); - printf("%.*s\n", (int)str.size(), str.data()); - fflush(stdout); -} - // format incomplete utf-8 multibyte character for output static std::string tokens_to_output_formatted_string(const llama_context *ctx, const llama_token token) { @@ -355,15 +138,6 @@ static json probs_vector_to_json(const llama_context *ctx, const std::vector -static T json_value(const json &body, const std::string &key, const T &default_value) -{ - // Fallback null to default value - return body.contains(key) && !body.at(key).is_null() - ? body.value(key, default_value) - : default_value; -} - struct llama_client_slot { int id; @@ -491,7 +265,7 @@ struct llama_client_slot } void release() { - if (state == IDLE || state == PROCESSING) + if (state == PROCESSING) { t_token_generation = (ggml_time_us() - t_start_genereration) / 1e3; command = RELEASE; @@ -539,7 +313,6 @@ struct llama_server_context bool all_slots_are_idle = false; bool add_bos_token = true; - int32_t id_gen; int32_t n_ctx; // total context for all clients / slots // system prompt @@ -554,13 +327,8 @@ struct llama_server_context // slots / clients std::vector slots; - std::vector queue_tasks; - std::vector queue_results; - std::vector queue_multitasks; - std::mutex mutex_tasks; // also guards id_gen, and queue_multitasks - std::condition_variable condition_tasks; - std::mutex mutex_results; - std::condition_variable condition_results; + llama_server_queue queue_tasks; + llama_server_response queue_results; ~llama_server_context() { @@ -619,8 +387,6 @@ struct llama_server_context } void initialize() { - id_gen = 0; - // create slots all_slots_are_idle = true; @@ -1183,39 +949,13 @@ struct llama_server_context void send_error(task_server& task, const std::string &error) { LOG_TEE("task %i - error: %s\n", task.id, error.c_str()); - std::unique_lock lock(mutex_results); task_result res; res.id = task.id; res.multitask_id = task.multitask_id; res.stop = false; res.error = true; res.result_json = { { "content", error } }; - queue_results.push_back(res); - condition_results.notify_all(); - } - - void add_multi_task(int id, std::vector& sub_ids) - { - std::lock_guard lock(mutex_tasks); - task_multi multi; - multi.id = id; - std::copy(sub_ids.begin(), sub_ids.end(), std::inserter(multi.subtasks_remaining, multi.subtasks_remaining.end())); - queue_multitasks.push_back(multi); - condition_tasks.notify_one(); - } - - void update_multi_task(int multitask_id, int subtask_id, task_result& result) - { - std::lock_guard lock(mutex_tasks); - for (auto& multitask : queue_multitasks) - { - if (multitask.id == multitask_id) - { - multitask.subtasks_remaining.erase(subtask_id); - multitask.results.push_back(result); - condition_tasks.notify_one(); - } - } + queue_results.send(res); } json get_model_props() @@ -1261,7 +1001,6 @@ struct llama_server_context void send_partial_response(llama_client_slot &slot, completion_token_output tkn) { - std::unique_lock lock(mutex_results); task_result res; res.id = slot.task_id; res.multitask_id = slot.multitask_id; @@ -1296,13 +1035,11 @@ struct llama_server_context res.result_json["model"] = slot.oaicompat_model; } - queue_results.push_back(res); - condition_results.notify_all(); + queue_results.send(res); } void send_final_response(llama_client_slot &slot) { - std::unique_lock lock(mutex_results); task_result res; res.id = slot.task_id; res.multitask_id = slot.multitask_id; @@ -1351,22 +1088,11 @@ struct llama_server_context res.result_json["model"] = slot.oaicompat_model; } - queue_results.push_back(res); - condition_results.notify_all(); - - // done with results, unlock - lock.unlock(); - - // parent multitask, if any, needs to be updated - if (slot.multitask_id != -1) - { - update_multi_task(slot.multitask_id, slot.task_id, res); - } + queue_results.send(res); } void send_embedding(llama_client_slot &slot) { - std::unique_lock lock(mutex_results); task_result res; res.id = slot.task_id; res.multitask_id = slot.multitask_id; @@ -1393,15 +1119,13 @@ struct llama_server_context {"embedding", embedding }, }; } - queue_results.push_back(res); - condition_results.notify_all(); + queue_results.send(res); } - int request_completion(json data, bool infill, bool embedding, int multitask_id) + void request_completion(int task_id, json data, bool infill, bool embedding, int multitask_id) { - std::unique_lock lock(mutex_tasks); task_server task; - task.id = id_gen++; + task.id = task_id; task.target_id = 0; task.data = std::move(data); task.infill_mode = infill; @@ -1412,47 +1136,11 @@ struct llama_server_context // when a completion task's prompt array is not a singleton, we split it into multiple requests if (task.data.count("prompt") && task.data.at("prompt").size() > 1) { - lock.unlock(); // entering new func scope - return split_multiprompt_task(task); + split_multiprompt_task(task_id, task); } // otherwise, it's a single-prompt task, we actually queue it - queue_tasks.push_back(task); - condition_tasks.notify_one(); - return task.id; - } - - task_result next_result(int task_id) - { - while (true) - { - std::unique_lock lock(mutex_results); - condition_results.wait(lock, [&]{ - return !queue_results.empty(); - }); - - for (int i = 0; i < (int) queue_results.size(); i++) - { - // for now, tasks that have associated parent multitasks just get erased once multitask picks up the result - if (queue_results[i].multitask_id == task_id) - { - update_multi_task(task_id, queue_results[i].id, queue_results[i]); - queue_results.erase(queue_results.begin() + i); - continue; - } - - if (queue_results[i].id == task_id) - { - assert(queue_results[i].multitask_id == -1); - task_result res = queue_results[i]; - queue_results.erase(queue_results.begin() + i); - return res; - } - } - } - - // never reached - //return task_result{-1, false, false, {}}; + queue_tasks.post(task); } // for multiple images processing @@ -1525,150 +1213,117 @@ struct llama_server_context void request_cancel(int task_id) { - std::unique_lock lock(mutex_tasks); task_server task; - task.id = id_gen++; task.type = TASK_TYPE_CANCEL; task.target_id = task_id; - queue_tasks.push_back(task); - condition_tasks.notify_one(); + queue_tasks.post(task); } - int split_multiprompt_task(task_server& multiprompt_task) + void split_multiprompt_task(int multitask_id, task_server& multiprompt_task) { int prompt_count = multiprompt_task.data.at("prompt").size(); assert(prompt_count > 1); - int multitask_id = id_gen++; + // generate all the ID for subtask std::vector subtask_ids(prompt_count); for (int i = 0; i < prompt_count; i++) + { + subtask_ids[i] = queue_tasks.get_new_id(); + } + + // queue up the multitask so we can track its subtask progression + queue_tasks.add_multitask(multitask_id, subtask_ids); + + // add subtasks + for (int i = 0; i < prompt_count; i++) { json subtask_data = multiprompt_task.data; subtask_data["prompt"] = subtask_data["prompt"][i]; // subtasks inherit everything else (infill mode, embedding mode, etc.) - subtask_ids[i] = request_completion(subtask_data, multiprompt_task.infill_mode, multiprompt_task.embedding_mode, multitask_id); + request_completion(subtask_ids[i], subtask_data, multiprompt_task.infill_mode, multiprompt_task.embedding_mode, multitask_id); } - - // queue up the multitask so we can track its subtask progression - add_multi_task(multitask_id, subtask_ids); - return multitask_id; } - void process_tasks() + void process_single_task(task_server& task) { - std::unique_lock lock(mutex_tasks); - std::vector deferred_tasks; - while (!queue_tasks.empty()) + switch (task.type) { - task_server task = queue_tasks.front(); - queue_tasks.erase(queue_tasks.begin()); - switch (task.type) - { - case TASK_TYPE_COMPLETION: { - llama_client_slot *slot = get_slot(json_value(task.data, "slot_id", -1)); - if (slot == nullptr) - { - // if no slot is available, we defer this task for processing later - deferred_tasks.push_back(task); - break; - } - - if (task.data.contains("system_prompt")) - { - if (!all_slots_are_idle) { - send_error(task, "system prompt can only be updated when all slots are idle"); - break; - } - process_system_prompt_data(task.data["system_prompt"]); - - // reset cache_tokens for all slots - for (llama_client_slot &slot : slots) - { - slot.cache_tokens.clear(); - } - } - - slot->reset(); - - slot->infill = task.infill_mode; - slot->embedding = task.embedding_mode; - slot->task_id = task.id; - slot->multitask_id = task.multitask_id; - - if (!launch_slot_with_data(slot, task.data)) - { - // send error result - send_error(task, "internal_error"); - break; - } - } break; - case TASK_TYPE_CANCEL: { // release slot linked with the task id - for (auto & slot : slots) - { - if (slot.task_id == task.target_id) - { - slot.release(); - break; - } - } - } break; - } - } - - // add all the deferred tasks back the the queue - for (task_server &task : deferred_tasks) - { - queue_tasks.push_back(task); - } - - // remove finished multitasks from the queue of multitasks, and add the corresponding result to the result queue - std::vector agg_results; - auto queue_iterator = queue_multitasks.begin(); - while (queue_iterator != queue_multitasks.end()) - { - if (queue_iterator->subtasks_remaining.empty()) - { - // all subtasks done == multitask is done - task_result aggregate_result; - aggregate_result.id = queue_iterator->id; - aggregate_result.stop = true; - aggregate_result.error = false; - - // collect json results into one json result - std::vector result_jsons; - for (auto& subres : queue_iterator->results) + case TASK_TYPE_COMPLETION: { + llama_client_slot *slot = get_slot(json_value(task.data, "slot_id", -1)); + if (slot == nullptr) { - result_jsons.push_back(subres.result_json); - aggregate_result.error = aggregate_result.error && subres.error; + // if no slot is available, we defer this task for processing later + LOG_VERBOSE("no slot is available", {}); + queue_tasks.defer(task); + break; } - aggregate_result.result_json = json{ "results", result_jsons }; + if (task.data.contains("system_prompt")) + { + if (!all_slots_are_idle) { + send_error(task, "system prompt can only be updated when all slots are idle"); + break; + } + process_system_prompt_data(task.data["system_prompt"]); - agg_results.push_back(aggregate_result); + // reset cache_tokens for all slots + for (llama_client_slot &slot : slots) + { + slot.cache_tokens.clear(); + } + } - condition_results.notify_all(); + slot->reset(); - queue_iterator = queue_multitasks.erase(queue_iterator); - } - else - { - ++queue_iterator; - } + slot->infill = task.infill_mode; + slot->embedding = task.embedding_mode; + slot->task_id = task.id; + slot->multitask_id = task.multitask_id; + + if (!launch_slot_with_data(slot, task.data)) + { + // send error result + send_error(task, "internal_error"); + break; + } + } break; + case TASK_TYPE_CANCEL: { // release slot linked with the task id + for (auto & slot : slots) + { + if (slot.task_id == task.target_id) + { + slot.release(); + break; + } + } + } break; + case TASK_TYPE_NEXT_RESPONSE: { + // do nothing + } break; } + } - // done with tasks, unlock - lock.unlock(); + void on_finish_multitask(task_multi& multitask) + { + // all subtasks done == multitask is done + task_result result; + result.id = multitask.id; + result.stop = true; + result.error = false; - // copy aggregate results of complete multi-tasks to the results queue - std::lock_guard lock_results(mutex_results); - queue_results.insert(queue_results.end(), agg_results.begin(), agg_results.end()); + // collect json results into one json result + std::vector result_jsons; + for (auto& subres : multitask.results) + { + result_jsons.push_back(subres.result_json); + result.error = result.error && subres.error; + } + result.result_json = json{ { "results", result_jsons } }; + queue_results.send(result); } bool update_slots() { - // attend tasks - process_tasks(); - if (system_need_update) { LOG_TEE("updating system prompt\n"); @@ -1684,10 +1339,12 @@ struct llama_server_context LOG_TEE("all slots are idle and system prompt is empty, clear the KV cache\n"); kv_cache_clear(); } - std::unique_lock lock(mutex_tasks); - condition_tasks.wait(lock, [&]{ - return !queue_tasks.empty(); - }); + return true; + } else { + task_server task; + task.type = TASK_TYPE_NEXT_RESPONSE; + task.target_id = -1; + queue_tasks.post(task); } for (llama_client_slot &slot : slots) @@ -1732,6 +1389,7 @@ struct llama_server_context slot.t_last_used = ggml_time_us(); LOG_TEE("slot %d released (%d tokens in cache)\n", slot.id, (int) slot.cache_tokens.size()); + queue_tasks.notify_slot_changed(); continue; } @@ -1997,6 +1655,10 @@ struct llama_server_context } return true; } + + void run_on_all_tasks_finished() { + update_slots(); + } }; static void server_print_usage(const char *argv0, const gpt_params ¶ms, @@ -2541,239 +2203,7 @@ static void server_params_parse(int argc, char **argv, server_params &sparams, } } -static std::string random_string() -{ - static const std::string str("0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz"); - - std::random_device rd; - std::mt19937 generator(rd()); - - std::string result(32, ' '); - - for (int i = 0; i < 32; ++i) { - result[i] = str[generator() % str.size()]; - } - - return result; -} - -static std::string gen_chatcmplid() -{ - std::stringstream chatcmplid; - chatcmplid << "chatcmpl-" << random_string(); - return chatcmplid.str(); -} - -std::string format_chatml(std::vector messages) -{ - std::ostringstream chatml_msgs; - - for (auto it = messages.begin(); it != messages.end(); ++it) { - chatml_msgs << "<|im_start|>" - << json_value(*it, "role", std::string("user")) << '\n'; - chatml_msgs << json_value(*it, "content", std::string("")) - << "<|im_end|>\n"; - } - - chatml_msgs << "<|im_start|>assistant" << '\n'; - - return chatml_msgs.str(); -} - /* llama.cpp completion api semantics */ -json oaicompat_completion_params_parse( - const json &body /* openai api json semantics */) -{ - json llama_params; - - llama_params["__oaicompat"] = true; - - // Map OpenAI parameters to llama.cpp parameters - // - // For parameters that are defined by the OpenAI documentation (e.g. - // temperature), we explicitly specify OpenAI's intended default; we - // need to do that because sometimes OpenAI disagrees with llama.cpp - // - // https://platform.openai.com/docs/api-reference/chat/create - llama_sampling_params default_sparams; - llama_params["model"] = json_value(body, "model", std::string("unknown")); - llama_params["prompt"] = format_chatml(body["messages"]); // OpenAI 'messages' to llama.cpp 'prompt' - llama_params["cache_prompt"] = json_value(body, "cache_prompt", false); - llama_params["temperature"] = json_value(body, "temperature", 0.0); - llama_params["top_k"] = json_value(body, "top_k", default_sparams.top_k); - llama_params["top_p"] = json_value(body, "top_p", 1.0); - llama_params["n_predict"] = json_value(body, "max_tokens", -1); - llama_params["logit_bias"] = json_value(body, "logit_bias",json::object()); - llama_params["frequency_penalty"] = json_value(body, "frequency_penalty", 0.0); - llama_params["presence_penalty"] = json_value(body, "presence_penalty", 0.0); - llama_params["seed"] = json_value(body, "seed", LLAMA_DEFAULT_SEED); - llama_params["stream"] = json_value(body, "stream", false); - llama_params["mirostat"] = json_value(body, "mirostat", default_sparams.mirostat); - llama_params["mirostat_tau"] = json_value(body, "mirostat_tau", default_sparams.mirostat_tau); - llama_params["mirostat_eta"] = json_value(body, "mirostat_eta", default_sparams.mirostat_eta); - llama_params["penalize_nl"] = json_value(body, "penalize_nl", default_sparams.penalize_nl); - llama_params["typical_p"] = json_value(body, "typical_p", default_sparams.typical_p); - llama_params["repeat_last_n"] = json_value(body, "repeat_last_n", default_sparams.penalty_last_n); - llama_params["ignore_eos"] = json_value(body, "ignore_eos", false); - llama_params["tfs_z"] = json_value(body, "tfs_z", default_sparams.tfs_z); - - if (body.count("grammar") != 0) { - llama_params["grammar"] = json_value(body, "grammar", json::object()); - } - - // Handle 'stop' field - if (body.contains("stop") && body["stop"].is_string()) { - llama_params["stop"] = json::array({body["stop"].get()}); - } else { - llama_params["stop"] = json_value(body, "stop", json::array()); - } - - // Ensure there is ChatML-specific end sequence among stop words - llama_params["stop"].push_back("<|im_end|>"); - - return llama_params; -} - -static json format_final_response_oaicompat(const json &request, const task_result &response, bool streaming = false) -{ - json result = response.result_json; - - bool stopped_word = result.count("stopped_word") != 0; - bool stopped_eos = json_value(result, "stopped_eos", false); - int num_tokens_predicted = json_value(result, "tokens_predicted", 0); - int num_prompt_tokens = json_value(result, "tokens_evaluated", 0); - std::string content = json_value(result, "content", std::string("")); - - std::string finish_reason = "length"; - if (stopped_word || stopped_eos) { - finish_reason = "stop"; - } - - json choices = - streaming ? json::array({json{{"finish_reason", finish_reason}, - {"index", 0}, - {"delta", json::object()}}}) - : json::array({json{{"finish_reason", finish_reason}, - {"index", 0}, - {"message", json{{"content", content}, - {"role", "assistant"}}}}}); - - std::time_t t = std::time(0); - - json res = - json{{"choices", choices}, - {"created", t}, - {"model", - json_value(request, "model", std::string(DEFAULT_OAICOMPAT_MODEL))}, - {"object", streaming ? "chat.completion.chunk" : "chat.completion"}, - {"usage", - json{{"completion_tokens", num_tokens_predicted}, - {"prompt_tokens", num_prompt_tokens}, - {"total_tokens", num_tokens_predicted + num_prompt_tokens}}}, - {"id", gen_chatcmplid()}}; - - if (server_verbose) { - res["__verbose"] = result; - } - - if (result.contains("completion_probabilities")) { - res["completion_probabilities"] = json_value(result, "completion_probabilities", json::array()); - } - - return res; -} - -// return value is vector as there is one case where we might need to generate two responses -static std::vector format_partial_response_oaicompat(const task_result &response) { - json result = response.result_json; - - if (!result.contains("model") || !result.contains("oaicompat_token_ctr")) { - return std::vector({response.result_json}); - } - - bool first = json_value(result, "oaicompat_token_ctr", 0) == 0; - std::string modelname = json_value(result, "model", std::string(DEFAULT_OAICOMPAT_MODEL)); - - bool stopped_word = json_value(result, "stopped_word", false); - bool stopped_eos = json_value(result, "stopped_eos", false); - bool stopped_limit = json_value(result, "stopped_limit", false); - std::string content = json_value(result, "content", std::string("")); - - std::string finish_reason; - if (stopped_word || stopped_eos) { - finish_reason = "stop"; - } - if (stopped_limit) { - finish_reason = "length"; - } - - std::time_t t = std::time(0); - - json choices; - - if (!finish_reason.empty()) { - choices = json::array({json{{"finish_reason", finish_reason}, - {"index", 0}, - {"delta", json::object()}}}); - } else { - if (first) { - if (content.empty()) { - choices = json::array({json{{"finish_reason", nullptr}, - {"index", 0}, - {"delta", json{{"role", "assistant"}}}}}); - } else { - // We have to send this as two updates to conform to openai behavior - json initial_ret = json{{"choices", json::array({json{ - {"finish_reason", nullptr}, - {"index", 0}, - {"delta", json{ - {"role", "assistant"} - }}}})}, - {"created", t}, - {"id", gen_chatcmplid()}, - {"model", modelname}, - {"object", "chat.completion.chunk"}}; - - json second_ret = json{ - {"choices", json::array({json{{"finish_reason", nullptr}, - {"index", 0}, - {"delta", json{ - {"content", content}}} - }})}, - {"created", t}, - {"id", gen_chatcmplid()}, - {"model", modelname}, - {"object", "chat.completion.chunk"}}; - - return std::vector({initial_ret, second_ret}); - } - } else { - // Some idiosyncrasy in task processing logic makes several trailing calls - // with empty content, we ignore these at the calee site. - if (content.empty()) { - return std::vector({json::object()}); - } - - choices = json::array({json{ - {"finish_reason", nullptr}, - {"index", 0}, - {"delta", - json{ - {"content", content}, - }}, - }}); - } - } - - json ret = json{{"choices", choices}, - {"created", t}, - {"id", gen_chatcmplid()}, - {"model", modelname}, - {"object", "chat.completion.chunk"}}; - - return std::vector({ret}); -} - static json format_partial_response( llama_server_context &llama, llama_client_slot *slot, const std::string &content, const std::vector &probs ) { @@ -3069,10 +2499,12 @@ int main(int argc, char **argv) return; } json data = json::parse(req.body); - const int task_id = llama.request_completion(data, false, false, -1); + const int task_id = llama.queue_tasks.get_new_id(); + llama.queue_results.add_waiting_task_id(task_id); + llama.request_completion(task_id, data, false, false, -1); if (!json_value(data, "stream", false)) { std::string completion_text; - task_result result = llama.next_result(task_id); + task_result result = llama.queue_results.recv(task_id); if (!result.error && result.stop) { res.set_content(result.result_json.dump(-1, ' ', false, json::error_handler_t::replace), "application/json; charset=utf-8"); } @@ -3080,14 +2512,14 @@ int main(int argc, char **argv) { res.status = 404; res.set_content(result.result_json["content"], "text/plain; charset=utf-8"); - return; } + llama.queue_results.remove_waiting_task_id(task_id); } else { const auto chunked_content_provider = [task_id, &llama](size_t, httplib::DataSink & sink) { while (true) { - task_result result = llama.next_result(task_id); + task_result result = llama.queue_results.recv(task_id); if (!result.error) { const std::string str = "data: " + @@ -3098,6 +2530,7 @@ int main(int argc, char **argv) }); if (!sink.write(str.c_str(), str.size())) { + llama.queue_results.remove_waiting_task_id(task_id); return false; } if (result.stop) { @@ -3113,11 +2546,14 @@ int main(int argc, char **argv) }); if (!sink.write(str.c_str(), str.size())) { + llama.queue_results.remove_waiting_task_id(task_id); return false; } break; } } + + llama.queue_results.remove_waiting_task_id(task_id); sink.done(); return true; }; @@ -3126,6 +2562,7 @@ int main(int argc, char **argv) { // cancel llama.request_cancel(task_id); + llama.queue_results.remove_waiting_task_id(task_id); }; res.set_chunked_content_provider("text/event-stream", chunked_content_provider, on_complete); @@ -3162,11 +2599,13 @@ int main(int argc, char **argv) } json data = oaicompat_completion_params_parse(json::parse(req.body)); - const int task_id = llama.request_completion(data, false, false, -1); + const int task_id = llama.queue_tasks.get_new_id(); + llama.queue_results.add_waiting_task_id(task_id); + llama.request_completion(task_id, data, false, false, -1); if (!json_value(data, "stream", false)) { std::string completion_text; - task_result result = llama.next_result(task_id); + task_result result = llama.queue_results.recv(task_id); if (!result.error && result.stop) { json oaicompat_result = format_final_response_oaicompat(data, result); @@ -3177,12 +2616,12 @@ int main(int argc, char **argv) } else { res.status = 500; res.set_content(result.result_json["content"], "text/plain; charset=utf-8"); - return; } + llama.queue_results.remove_waiting_task_id(task_id); } else { const auto chunked_content_provider = [task_id, &llama](size_t, httplib::DataSink &sink) { while (true) { - task_result llama_result = llama.next_result(task_id); + task_result llama_result = llama.queue_results.recv(task_id); if (!llama_result.error) { std::vector result_array = format_partial_response_oaicompat( llama_result); @@ -3195,6 +2634,7 @@ int main(int argc, char **argv) "\n\n"; LOG_VERBOSE("data stream", {{"to_send", str}}); if (!sink.write(str.c_str(), str.size())) { + llama.queue_results.remove_waiting_task_id(task_id); return false; } } @@ -3210,18 +2650,21 @@ int main(int argc, char **argv) "\n\n"; LOG_VERBOSE("data stream", {{"to_send", str}}); if (!sink.write(str.c_str(), str.size())) { + llama.queue_results.remove_waiting_task_id(task_id); return false; } break; } } sink.done(); + llama.queue_results.remove_waiting_task_id(task_id); return true; }; auto on_complete = [task_id, &llama](bool) { // cancel request llama.request_cancel(task_id); + llama.queue_results.remove_waiting_task_id(task_id); }; res.set_chunked_content_provider("text/event-stream", chunked_content_provider, on_complete); @@ -3235,10 +2678,12 @@ int main(int argc, char **argv) return; } json data = json::parse(req.body); - const int task_id = llama.request_completion(data, true, false, -1); + const int task_id = llama.queue_tasks.get_new_id(); + llama.queue_results.add_waiting_task_id(task_id); + llama.request_completion(task_id, data, true, false, -1); if (!json_value(data, "stream", false)) { std::string completion_text; - task_result result = llama.next_result(task_id); + task_result result = llama.queue_results.recv(task_id); if (!result.error && result.stop) { res.set_content(result.result_json.dump(-1, ' ', false, json::error_handler_t::replace), "application/json; charset=utf-8"); @@ -3247,13 +2692,13 @@ int main(int argc, char **argv) { res.status = 404; res.set_content(result.result_json["content"], "text/plain; charset=utf-8"); - return; } + llama.queue_results.remove_waiting_task_id(task_id); } else { const auto chunked_content_provider = [task_id, &llama](size_t, httplib::DataSink & sink) { while (true) { - task_result result = llama.next_result(task_id); + task_result result = llama.queue_results.recv(task_id); if (!result.error) { const std::string str = "data: " + @@ -3264,6 +2709,7 @@ int main(int argc, char **argv) }); if (!sink.write(str.c_str(), str.size())) { + llama.queue_results.remove_waiting_task_id(task_id); return false; } if (result.stop) @@ -3277,8 +2723,8 @@ int main(int argc, char **argv) } } + llama.queue_results.remove_waiting_task_id(task_id); sink.done(); - return true; }; @@ -3352,23 +2798,46 @@ int main(int argc, char **argv) image_data = ""; } - const int task_id = llama.request_completion({ {"prompt", prompt}, { "n_predict", 0}, {"image_data", image_data} }, false, true, -1); - task_result result = llama.next_result(task_id); + // create and queue the task + const int task_id = llama.queue_tasks.get_new_id(); + llama.queue_results.add_waiting_task_id(task_id); + llama.request_completion(task_id, { {"prompt", prompt}, { "n_predict", 0}, {"image_data", image_data} }, false, true, -1); + + // get the result + task_result result = llama.queue_results.recv(task_id); + llama.queue_results.remove_waiting_task_id(task_id); + + // send the result return res.set_content(result.result_json.dump(), "application/json; charset=utf-8"); }); // GG: if I put the main loop inside a thread, it crashes on the first request when build in Debug!? // "Bus error: 10" - this is on macOS, it does not crash on Linux //std::thread t2([&]() - { + /*{ bool running = true; while (running) { running = llama.update_slots(); } - } + }*/ //); + llama.queue_tasks.on_new_task(std::bind( + &llama_server_context::process_single_task, &llama, std::placeholders::_1)); + llama.queue_tasks.on_finish_multitask(std::bind( + &llama_server_context::on_finish_multitask, &llama, std::placeholders::_1)); + llama.queue_tasks.on_all_tasks_finished(std::bind( + &llama_server_context::run_on_all_tasks_finished, &llama)); + llama.queue_results.on_multitask_update(std::bind( + &llama_server_queue::update_multitask, + &llama.queue_tasks, + std::placeholders::_1, + std::placeholders::_2, + std::placeholders::_3 + )); + llama.queue_tasks.start_loop(); + t.join(); llama_backend_free(); diff --git a/examples/server/utils.hpp b/examples/server/utils.hpp new file mode 100644 index 000000000..e2b6065f7 --- /dev/null +++ b/examples/server/utils.hpp @@ -0,0 +1,507 @@ +#pragma once + +#include +#include +#include +#include +#include +#include + +#include "json.hpp" + +#include "../llava/clip.h" + +using json = nlohmann::json; + +extern bool server_verbose; + +#ifndef SERVER_VERBOSE +#define SERVER_VERBOSE 1 +#endif + +#if SERVER_VERBOSE != 1 +#define LOG_VERBOSE(MSG, ...) +#else +#define LOG_VERBOSE(MSG, ...) \ + do \ + { \ + if (server_verbose) \ + { \ + server_log("VERBOSE", __func__, __LINE__, MSG, __VA_ARGS__); \ + } \ + } while (0) +#endif + +#define LOG_ERROR( MSG, ...) server_log("ERROR", __func__, __LINE__, MSG, __VA_ARGS__) +#define LOG_WARNING(MSG, ...) server_log("WARNING", __func__, __LINE__, MSG, __VA_ARGS__) +#define LOG_INFO( MSG, ...) server_log("INFO", __func__, __LINE__, MSG, __VA_ARGS__) + +// +// parallel +// + +enum server_state { + SERVER_STATE_LOADING_MODEL, // Server is starting up, model not fully loaded yet + SERVER_STATE_READY, // Server is ready and model is loaded + SERVER_STATE_ERROR // An error occurred, load_model failed +}; + +enum task_type { + TASK_TYPE_COMPLETION, + TASK_TYPE_CANCEL, + TASK_TYPE_NEXT_RESPONSE +}; + +struct task_server { + int id = -1; // to be filled by llama_server_queue + int target_id; + task_type type; + json data; + bool infill_mode = false; + bool embedding_mode = false; + int multitask_id = -1; +}; + +struct task_result { + int id; + int multitask_id = -1; + bool stop; + bool error; + json result_json; +}; + +struct task_multi { + int id; + std::set subtasks_remaining{}; + std::vector results{}; +}; + +// TODO: can become bool if we can't find use of more states +enum slot_state +{ + IDLE, + PROCESSING, +}; + +enum slot_command +{ + NONE, + LOAD_PROMPT, + RELEASE, +}; + +struct slot_params +{ + bool stream = true; + bool cache_prompt = false; // remember the prompt to avoid reprocessing all prompt + + uint32_t seed = -1; // RNG seed + int32_t n_keep = 0; // number of tokens to keep from initial prompt + int32_t n_predict = -1; // new tokens to predict + + std::vector antiprompt; + + json input_prefix; + json input_suffix; +}; + +struct slot_image +{ + int32_t id; + + bool request_encode_image = false; + float * image_embedding = nullptr; + int32_t image_tokens = 0; + + clip_image_u8 * img_data; + + std::string prefix_prompt; // before of this image +}; + +// completion token output with probabilities +struct completion_token_output +{ + struct token_prob + { + llama_token tok; + float prob; + }; + + std::vector probs; + llama_token tok; + std::string text_to_send; +}; + +static inline void server_log(const char *level, const char *function, int line, + const char *message, const nlohmann::ordered_json &extra) +{ + nlohmann::ordered_json log + { + {"timestamp", time(nullptr)}, + {"level", level}, + {"function", function}, + {"line", line}, + {"message", message}, + }; + + if (!extra.empty()) + { + log.merge_patch(extra); + } + + const std::string str = log.dump(-1, ' ', false, json::error_handler_t::replace); + printf("%.*s\n", (int)str.size(), str.data()); + fflush(stdout); +} + +// +// server utils +// + +template +static T json_value(const json &body, const std::string &key, const T &default_value) +{ + // Fallback null to default value + return body.contains(key) && !body.at(key).is_null() + ? body.value(key, default_value) + : default_value; +} + +inline std::string format_chatml(std::vector messages) +{ + std::ostringstream chatml_msgs; + + for (auto it = messages.begin(); it != messages.end(); ++it) { + chatml_msgs << "<|im_start|>" + << json_value(*it, "role", std::string("user")) << '\n'; + chatml_msgs << json_value(*it, "content", std::string("")) + << "<|im_end|>\n"; + } + + chatml_msgs << "<|im_start|>assistant" << '\n'; + + return chatml_msgs.str(); +} + +// +// work queue utils +// + +struct llama_server_queue { + int id = 0; + std::mutex mutex_tasks; + // queues + std::vector queue_tasks; + std::vector queue_tasks_deferred; + std::vector queue_multitasks; + std::condition_variable condition_tasks; + // callback functions + std::function callback_new_task; + std::function callback_finish_multitask; + std::function callback_all_task_finished; + + // Add a new task to the end of the queue + int post(task_server task) { + std::unique_lock lock(mutex_tasks); + if (task.id == -1) { + task.id = id++; + } + queue_tasks.push_back(std::move(task)); + condition_tasks.notify_one(); + return task.id; + } + + // Add a new task, but defer until one slot is available + void defer(task_server task) { + std::unique_lock lock(mutex_tasks); + queue_tasks_deferred.push_back(std::move(task)); + } + + // Get the next id for creating anew task + int get_new_id() { + std::unique_lock lock(mutex_tasks); + return id++; + } + + // Register function to process a new task + void on_new_task(std::function callback) { + callback_new_task = callback; + } + + // Register function to process a multitask + void on_finish_multitask(std::function callback) { + callback_finish_multitask = callback; + } + + // Register the function to be called when the batch of tasks is finished + void on_all_tasks_finished(std::function callback) { + callback_all_task_finished = callback; + } + + // Call when the state of one slot is changed + void notify_slot_changed() { + // move deferred tasks back to main loop + std::unique_lock lock(mutex_tasks); + for (auto & task : queue_tasks_deferred) { + queue_tasks.push_back(std::move(task)); + } + queue_tasks_deferred.clear(); + } + + // Start the main loop. This call is blocking + void start_loop() { + while (true) { + // new task arrived + LOG_VERBOSE("have new task", {}); + { + while (true) + { + std::unique_lock lock(mutex_tasks); + if (queue_tasks.empty()) { + lock.unlock(); + break; + } + task_server task = queue_tasks.front(); + queue_tasks.erase(queue_tasks.begin()); + lock.unlock(); + LOG_VERBOSE("callback_new_task", {}); + callback_new_task(task); + } + LOG_VERBOSE("callback_all_task_finished", {}); + // process and update all the multitasks + auto queue_iterator = queue_multitasks.begin(); + while (queue_iterator != queue_multitasks.end()) + { + if (queue_iterator->subtasks_remaining.empty()) + { + // all subtasks done == multitask is done + task_multi current_multitask = *queue_iterator; + callback_finish_multitask(current_multitask); + // remove this multitask + queue_iterator = queue_multitasks.erase(queue_iterator); + } + else + { + ++queue_iterator; + } + } + // all tasks in the current loop is finished + callback_all_task_finished(); + } + LOG_VERBOSE("wait for new task", {}); + // wait for new task + { + std::unique_lock lock(mutex_tasks); + if (queue_tasks.empty()) { + condition_tasks.wait(lock, [&]{ + return !queue_tasks.empty(); + }); + } + } + } + } + + // + // functions to manage multitasks + // + + // add a multitask by specifying the id of all subtask (subtask is a task_server) + void add_multitask(int multitask_id, std::vector& sub_ids) + { + std::lock_guard lock(mutex_tasks); + task_multi multi; + multi.id = multitask_id; + std::copy(sub_ids.begin(), sub_ids.end(), std::inserter(multi.subtasks_remaining, multi.subtasks_remaining.end())); + queue_multitasks.push_back(multi); + } + + // updatethe remaining subtasks, while appending results to multitask + void update_multitask(int multitask_id, int subtask_id, task_result& result) + { + std::lock_guard lock(mutex_tasks); + for (auto& multitask : queue_multitasks) + { + if (multitask.id == multitask_id) + { + multitask.subtasks_remaining.erase(subtask_id); + multitask.results.push_back(result); + } + } + } +}; + +struct llama_server_response { + typedef std::function callback_multitask_t; + callback_multitask_t callback_update_multitask; + // for keeping track of all tasks waiting for the result + std::set waiting_task_ids; + // the main result queue + std::vector queue_results; + std::mutex mutex_results; + std::condition_variable condition_results; + + void add_waiting_task_id(int task_id) { + std::unique_lock lock(mutex_results); + waiting_task_ids.insert(task_id); + } + + void remove_waiting_task_id(int task_id) { + std::unique_lock lock(mutex_results); + waiting_task_ids.erase(task_id); + } + + // This function blocks the thread until there is a response for this task_id + task_result recv(int task_id) { + while (true) + { + std::unique_lock lock(mutex_results); + condition_results.wait(lock, [&]{ + return !queue_results.empty(); + }); + LOG_VERBOSE("condition_results unblock", {}); + + for (int i = 0; i < (int) queue_results.size(); i++) + { + if (queue_results[i].id == task_id) + { + assert(queue_results[i].multitask_id == -1); + task_result res = queue_results[i]; + queue_results.erase(queue_results.begin() + i); + return res; + } + } + } + + // should never reach here + } + + // Register the function to update multitask + void on_multitask_update(callback_multitask_t callback) { + callback_update_multitask = callback; + } + + // Send a new result to a waiting task_id + void send(task_result result) { + std::unique_lock lock(mutex_results); + LOG_VERBOSE("send new result", {}); + for (auto& task_id : waiting_task_ids) { + // LOG_TEE("waiting task id %i \n", task_id); + // for now, tasks that have associated parent multitasks just get erased once multitask picks up the result + if (result.multitask_id == task_id) + { + LOG_VERBOSE("callback_update_multitask", {}); + callback_update_multitask(task_id, result.id, result); + continue; + } + + if (result.id == task_id) + { + LOG_VERBOSE("queue_results.push_back", {}); + queue_results.push_back(result); + condition_results.notify_one(); + return; + } + } + } +}; + +// +// base64 utils (TODO: move to common in the future) +// + +static const std::string base64_chars = + "ABCDEFGHIJKLMNOPQRSTUVWXYZ" + "abcdefghijklmnopqrstuvwxyz" + "0123456789+/"; + +static inline bool is_base64(uint8_t c) +{ + return (isalnum(c) || (c == '+') || (c == '/')); +} + +static inline std::vector base64_decode(const std::string & encoded_string) +{ + int i = 0; + int j = 0; + int in_ = 0; + + int in_len = encoded_string.size(); + + uint8_t char_array_4[4]; + uint8_t char_array_3[3]; + + std::vector ret; + + while (in_len-- && (encoded_string[in_] != '=') && is_base64(encoded_string[in_])) + { + char_array_4[i++] = encoded_string[in_]; in_++; + if (i == 4) + { + for (i = 0; i <4; i++) + { + char_array_4[i] = base64_chars.find(char_array_4[i]); + } + + char_array_3[0] = ((char_array_4[0] ) << 2) + ((char_array_4[1] & 0x30) >> 4); + char_array_3[1] = ((char_array_4[1] & 0xf) << 4) + ((char_array_4[2] & 0x3c) >> 2); + char_array_3[2] = ((char_array_4[2] & 0x3) << 6) + char_array_4[3]; + + for (i = 0; (i < 3); i++) + { + ret.push_back(char_array_3[i]); + } + i = 0; + } + } + + if (i) + { + for (j = i; j <4; j++) + { + char_array_4[j] = 0; + } + + for (j = 0; j <4; j++) + { + char_array_4[j] = base64_chars.find(char_array_4[j]); + } + + char_array_3[0] = ((char_array_4[0] ) << 2) + ((char_array_4[1] & 0x30) >> 4); + char_array_3[1] = ((char_array_4[1] & 0xf) << 4) + ((char_array_4[2] & 0x3c) >> 2); + char_array_3[2] = ((char_array_4[2] & 0x3) << 6) + char_array_4[3]; + + for (j = 0; (j < i - 1); j++) + { + ret.push_back(char_array_3[j]); + } + } + + return ret; +} + +// +// random string / id +// + +static std::string random_string() +{ + static const std::string str("0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz"); + + std::random_device rd; + std::mt19937 generator(rd()); + + std::string result(32, ' '); + + for (int i = 0; i < 32; ++i) { + result[i] = str[generator() % str.size()]; + } + + return result; +} + +static std::string gen_chatcmplid() +{ + std::stringstream chatcmplid; + chatcmplid << "chatcmpl-" << random_string(); + return chatcmplid.str(); +} From 3b7c914de25c6851396d7f9178249f1ed278120e Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Fri, 26 Jan 2024 14:48:15 +0200 Subject: [PATCH 51/66] tests : gitignore test-c.o --- tests/.gitignore | 1 + 1 file changed, 1 insertion(+) diff --git a/tests/.gitignore b/tests/.gitignore index 59be43b99..092dce742 100644 --- a/tests/.gitignore +++ b/tests/.gitignore @@ -1,2 +1,3 @@ * !*.* +test-c.o From 5f1925a8cef81eb9b372faaae34b0dd76d5361d4 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Fri, 26 Jan 2024 17:09:44 +0200 Subject: [PATCH 52/66] scripts : move run-with-preset.py from root to scripts folder --- run_with_preset.py => scripts/run-with-preset.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) rename run_with_preset.py => scripts/run-with-preset.py (98%) diff --git a/run_with_preset.py b/scripts/run-with-preset.py similarity index 98% rename from run_with_preset.py rename to scripts/run-with-preset.py index 9b4d7ecbe..a18252730 100755 --- a/run_with_preset.py +++ b/scripts/run-with-preset.py @@ -46,7 +46,7 @@ Formatting considerations: - To define multiple "reverse_prompt" properties simultaneously the expected format is a list of strings. - To define a tensor split, pass a list of floats. """ -usage = "run_with_preset.py [-h] [yaml_files ...] [-- ...]" +usage = "run-with-preset.py [-h] [yaml_files ...] [-- ...]" epilog = (" -- specify additional CLI ars to be passed to the binary (override all preset files). " "Unknown args will be ignored.") From 7032f4f6349c17a8352f9f93f7d2122f45469e59 Mon Sep 17 00:00:00 2001 From: snadampal <87143774+snadampal@users.noreply.github.com> Date: Fri, 26 Jan 2024 11:17:59 -0600 Subject: [PATCH 53/66] ggml : update softmax n_task calculation (#5126) updated the n_task calculation to use max number of threads possible. This has improved the prompt eval performance by around 5% for DOT kernels and by around 10% for MMLA kernels on AWS Graviton3. --- ggml.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/ggml.c b/ggml.c index ca98fde8a..ef6fd8caf 100644 --- a/ggml.c +++ b/ggml.c @@ -16597,7 +16597,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) { } break; case GGML_OP_SOFT_MAX: { - n_tasks = MIN(MIN(4, n_threads), ggml_nrows(node->src[0])); + n_tasks = MIN(n_threads, ggml_nrows(node->src[0])); } break; case GGML_OP_CONV_TRANSPOSE_1D: { From 15b4538ff29b280a395a1406d711497d8eaa2564 Mon Sep 17 00:00:00 2001 From: slaren Date: Fri, 26 Jan 2024 18:18:26 +0100 Subject: [PATCH 54/66] ggml-alloc : add 10% margin to the buffer sizes (#5149) --- ggml-alloc.c | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-) diff --git a/ggml-alloc.c b/ggml-alloc.c index 60141a34d..95a93c99d 100644 --- a/ggml-alloc.c +++ b/ggml-alloc.c @@ -335,7 +335,9 @@ bool ggml_tallocr_is_measure(ggml_tallocr_t alloc) { } size_t ggml_tallocr_max_size(ggml_tallocr_t alloc) { - return alloc->max_size; + // FIXME: changes in the tensor sizes compared to the measure graph may cause allocations to fail + // to avoid this, we add a 10% margin to the buffer size + return alloc->max_size + alloc->max_size/10; } // graph allocator From 62fead3ea0a30c8d424f4a8373fa14165c7c707f Mon Sep 17 00:00:00 2001 From: slaren Date: Fri, 26 Jan 2024 18:59:43 +0100 Subject: [PATCH 55/66] cuda : fix tensor size calculation for non-split buffer (#5145) --- ggml-backend.c | 4 +++- ggml-cuda.cu | 19 +++++-------------- 2 files changed, 8 insertions(+), 15 deletions(-) diff --git a/ggml-backend.c b/ggml-backend.c index 423512def..3fff5fc87 100644 --- a/ggml-backend.c +++ b/ggml-backend.c @@ -30,7 +30,9 @@ size_t ggml_backend_buft_get_alignment(ggml_backend_buffer_type_t buft) { GGML_CALL size_t ggml_backend_buft_get_alloc_size(ggml_backend_buffer_type_t buft, struct ggml_tensor * tensor) { // get_alloc_size is optional, defaults to ggml_nbytes if (buft->iface.get_alloc_size) { - return buft->iface.get_alloc_size(buft, tensor); + size_t size = buft->iface.get_alloc_size(buft, tensor); + assert(size >= ggml_nbytes(tensor)); + return size; } return ggml_nbytes(tensor); } diff --git a/ggml-cuda.cu b/ggml-cuda.cu index 05e5d18ab..0d599e20a 100644 --- a/ggml-cuda.cu +++ b/ggml-cuda.cu @@ -9790,8 +9790,8 @@ static void ggml_cuda_mul_mat_id(const ggml_tensor * src0, const ggml_tensor * s // TODO: mmq/mmv support #endif - const int64_t nb11 = src1->nb[1]; - const int64_t nb1 = dst->nb[1]; + const size_t nb11 = src1->nb[1]; + const size_t nb1 = dst->nb[1]; const struct ggml_tensor * ids = src0; const int32_t id = ((int32_t *) dst->op_params)[0]; @@ -10304,15 +10304,11 @@ GGML_CALL static void ggml_backend_cuda_buffer_init_tensor(ggml_backend_buffer_t if (ggml_is_quantized(tensor->type)) { // initialize padding to 0 to avoid possible NaN values - int64_t row_low = 0; - int64_t row_high = ggml_nrows(tensor); - int64_t nrows_split = row_high - row_low; - - size_t original_size = ggml_nbytes_split(tensor, nrows_split); + size_t original_size = ggml_nbytes(tensor); size_t padded_size = ggml_backend_buft_get_alloc_size(buffer->buft, tensor); if (padded_size > original_size && tensor->view_src == nullptr) { - CUDA_CHECK(cudaMemsetAsync((char *)tensor->data + original_size, 0, padded_size - original_size, g_cudaStreams[ctx->device][0])); + CUDA_CHECK(cudaMemset((char *)tensor->data + original_size, 0, padded_size - original_size)); } } } @@ -10415,12 +10411,7 @@ GGML_CALL static size_t ggml_backend_cuda_buffer_type_get_alignment(ggml_backend } GGML_CALL static size_t ggml_backend_cuda_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor * tensor) { - int64_t row_low = 0; - int64_t row_high = ggml_nrows(tensor); - int64_t nrows_split = row_high - row_low; - - size_t size = ggml_nbytes_split(tensor, nrows_split); - + size_t size = ggml_nbytes(tensor); int64_t ne0 = tensor->ne[0]; if (ggml_is_quantized(tensor->type)) { From bbe7c56c9993af86aa2d84cbe1fd69e1b4300cea Mon Sep 17 00:00:00 2001 From: Jared Van Bortel Date: Fri, 26 Jan 2024 15:34:06 -0500 Subject: [PATCH 56/66] cmake : pass CPU architecture flags to nvcc (#5146) --- CMakeLists.txt | 74 ++++++++++++++++++++++++++------------------------ 1 file changed, 39 insertions(+), 35 deletions(-) diff --git a/CMakeLists.txt b/CMakeLists.txt index af3665129..2b2ae532e 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -466,17 +466,17 @@ function(get_flags CCID CCVER) (CCID STREQUAL "Clang" AND CCVER VERSION_GREATER_EQUAL 3.8.0) OR (CCID STREQUAL "AppleClang" AND CCVER VERSION_GREATER_EQUAL 7.3.0) ) - set(C_FLAGS ${C_FLAGS} -Wdouble-promotion) + list(APPEND C_FLAGS -Wdouble-promotion) endif() elseif (CCID STREQUAL "GNU") set(C_FLAGS -Wdouble-promotion) set(CXX_FLAGS -Wno-array-bounds) if (CCVER VERSION_GREATER_EQUAL 7.1.0) - set(CXX_FLAGS ${CXX_FLAGS} -Wno-format-truncation) + list(APPEND CXX_FLAGS -Wno-format-truncation) endif() if (CCVER VERSION_GREATER_EQUAL 8.1.0) - set(CXX_FLAGS ${CXX_FLAGS} -Wextra-semi) + list(APPEND CXX_FLAGS -Wextra-semi) endif() elseif (CCID MATCHES "Intel") # enable max optimization level when using Intel compiler @@ -510,16 +510,18 @@ if (LLAMA_ALL_WARNINGS) endif() endif() +set(CUDA_CXX_FLAGS "") + if (LLAMA_CUBLAS) set(CUDA_FLAGS ${CXX_FLAGS} -use_fast_math) if (NOT MSVC) - set(CUDA_FLAGS ${CUDA_FLAGS} -Wno-pedantic) + list(APPEND CUDA_FLAGS -Wno-pedantic) endif() if (LLAMA_ALL_WARNINGS AND NOT MSVC) set(NVCC_CMD ${CMAKE_CUDA_COMPILER} .c) if (NOT CMAKE_CUDA_HOST_COMPILER STREQUAL "") - set(NVCC_CMD ${NVCC_CMD} -ccbin ${CMAKE_CUDA_HOST_COMPILER}) + list(APPEND NVCC_CMD -ccbin ${CMAKE_CUDA_HOST_COMPILER}) endif() execute_process( @@ -547,13 +549,8 @@ if (LLAMA_CUBLAS) message("-- CUDA host compiler is ${CUDA_CCID} ${CUDA_CCVER}") get_flags(${CUDA_CCID} ${CUDA_CCVER}) - list(JOIN GF_CXX_FLAGS " " CUDA_CXX_FLAGS) # pass host compiler flags as a single argument - if (NOT CUDA_CXX_FLAGS STREQUAL "") - set(CUDA_FLAGS ${CUDA_FLAGS} -Xcompiler ${CUDA_CXX_FLAGS}) - endif() + list(APPEND CUDA_CXX_FLAGS ${GF_CXX_FLAGS}) # This is passed to -Xcompiler later endif() - - add_compile_options("$<$:${CUDA_FLAGS}>") endif() if (WIN32) @@ -618,12 +615,7 @@ if (NOT MSVC) endif() endif() -function(add_compile_option_cpp ARG) - # Adds a compile option to C/C++ only, but not for Cuda. - # Use, e.g., for CPU-architecture flags. - add_compile_options($<$:${ARG}>) - add_compile_options($<$:${ARG}>) -endfunction() +set(ARCH_FLAGS "") if ((${CMAKE_SYSTEM_PROCESSOR} MATCHES "arm") OR (${CMAKE_SYSTEM_PROCESSOR} MATCHES "aarch64") OR ("${CMAKE_GENERATOR_PLATFORM_LWR}" MATCHES "arm64")) message(STATUS "ARM detected") @@ -636,19 +628,19 @@ if ((${CMAKE_SYSTEM_PROCESSOR} MATCHES "arm") OR (${CMAKE_SYSTEM_PROCESSOR} MATC else() check_cxx_compiler_flag(-mfp16-format=ieee COMPILER_SUPPORTS_FP16_FORMAT_I3E) if (NOT "${COMPILER_SUPPORTS_FP16_FORMAT_I3E}" STREQUAL "") - add_compile_options(-mfp16-format=ieee) + list(APPEND ARCH_FLAGS -mfp16-format=ieee) endif() if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv6") # Raspberry Pi 1, Zero - add_compile_options(-mfpu=neon-fp-armv8 -mno-unaligned-access) + list(APPEND ARCH_FLAGS -mfpu=neon-fp-armv8 -mno-unaligned-access) endif() if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv7") # Raspberry Pi 2 - add_compile_options(-mfpu=neon-fp-armv8 -mno-unaligned-access -funsafe-math-optimizations) + list(APPEND ARCH_FLAGS -mfpu=neon-fp-armv8 -mno-unaligned-access -funsafe-math-optimizations) endif() if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv8") # Raspberry Pi 3, 4, Zero 2 (32-bit) - add_compile_options(-mno-unaligned-access) + list(APPEND ARCH_FLAGS -mno-unaligned-access) endif() endif() elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "^(x86_64|i686|AMD64)$" OR "${CMAKE_GENERATOR_PLATFORM_LWR}" MATCHES "^(x86_64|i686|amd64|x64)$" ) @@ -659,7 +651,7 @@ elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "^(x86_64|i686|AMD64)$" OR "${CMAKE_GE include(cmake/FindSIMD.cmake) endif () if (LLAMA_AVX512) - add_compile_option_cpp(/arch:AVX512) + list(APPEND ARCH_FLAGS /arch:AVX512) # MSVC has no compile-time flags enabling specific # AVX512 extensions, neither it defines the # macros corresponding to the extensions. @@ -673,49 +665,61 @@ elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "^(x86_64|i686|AMD64)$" OR "${CMAKE_GE add_compile_definitions($<$:__AVX512VNNI__>) endif() elseif (LLAMA_AVX2) - add_compile_option_cpp(/arch:AVX2) + list(APPEND ARCH_FLAGS /arch:AVX2) elseif (LLAMA_AVX) - add_compile_option_cpp(/arch:AVX) + list(APPEND ARCH_FLAGS /arch:AVX) endif() else() if (LLAMA_NATIVE) - add_compile_option_cpp(-march=native) + list(APPEND ARCH_FLAGS -march=native) endif() if (LLAMA_F16C) - add_compile_option_cpp(-mf16c) + list(APPEND ARCH_FLAGS -mf16c) endif() if (LLAMA_FMA) - add_compile_option_cpp(-mfma) + list(APPEND ARCH_FLAGS -mfma) endif() if (LLAMA_AVX) - add_compile_option_cpp(-mavx) + list(APPEND ARCH_FLAGS -mavx) endif() if (LLAMA_AVX2) - add_compile_option_cpp(-mavx2) + list(APPEND ARCH_FLAGS -mavx2) endif() if (LLAMA_AVX512) - add_compile_option_cpp(-mavx512f) - add_compile_option_cpp(-mavx512bw) + list(APPEND ARCH_FLAGS -mavx512f) + list(APPEND ARCH_FLAGS -mavx512bw) endif() if (LLAMA_AVX512_VBMI) - add_compile_option_cpp(-mavx512vbmi) + list(APPEND ARCH_FLAGS -mavx512vbmi) endif() if (LLAMA_AVX512_VNNI) - add_compile_option_cpp(-mavx512vnni) + list(APPEND ARCH_FLAGS -mavx512vnni) endif() endif() elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64") message(STATUS "PowerPC detected") if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64le") - add_compile_options(-mcpu=powerpc64le) + list(APPEND ARCH_FLAGS -mcpu=powerpc64le) else() - add_compile_options(-mcpu=native -mtune=native) + list(APPEND ARCH_FLAGS -mcpu=native -mtune=native) #TODO: Add targets for Power8/Power9 (Altivec/VSX) and Power10(MMA) and query for big endian systems (ppc64/le/be) endif() else() message(STATUS "Unknown architecture") endif() +add_compile_options("$<$:${ARCH_FLAGS}>") +add_compile_options("$<$:${ARCH_FLAGS}>") + +if (LLAMA_CUBLAS) + list(APPEND CUDA_CXX_FLAGS ${ARCH_FLAGS}) + list(JOIN CUDA_CXX_FLAGS " " CUDA_CXX_FLAGS_JOINED) # pass host compiler flags as a single argument + if (NOT CUDA_CXX_FLAGS_JOINED STREQUAL "") + list(APPEND CUDA_FLAGS -Xcompiler ${CUDA_CXX_FLAGS_JOINED}) + endif() + add_compile_options("$<$:${CUDA_FLAGS}>") +endif() + if (MINGW) # Target Windows 8 for PrefetchVirtualMemory add_compile_definitions(_WIN32_WINNT=${LLAMA_WIN_VER}) From a1d6df129bcd3d42cda38c09217d8d4ec4ea3bdd Mon Sep 17 00:00:00 2001 From: 0cc4m Date: Fri, 26 Jan 2024 23:07:32 +0100 Subject: [PATCH 57/66] Add OpenCL add kernel (#5151) * Add OpenCL add kernel * Put add kernel into different string to stay within MSVC string length limit, disable float16 support due to bad results --- ggml-opencl.cpp | 87 +++++++++++++++++++++++++++++++++++++++++++++++-- ggml-opencl.h | 1 + ggml.c | 11 +++++++ 3 files changed, 96 insertions(+), 3 deletions(-) diff --git a/ggml-opencl.cpp b/ggml-opencl.cpp index 2bb93638f..bf9ad964f 100644 --- a/ggml-opencl.cpp +++ b/ggml-opencl.cpp @@ -714,7 +714,6 @@ __kernel void dequantize_mul_mat_vec_q6_K(__global const struct block_q6_K * xx, dst[row] = tmp[0]; } } - ); @@ -784,6 +783,7 @@ __kernel void KERNEL_NAME(__global X_TYPE* x, __local float* tmp, __global float dst[row] = tmp[0]; } } + ); @@ -799,6 +799,18 @@ __kernel void KERNEL_NAME(__global TYPE* x, const int x_offset, __global TYPE* y } ); +std::string add_template = MULTILINE_QUOTE( +__kernel void add_f32(__global float * x, const int x_offset, __global float * y, const int y_offset, __global float * dst, const int dst_offset, const int ky) { + const int i = get_group_id(0)*get_local_size(0) + get_local_id(0); + + if (i >= get_global_size(0)) { + return; + } + + dst[dst_offset + i] = x[x_offset + i] + y[y_offset + i%ky]; +} +); + #define CL_CHECK(err) \ do { \ cl_int err_ = (err); \ @@ -878,6 +890,7 @@ static std::string generate_kernels() { } src << mul_kernel << '\n'; } + src << add_template << '\n'; return src.str(); } @@ -893,6 +906,7 @@ static cl_kernel dequantize_mul_mat_vec_q4_0_cl, dequantize_mul_mat_vec_q4_1_cl, static cl_kernel dequantize_block_q2_k_cl, dequantize_block_q3_k_cl, dequantize_block_q4_k_cl, dequantize_block_q5_k_cl, dequantize_block_q6_k_cl; static cl_kernel dequantize_mul_mat_vec_q2_K_cl, dequantize_mul_mat_vec_q3_K_cl, dequantize_mul_mat_vec_q4_K_cl, dequantize_mul_mat_vec_q5_K_cl, dequantize_mul_mat_vec_q6_K_cl; static cl_kernel mul_f32_cl; +static cl_kernel add_f32_cl; static bool fp16_support; static cl_program build_program_from_source(cl_context ctx, cl_device_id dev, const char* program_buffer) { @@ -1100,9 +1114,10 @@ void ggml_cl_init(void) { char *ext_buffer = (char *)alloca(ext_str_size + 1); clGetDeviceInfo(device, CL_DEVICE_EXTENSIONS, ext_str_size, ext_buffer, NULL); ext_buffer[ext_str_size] = '\0'; // ensure it is null terminated + // Disabled due to faulty outputs // Check if ext_buffer contains cl_khr_fp16 - fp16_support = strstr(ext_buffer, "cl_khr_fp16") != NULL; - fprintf(stderr, "ggml_opencl: device FP16 support: %s\n", fp16_support ? "true" : "false"); + fp16_support = false; // strstr(ext_buffer, "cl_khr_fp16") != NULL; + // fprintf(stderr, "ggml_opencl: device FP16 support: %s\n", fp16_support ? "true" : "false"); cl_context_properties properties[] = { (intptr_t)CL_CONTEXT_PLATFORM, (intptr_t)platform, 0 @@ -1150,6 +1165,8 @@ void ggml_cl_init(void) { // mul kernel CL_CHECK((mul_f32_cl = clCreateKernel(program, "mul_f32", &err), err)); + + CL_CHECK((add_f32_cl = clCreateKernel(program, "add_f32", &err), err)); } static cl_kernel* ggml_get_to_fp32_cl(ggml_type type) { @@ -1458,6 +1475,70 @@ void ggml_cl_mul(const struct ggml_tensor * src0, const struct ggml_tensor * src ggml_cl_mul_f32(src0, src1, dst); } +static void ggml_cl_add_f32(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { + GGML_ASSERT(src1->backend == GGML_BACKEND_GPU); + const int64_t ne00 = src0->ne[0]; + const int64_t ne01 = src0->ne[1]; + const int64_t ne02 = src0->ne[2]; + const int64_t ne03 = src0->ne[3]; + const int64_t ne10 = src1->ne[0]; + const int64_t ne11 = src1->ne[1]; + const int64_t ne12 = src1->ne[2]; + const int64_t ne13 = src1->ne[3]; + const int nb2 = dst->nb[2]; + const int nb3 = dst->nb[3]; + size_t x_size; + size_t d_size; + + cl_mem d_X = ggml_cl_pool_malloc(ne00 * ne01 * sizeof(float), &x_size); // src0 + cl_mem d_Y = (cl_mem) src1->extra; // src1 is already on device, broadcasted. + cl_mem d_D = ggml_cl_pool_malloc(ne00 * ne01 * sizeof(float), &d_size); // dst + + + for (int64_t i03 = 0; i03 < ne03; i03++) { + for (int64_t i02 = 0; i02 < ne02; i02++) { + cl_event ev; + + // copy src0 to device + CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_X, 0, src0, i03, i02, &ev)); + + const int64_t i13 = i03%ne13; + const int64_t i12 = i02%ne12; + const int i1 = i13*ne12*ne11 + i12*ne11; + + cl_int x_offset = 0; + cl_int y_offset = i1*ne10; + cl_int d_offset = 0; + + size_t global = ne00 * ne01; + cl_int ky = ne10 * ne11; + + CL_CHECK(clSetKernelArg(add_f32_cl, 0, sizeof(cl_mem), &d_X)); + CL_CHECK(clSetKernelArg(add_f32_cl, 1, sizeof(cl_int), &x_offset)); + CL_CHECK(clSetKernelArg(add_f32_cl, 2, sizeof(cl_mem), &d_Y)); + CL_CHECK(clSetKernelArg(add_f32_cl, 3, sizeof(cl_int), &y_offset)); + CL_CHECK(clSetKernelArg(add_f32_cl, 4, sizeof(cl_mem), &d_D)); + CL_CHECK(clSetKernelArg(add_f32_cl, 5, sizeof(cl_int), &d_offset)); + CL_CHECK(clSetKernelArg(add_f32_cl, 6, sizeof(cl_int), &ky)); + CL_CHECK(clEnqueueNDRangeKernel(queue, add_f32_cl, 1, NULL, &global, NULL, 1, &ev, NULL)); + + CL_CHECK(clReleaseEvent(ev)); + CL_CHECK(clFinish(queue)); + + // copy dst to host + float * d = (float *) ((char *) dst->data + i02*nb2 + i03*nb3); + CL_CHECK(clEnqueueReadBuffer(queue, d_D, true, 0, sizeof(float) * ne00*ne01, d, 0, NULL, NULL)); + } + } + ggml_cl_pool_free(d_X, x_size); + ggml_cl_pool_free(d_D, d_size); +} + +void ggml_cl_add(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst) { + GGML_ASSERT(src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32); + ggml_cl_add_f32(src0, src1, dst); +} + static void ggml_cl_mul_mat_f32(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { const int64_t ne00 = src0->ne[0]; const int64_t ne01 = src0->ne[1]; diff --git a/ggml-opencl.h b/ggml-opencl.h index 919b00d63..257a6be6a 100644 --- a/ggml-opencl.h +++ b/ggml-opencl.h @@ -10,6 +10,7 @@ extern "C" { GGML_API void ggml_cl_init(void); GGML_API void ggml_cl_mul(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst); +GGML_API void ggml_cl_add(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst); GGML_API bool ggml_cl_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, const struct ggml_tensor * dst); GGML_API size_t ggml_cl_mul_mat_get_wsize(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst); GGML_API void ggml_cl_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst, void * wdata, size_t wsize); diff --git a/ggml.c b/ggml.c index ef6fd8caf..8f57003e0 100644 --- a/ggml.c +++ b/ggml.c @@ -7207,6 +7207,17 @@ static void ggml_compute_forward_add_f32( const int ith = params->ith; const int nth = params->nth; +#ifdef GGML_USE_CLBLAST + if (src1->backend == GGML_BACKEND_GPU) { + // TODO: OpenCL kernel support full broadcast + GGML_ASSERT(ggml_can_repeat_rows(src1, src0)); + if (ith == 0) { + ggml_cl_add(src0, src1, dst); + } + return; + } +#endif + const int nr = ggml_nrows(src0); GGML_TENSOR_BINARY_OP_LOCALS From ec903c034131848da9222536ff18da07ec0882a0 Mon Sep 17 00:00:00 2001 From: Maximilian Winter Date: Sat, 27 Jan 2024 14:38:05 +0100 Subject: [PATCH 58/66] server : add self-extend support (#5104) * Ported self extension to server example * Update server.cpp * Fixed prompt caching without self extend * Update server.cpp * Added description to server readme. * Update server.cpp * Update server.cpp * Update server.cpp * Update server.cpp * Update README.md * Changed descriptions * server : formatting * Update examples/server/server.cpp Co-authored-by: Georgi Gerganov * Update examples/server/server.cpp Co-authored-by: Georgi Gerganov * Update server.cpp * Update server.cpp --------- Co-authored-by: Georgi Gerganov --- examples/server/README.md | 3 +- examples/server/server.cpp | 172 +++++++++++++++++++++++++++++++------ 2 files changed, 147 insertions(+), 28 deletions(-) diff --git a/examples/server/README.md b/examples/server/README.md index fd3034b99..1c92a2041 100644 --- a/examples/server/README.md +++ b/examples/server/README.md @@ -30,7 +30,8 @@ Command line options: - `-cb`, `--cont-batching`: enable continuous batching (a.k.a dynamic batching) (default: disabled) - `-spf FNAME`, `--system-prompt-file FNAME` Set a file to load "a system prompt (initial prompt of all slots), this is useful for chat applications. [See more](#change-system-prompt-on-runtime) - `--mmproj MMPROJ_FILE`: Path to a multimodal projector file for LLaVA. - +- `--grp-attn-n`: Set the group attention factor to extend context size through self-extend(default: 1=disabled), used together with group attention width `--grp-attn-w` +- `--grp-attn-w`: Set the group attention width to extend context size through self-extend(default: 512), used together with group attention factor `--grp-attn-n` ## Build server is build alongside everything else from the root of the project diff --git a/examples/server/server.cpp b/examples/server/server.cpp index 392836132..af63f2f6f 100644 --- a/examples/server/server.cpp +++ b/examples/server/server.cpp @@ -184,6 +184,12 @@ struct llama_client_slot struct llama_sampling_params sparams; llama_sampling_context *ctx_sampling = nullptr; + int32_t ga_i = 0; // group-attention state + int32_t ga_n = 1;// group-attention factor + int32_t ga_w = 512; // group-attention width + + int32_t n_past_se = 0; // self-extend + // multimodal std::vector images; @@ -212,7 +218,8 @@ struct llama_client_slot sent_count = 0; sent_token_probs_index = 0; infill = false; - + ga_i = 0; + n_past_se = 0; generated_token_probs.clear(); for (slot_image & img : images) @@ -399,9 +406,26 @@ struct llama_server_context slot.id = i; slot.n_ctx = n_ctx_slot; - slot.reset(); LOG_TEE(" -> Slot %i - max context: %i\n", slot.id, n_ctx_slot); + + const int ga_n = params.grp_attn_n; + const int ga_w = params.grp_attn_w; + + if (ga_n != 1) { + GGML_ASSERT(ga_n > 0 && "ga_n must be positive"); // NOLINT + GGML_ASSERT(ga_w % ga_n == 0 && "ga_w must be a multiple of ga_n"); // NOLINT + //GGML_ASSERT(n_ctx_train % ga_w == 0 && "n_ctx_train must be a multiple of ga_w"); // NOLINT + //GGML_ASSERT(n_ctx >= n_ctx_train * ga_n && "n_ctx must be at least n_ctx_train * ga_n"); // NOLINT + LOG_TEE(" -> Slot %i - self-extend: ga_n = %d, ga_w = %d\n", slot.id, ga_n, ga_w); + } + + slot.ga_i = 0; + slot.ga_n = ga_n; + slot.ga_w = ga_w; + + slot.reset(); + slots.push_back(slot); } @@ -1349,32 +1373,35 @@ struct llama_server_context for (llama_client_slot &slot : slots) { - if (slot.is_processing() && slot.cache_tokens.size() >= (size_t) slot.n_ctx) + if (slot.ga_n == 1) { - // Shift context - const int n_left = slot.n_past - slot.params.n_keep - 1; - const int n_discard = n_left / 2; - - LOG_TEE("slot %d: context shift - n_keep = %d, n_left = %d, n_discard = %d\n", slot.id, slot.params.n_keep, n_left, n_discard); - llama_kv_cache_seq_rm (ctx, slot.id, slot.params.n_keep + 1 , slot.params.n_keep + n_discard + 1); - llama_kv_cache_seq_shift(ctx, slot.id, slot.params.n_keep + 1 + n_discard, slot.n_past, -n_discard); - - for (size_t i = slot.params.n_keep + 1 + n_discard; i < slot.cache_tokens.size(); i++) + if (slot.is_processing() && slot.cache_tokens.size() >= (size_t) slot.n_ctx) { - slot.cache_tokens[i - n_discard] = slot.cache_tokens[i]; + // Shift context + const int n_left = slot.n_past - slot.params.n_keep - 1; + const int n_discard = n_left / 2; + + LOG_TEE("slot %d: context shift - n_keep = %d, n_left = %d, n_discard = %d\n", slot.id, slot.params.n_keep, n_left, n_discard); + llama_kv_cache_seq_rm (ctx, slot.id, slot.params.n_keep + 1 , slot.params.n_keep + n_discard + 1); + llama_kv_cache_seq_shift(ctx, slot.id, slot.params.n_keep + 1 + n_discard, slot.n_past, -n_discard); + + for (size_t i = slot.params.n_keep + 1 + n_discard; i < slot.cache_tokens.size(); i++) + { + slot.cache_tokens[i - n_discard] = slot.cache_tokens[i]; + } + + slot.cache_tokens.resize(slot.cache_tokens.size() - n_discard); + + slot.n_past -= n_discard; + + slot.truncated = true; + + LOG_VERBOSE("context shift", { + { "n_ctx", n_ctx }, + { "n_keep", params.n_keep }, + { "n_left", n_left }, + }); } - - slot.cache_tokens.resize(slot.cache_tokens.size() - n_discard); - - slot.n_past -= n_discard; - - slot.truncated = true; - - LOG_VERBOSE("context shift", { - {"n_ctx", n_ctx}, - {"n_keep", params.n_keep}, - {"n_left", n_left}, - }); } } @@ -1401,7 +1428,8 @@ struct llama_server_context slot.i_batch = batch.n_tokens; - llama_batch_add(batch, slot.sampled, system_tokens.size() + slot.n_past, { slot.id }, true); + const int32_t slot_npast = slot.n_past_se > 0 ? slot.n_past_se : slot.n_past; + llama_batch_add(batch, slot.sampled, system_tokens.size() + slot_npast, { slot.id }, true); slot.n_past += 1; } @@ -1499,6 +1527,8 @@ struct llama_server_context llama_sampling_reset(slot.ctx_sampling); slot.n_past = 0; + slot.n_past_se = 0; + slot.ga_i = 0; slot.num_prompt_tokens_processed = slot.num_prompt_tokens; } else @@ -1512,6 +1542,25 @@ struct llama_server_context slot.n_past = common_part(slot.cache_tokens, prompt_tokens); slot.num_prompt_tokens_processed = slot.num_prompt_tokens - slot.n_past; + if (slot.ga_n != 1) + { + int ga_i = 0; + int32_t ga_n = slot.ga_n; + int32_t ga_w = slot.ga_w; + int32_t slot_npast = 0; + for (int k = 0; k < slot.n_past; ++k) + { + while (slot_npast >= ga_i + ga_w) { + const int bd = (ga_w/ga_n)*(ga_n - 1); + slot_npast -= bd; + ga_i += ga_w/ga_n; + } + slot_npast++; + } + slot.n_past_se = slot_npast; + slot.ga_i = ga_i; + } + LOG_TEE("slot %d : in cache: %i tokens | to process: %i tokens\n", slot.id, slot.n_past, slot.num_prompt_tokens_processed); } @@ -1526,6 +1575,10 @@ struct llama_server_context // we have to evaluate at least 1 token to generate logits. LOG_TEE("slot %d : we have to evaluate at least 1 token to generate logits\n", slot.id); slot.n_past--; + if (slot.ga_i > 0) + { + slot.n_past_se--; + } } LOG_VERBOSE("prompt ingested", { @@ -1538,9 +1591,22 @@ struct llama_server_context // process the prefix of first image std::vector prefix_tokens = has_images ? tokenize(slot.images[0].prefix_prompt, add_bos_token) : prompt_tokens; + int32_t slot_npast = slot.n_past_se > 0 ? slot.n_past_se : slot.n_past; + int ga_i = slot.ga_i; + int32_t ga_n = slot.ga_n; + int32_t ga_w = slot.ga_w; for (; slot.n_past < (int) prefix_tokens.size(); ++slot.n_past) { - llama_batch_add(batch, prefix_tokens[slot.n_past], system_tokens.size() + slot.n_past, { slot.id }, false); + if (slot.ga_n != 1) + { + while (slot_npast >= ga_i + ga_w) { + const int bd = (ga_w/ga_n)*(ga_n - 1); + slot_npast -= bd; + ga_i += ga_w/ga_n; + } + } + llama_batch_add(batch, prefix_tokens[slot.n_past], system_tokens.size() + slot_npast, {slot.id }, false); + slot_npast += 1; } if (has_images && !ingest_images(slot, n_batch)) @@ -1570,6 +1636,36 @@ struct llama_server_context for (int32_t i = 0; i < (int32_t) batch.n_tokens; i += n_batch) { const int32_t n_tokens = std::min(n_batch, (int32_t) (batch.n_tokens - i)); + + for (auto & slot : slots) + { + if (slot.ga_n != 1) + { + // context extension via Self-Extend + while (slot.n_past_se >= slot.ga_i + slot.ga_w) + { + const int ib = (slot.ga_n * slot.ga_i) / slot.ga_w; + const int bd = (slot.ga_w / slot.ga_n) * (slot.ga_n - 1); + const int dd = (slot.ga_w / slot.ga_n) - ib * bd - slot.ga_w; + + LOG_TEE("\n"); + LOG_TEE("shift: [%6d, %6d] + %6d -> [%6d, %6d]\n", slot.ga_i, slot.n_past_se, ib * bd, slot.ga_i + ib * bd, slot.n_past_se + ib * bd); + LOG_TEE("div: [%6d, %6d] / %6d -> [%6d, %6d]\n", slot.ga_i + ib * bd, slot.ga_i + ib * bd + slot.ga_w, slot.ga_n, (slot.ga_i + ib * bd) / slot.ga_n, (slot.ga_i + ib * bd + slot.ga_w) / slot.ga_n); + LOG_TEE("shift: [%6d, %6d] + %6d -> [%6d, %6d]\n", slot.ga_i + ib * bd + slot.ga_w, slot.n_past_se + ib * bd, dd, slot.ga_i + ib * bd + slot.ga_w + dd, slot.n_past_se + ib * bd + dd); + + llama_kv_cache_seq_shift(ctx, slot.id, slot.ga_i, slot.n_past_se, ib * bd); + llama_kv_cache_seq_div(ctx, slot.id, slot.ga_i + ib * bd, slot.ga_i + ib * bd + slot.ga_w,slot.ga_n); + llama_kv_cache_seq_shift(ctx, slot.id, slot.ga_i + ib * bd + slot.ga_w,slot.n_past_se + ib * bd, dd); + + slot.n_past_se -= bd; + + slot.ga_i += slot.ga_w / slot.ga_n; + + LOG_TEE("\nn_past_old = %d, n_past = %d, ga_i = %d\n\n", slot.n_past_se + bd, slot.n_past_se, slot.ga_i); + } + slot.n_past_se += n_tokens; + } + } llama_batch batch_view = { n_tokens, @@ -1583,6 +1679,7 @@ struct llama_server_context }; const int ret = llama_decode(ctx, batch_view); + if (ret != 0) { if (n_batch == 1 || ret < 0) @@ -1728,6 +1825,8 @@ static void server_print_usage(const char *argv0, const gpt_params ¶ms, printf(" --override-kv KEY=TYPE:VALUE\n"); printf(" advanced option to override model metadata by key. may be specified multiple times.\n"); printf(" types: int, float, bool. example: --override-kv tokenizer.ggml.add_bos_token=bool:false\n"); + printf(" -gan N, --grp-attn-n N Set the group attention factor to extend context size through self-extend(default: 1=disabled), used together with group attention width `--grp-attn-w`"); + printf(" -gaw N, --grp-attn-w N Set the group attention width to extend context size through self-extend(default: 512), used together with group attention factor `--grp-attn-n`"); printf("\n"); } @@ -1913,6 +2012,25 @@ static void server_params_parse(int argc, char **argv, server_params &sparams, } params.n_threads = std::stoi(argv[i]); } + else if (arg == "--grp-attn-n" || arg == "-gan") + { + if (++i >= argc) { + invalid_param = true; + break; + } + + params.grp_attn_n = std::stoi(argv[i]); + } + else if (arg == "--grp-attn-w" || arg == "-gaw") + { + if (++i >= argc) + { + invalid_param = true; + break; + } + + params.grp_attn_w = std::stoi(argv[i]); + } else if (arg == "--threads-batch" || arg == "-tb") { if (++i >= argc) From 35a2ee914308c85ab5cb576467381443ad23f0ac Mon Sep 17 00:00:00 2001 From: Michael Klimenko Date: Sat, 27 Jan 2024 15:25:55 +0100 Subject: [PATCH 59/66] Remove unused data and add fixes (#5154) * Remove unused data and add fixes * Add missing file * Address review comments * Replace the scope of vq allocation --- common/sampling.cpp | 1 + examples/infill/infill.cpp | 2 +- examples/llava/clip.cpp | 1 - examples/server/server.cpp | 2 +- pocs/vdot/vdot.cpp | 1 - tests/test-backend-ops.cpp | 1 - tests/test-llama-grammar.cpp | 1 - 7 files changed, 3 insertions(+), 6 deletions(-) diff --git a/common/sampling.cpp b/common/sampling.cpp index efd7eab6e..e8675a8c0 100644 --- a/common/sampling.cpp +++ b/common/sampling.cpp @@ -13,6 +13,7 @@ struct llama_sampling_context * llama_sampling_init(const struct llama_sampling_ // will be empty (default) if there are parse errors if (result->parsed_grammar.rules.empty()) { fprintf(stderr, "%s: failed to parse grammar\n", __func__); + delete result; return nullptr; } diff --git a/examples/infill/infill.cpp b/examples/infill/infill.cpp index 4a7827876..72fb133b4 100644 --- a/examples/infill/infill.cpp +++ b/examples/infill/infill.cpp @@ -241,7 +241,7 @@ int main(int argc, char ** argv) { LOG("add_bos: %d\n", add_bos); bool suff_rm_leading_spc = params.escape; - if (suff_rm_leading_spc && params.input_suffix.find_first_of(" ") == 0 && params.input_suffix.size() > 1) { + if (suff_rm_leading_spc && params.input_suffix.find_first_of(' ') == 0 && params.input_suffix.size() > 1) { params.input_suffix.erase(0, 1); suff_rm_leading_spc = false; } diff --git a/examples/llava/clip.cpp b/examples/llava/clip.cpp index 4a0338a37..f2cd86afe 100644 --- a/examples/llava/clip.cpp +++ b/examples/llava/clip.cpp @@ -1277,7 +1277,6 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i ".*weight", }; - std::vector read_data(512); std::vector work(512); std::vector conv_buf(512); std::vector hist_all(1 << 4, 0); diff --git a/examples/server/server.cpp b/examples/server/server.cpp index af63f2f6f..f58a2acaa 100644 --- a/examples/server/server.cpp +++ b/examples/server/server.cpp @@ -681,7 +681,7 @@ struct llama_server_context while ((pos = prompt.find(pattern, pos)) != std::string::npos) { size_t end_prefix = pos; pos += pattern.length(); - size_t end_pos = prompt.find("]", pos); + size_t end_pos = prompt.find(']', pos); if (end_pos != std::string::npos) { std::string image_id = prompt.substr(pos, end_pos - pos); diff --git a/pocs/vdot/vdot.cpp b/pocs/vdot/vdot.cpp index e96372c4b..73ffcd1ca 100644 --- a/pocs/vdot/vdot.cpp +++ b/pocs/vdot/vdot.cpp @@ -243,7 +243,6 @@ int main(int argc, char** argv) { if (useQ4_1) q41.resize(n4); else q40.resize(n4); std::vector q8(n8); - std::vector H(16, 0); double sumt = 0, sumt2 = 0, maxt = 0; double sumqt = 0, sumqt2 = 0, maxqt = 0; double sum = 0, sumq = 0, exactSum = 0; diff --git a/tests/test-backend-ops.cpp b/tests/test-backend-ops.cpp index 55ce14e0d..e3c656f56 100644 --- a/tests/test-backend-ops.cpp +++ b/tests/test-backend-ops.cpp @@ -102,7 +102,6 @@ static std::vector tensor_to_float(const ggml_tensor * t) { } else if (t->type == GGML_TYPE_I8) { tv.push_back((float)*(int8_t *) &buf[i]); } else if (quantized) { - std::vector vq(ggml_blck_size(t->type)); tt.to_float(&buf[i], vq.data(), ggml_blck_size(t->type)); tv.insert(tv.end(), vq.begin(), vq.end()); } else { diff --git a/tests/test-llama-grammar.cpp b/tests/test-llama-grammar.cpp index 73dd33dd2..78fc41117 100644 --- a/tests/test-llama-grammar.cpp +++ b/tests/test-llama-grammar.cpp @@ -190,7 +190,6 @@ int main() index++; } - std::vector> next_stacks; std::vector next_candidates; next_candidates.resize(24); From e9764230054e01553bdead6f2bfd8e001869599d Mon Sep 17 00:00:00 2001 From: Judd Date: Fri, 26 Jan 2024 21:04:01 +0800 Subject: [PATCH 60/66] ggml : check ggml_add src1 type (ggml/708) Co-authored-by: Judd --- ggml.c | 7 ++++++- 1 file changed, 6 insertions(+), 1 deletion(-) diff --git a/ggml.c b/ggml.c index 8f57003e0..11a3114e7 100644 --- a/ggml.c +++ b/ggml.c @@ -7498,7 +7498,12 @@ static void ggml_compute_forward_add( switch (src0->type) { case GGML_TYPE_F32: { - ggml_compute_forward_add_f32(params, src0, src1, dst); + if (src1->type == GGML_TYPE_F32) { + ggml_compute_forward_add_f32(params, src0, src1, dst); + } + else { + GGML_ASSERT(false); + } } break; case GGML_TYPE_F16: { From 753eafed0ebd07af6903771327a1786a7c02cf98 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Sat, 27 Jan 2024 16:59:20 +0200 Subject: [PATCH 61/66] sync : ggml --- examples/server/utils.hpp | 1 + scripts/sync-ggml.last | 2 +- 2 files changed, 2 insertions(+), 1 deletion(-) diff --git a/examples/server/utils.hpp b/examples/server/utils.hpp index e2b6065f7..70cce0721 100644 --- a/examples/server/utils.hpp +++ b/examples/server/utils.hpp @@ -249,6 +249,7 @@ struct llama_server_queue { } // Start the main loop. This call is blocking + [[noreturn]] void start_loop() { while (true) { // new task arrived diff --git a/scripts/sync-ggml.last b/scripts/sync-ggml.last index 4d52d946b..efde0069f 100644 --- a/scripts/sync-ggml.last +++ b/scripts/sync-ggml.last @@ -1 +1 @@ -6c1ce0bd591a430c1d3f6797d905194581c878c1 +c2448f88d17395452a587d0176d19ed87e0f7ce1 From 6db2b41a76ee78d5efdd5c3cddd5d7ad3f646855 Mon Sep 17 00:00:00 2001 From: John <78893154+cmp-nct@users.noreply.github.com> Date: Sat, 27 Jan 2024 16:09:18 +0100 Subject: [PATCH 62/66] llava : support for Yi-VL and fix for mobileVLM (#5093) * Support for Yi-VL, templating fix for mobileVLM * ws * Update examples/llava/clip.cpp Co-authored-by: Georgi Gerganov * Update llava-cli.cpp * Update clip.cpp bugfix for new conversions --------- Co-authored-by: Georgi Gerganov --- examples/llava/clip.cpp | 70 ++++++++++++++++++++++++++++++++---- examples/llava/llava-cli.cpp | 32 +++++++++++++++-- 2 files changed, 92 insertions(+), 10 deletions(-) diff --git a/examples/llava/clip.cpp b/examples/llava/clip.cpp index f2cd86afe..9129052a2 100644 --- a/examples/llava/clip.cpp +++ b/examples/llava/clip.cpp @@ -98,6 +98,7 @@ static std::string format(const char * fmt, ...) { enum projector_type { PROJECTOR_TYPE_MLP, + PROJECTOR_TYPE_MLP_NORM, PROJECTOR_TYPE_LDP, PROJECTOR_TYPE_UNKNOWN, }; @@ -304,10 +305,18 @@ struct clip_vision_model { struct ggml_tensor * projection; // LLaVA projection - struct ggml_tensor * mm_0_w; - struct ggml_tensor * mm_0_b; - struct ggml_tensor * mm_2_w; - struct ggml_tensor * mm_2_b; + struct ggml_tensor * mm_0_w = NULL; + struct ggml_tensor * mm_0_b = NULL; + struct ggml_tensor * mm_2_w = NULL; + struct ggml_tensor * mm_2_b = NULL; + + // Yi type models with mlp+normalization projection + struct ggml_tensor * mm_1_w = NULL; // Yi type models have 0, 1, 3, 4 + struct ggml_tensor * mm_1_b = NULL; + struct ggml_tensor * mm_3_w = NULL; + struct ggml_tensor * mm_3_b = NULL; + struct ggml_tensor * mm_4_w = NULL; + struct ggml_tensor * mm_4_b = NULL; // MobileVLM projection struct ggml_tensor * mm_model_mlp_1_w; @@ -460,6 +469,7 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32 // pre-layernorm { embeddings = ggml_norm(ctx0, embeddings, eps); + ggml_set_name(embeddings, "pre_ln"); embeddings = ggml_add(ctx0, ggml_mul(ctx0, embeddings, model.pre_ln_w), model.pre_ln_b); } @@ -575,6 +585,27 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32 embeddings = ggml_mul_mat(ctx0, model.mm_2_w, embeddings); embeddings = ggml_add(ctx0, embeddings, model.mm_2_b); + + } else if (ctx->proj_type == PROJECTOR_TYPE_MLP_NORM) { + embeddings = ggml_mul_mat(ctx0, model.mm_0_w, embeddings); + embeddings = ggml_add(ctx0, embeddings, model.mm_0_b); + // ggml_tensor_printf(embeddings, "mm_0_w",0,true,false); + // First LayerNorm + embeddings = ggml_norm(ctx0, embeddings, eps); + embeddings = ggml_add(ctx0, ggml_mul(ctx0, embeddings, model.mm_1_w), + model.mm_1_b); + + // GELU activation + embeddings = ggml_gelu(ctx0, embeddings); + + // Second linear layer + embeddings = ggml_mul_mat(ctx0, model.mm_3_w, embeddings); + embeddings = ggml_add(ctx0, embeddings, model.mm_3_b); + + // Second LayerNorm + embeddings = ggml_norm(ctx0, embeddings, eps); + embeddings = ggml_add(ctx0, ggml_mul(ctx0, embeddings, model.mm_4_w), + model.mm_4_b); } else if (ctx->proj_type == PROJECTOR_TYPE_LDP) { // MobileVLM projector @@ -808,6 +839,11 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) { else { new_clip->proj_type = PROJECTOR_TYPE_MLP; } + if (new_clip->proj_type == PROJECTOR_TYPE_MLP) { + if (gguf_find_tensor(ctx, format(TN_LLAVA_PROJ, 3, "weight").c_str()) != -1) { + new_clip->proj_type = PROJECTOR_TYPE_MLP_NORM; + } + } } #ifdef GGML_USE_CUBLAS @@ -956,11 +992,29 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) { vision_model.pre_ln_b = get_tensor(new_clip->ctx_data, format(TN_LN_PRE, "v", "bias")); // LLaVA projection - if (new_clip->proj_type == PROJECTOR_TYPE_MLP) { + if (new_clip->proj_type == PROJECTOR_TYPE_MLP || new_clip->proj_type == PROJECTOR_TYPE_MLP_NORM) { vision_model.mm_0_w = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 0, "weight")); vision_model.mm_0_b = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 0, "bias")); - vision_model.mm_2_w = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 2, "weight")); - vision_model.mm_2_b = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 2, "bias")); + try { + // Yi-type llava + vision_model.mm_1_w = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 1, "weight")); + vision_model.mm_1_b = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 1, "bias")); + } catch (std::runtime_error & e) { } + try { + // missing in Yi-type llava + vision_model.mm_2_w = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 2, "weight")); + vision_model.mm_2_b = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 2, "bias")); + } catch (std::runtime_error & e) { } + try { + // Yi-type llava + vision_model.mm_3_w = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 3, "weight")); + vision_model.mm_3_b = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 3, "bias")); + } catch (std::runtime_error & e) { } + try { + // Yi-type llava + vision_model.mm_4_w = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 4, "weight")); + vision_model.mm_4_b = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 4, "bias")); + } catch (std::runtime_error & e) { } } else if (new_clip->proj_type == PROJECTOR_TYPE_LDP) { // MobileVLM projection @@ -1432,6 +1486,8 @@ int clip_n_mmproj_embd(const struct clip_ctx * ctx) { } else if (ctx->proj_type == PROJECTOR_TYPE_MLP) { return ctx->vision_model.mm_2_b->ne[0]; + } else if (ctx->proj_type == PROJECTOR_TYPE_MLP_NORM) { + return ctx->vision_model.mm_3_b->ne[0]; } else { std::string proj_type = PROJECTOR_TYPE_NAMES[ctx->proj_type]; diff --git a/examples/llava/llava-cli.cpp b/examples/llava/llava-cli.cpp index d94795fe3..6ac70ba69 100644 --- a/examples/llava/llava-cli.cpp +++ b/examples/llava/llava-cli.cpp @@ -148,10 +148,35 @@ static void process_prompt(struct llava_context * ctx_llava, struct llava_image_ const int max_tgt_len = params->n_predict < 0 ? 256 : params->n_predict; const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx_llava->ctx_llama)); - // llava chat format is "\nUSER:\n\nASSISTANT:" - eval_string(ctx_llava->ctx_llama, "A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions.\nUSER:", params->n_batch, &n_past, add_bos); + std::string system_prompt, user_prompt; + size_t image_pos = prompt.find(""); + if (image_pos != std::string::npos) { + // new templating mode: Provide the full prompt including system message and use as a placeholder for the image + + system_prompt = prompt.substr(0, image_pos); + user_prompt = prompt.substr(image_pos + std::string("").length()); + // We replace \n with actual newlines in user_prompt, just in case -e was not used in templating string + size_t pos = 0; + while ((pos = user_prompt.find("\\n", pos)) != std::string::npos) { + user_prompt.replace(pos, 2, "\n"); + pos += 1; // Advance past the replaced newline + } + while ((pos = system_prompt.find("\\n", pos)) != std::string::npos) { + system_prompt.replace(pos, 2, "\n"); + pos += 1; // Advance past the replaced newline + } + + printf("system_prompt: %s\n", system_prompt.c_str()); + printf("user_prompt: %s\n", user_prompt.c_str()); + } else { + // llava-1.5 native mode + system_prompt = "A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions.\nUSER:"; + user_prompt = prompt + "\nASSISTANT:"; + } + + eval_string(ctx_llava->ctx_llama, system_prompt.c_str(), params->n_batch, &n_past, add_bos); llava_eval_image_embed(ctx_llava->ctx_llama, image_embed, params->n_batch, &n_past); - eval_string(ctx_llava->ctx_llama, (prompt + "\nASSISTANT:").c_str(), params->n_batch, &n_past, false); + eval_string(ctx_llava->ctx_llama, user_prompt.c_str(), params->n_batch, &n_past, false); // generate the response @@ -162,6 +187,7 @@ static void process_prompt(struct llava_context * ctx_llava, struct llava_image_ for (int i = 0; i < max_tgt_len; i++) { const char * tmp = sample(ctx_sampling, ctx_llava->ctx_llama, &n_past); if (strcmp(tmp, "") == 0) break; + if (strstr(tmp, "###")) break; // Yi-VL behavior printf("%s", tmp); fflush(stdout); From 39baaf55a160909bb9428bd981014218761a20cb Mon Sep 17 00:00:00 2001 From: Kyle Mistele Date: Sun, 28 Jan 2024 01:55:31 -0600 Subject: [PATCH 63/66] docker : add server-first container images (#5157) * feat: add Dockerfiles for each platform that user ./server instead of ./main * feat: update .github/workflows/docker.yml to build server-first docker containers * doc: add information about running the server with Docker to README.md * doc: add information about running with docker to the server README * doc: update n-gpu-layers to show correct GPU usage * fix(doc): update container tag from `server` to `server-cuda` for README example on running server container with CUDA --- .devops/server-cuda.Dockerfile | 32 +++++++++++++++++++++++ .devops/server-intel.Dockerfile | 25 ++++++++++++++++++ .devops/server-rocm.Dockerfile | 45 +++++++++++++++++++++++++++++++++ .devops/server.Dockerfile | 20 +++++++++++++++ .github/workflows/docker.yml | 4 +++ README.md | 14 +++++++++- examples/server/README.md | 8 ++++++ 7 files changed, 147 insertions(+), 1 deletion(-) create mode 100644 .devops/server-cuda.Dockerfile create mode 100644 .devops/server-intel.Dockerfile create mode 100644 .devops/server-rocm.Dockerfile create mode 100644 .devops/server.Dockerfile diff --git a/.devops/server-cuda.Dockerfile b/.devops/server-cuda.Dockerfile new file mode 100644 index 000000000..4f83904bc --- /dev/null +++ b/.devops/server-cuda.Dockerfile @@ -0,0 +1,32 @@ +ARG UBUNTU_VERSION=22.04 +# This needs to generally match the container host's environment. +ARG CUDA_VERSION=11.7.1 +# Target the CUDA build image +ARG BASE_CUDA_DEV_CONTAINER=nvidia/cuda:${CUDA_VERSION}-devel-ubuntu${UBUNTU_VERSION} +# Target the CUDA runtime image +ARG BASE_CUDA_RUN_CONTAINER=nvidia/cuda:${CUDA_VERSION}-runtime-ubuntu${UBUNTU_VERSION} + +FROM ${BASE_CUDA_DEV_CONTAINER} as build + +# Unless otherwise specified, we make a fat build. +ARG CUDA_DOCKER_ARCH=all + +RUN apt-get update && \ + apt-get install -y build-essential git + +WORKDIR /app + +COPY . . + +# Set nvcc architecture +ENV CUDA_DOCKER_ARCH=${CUDA_DOCKER_ARCH} +# Enable cuBLAS +ENV LLAMA_CUBLAS=1 + +RUN make + +FROM ${BASE_CUDA_RUN_CONTAINER} as runtime + +COPY --from=build /app/server /server + +ENTRYPOINT [ "/server" ] diff --git a/.devops/server-intel.Dockerfile b/.devops/server-intel.Dockerfile new file mode 100644 index 000000000..e343d278c --- /dev/null +++ b/.devops/server-intel.Dockerfile @@ -0,0 +1,25 @@ +ARG ONEAPI_VERSION=2024.0.1-devel-ubuntu22.04 +ARG UBUNTU_VERSION=22.04 + +FROM intel/hpckit:$ONEAPI_VERSION as build + +RUN apt-get update && \ + apt-get install -y git + +WORKDIR /app + +COPY . . + +# for some reasons, "-DLLAMA_BLAS=ON -DLLAMA_BLAS_VENDOR=Intel10_64lp -DLLAMA_NATIVE=ON" give worse performance +RUN mkdir build && \ + cd build && \ + cmake .. -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx && \ + cmake --build . --config Release --target main server + +FROM ubuntu:$UBUNTU_VERSION as runtime + +COPY --from=build /app/build/bin/server /server + +ENV LC_ALL=C.utf8 + +ENTRYPOINT [ "/server" ] diff --git a/.devops/server-rocm.Dockerfile b/.devops/server-rocm.Dockerfile new file mode 100644 index 000000000..e9a31647c --- /dev/null +++ b/.devops/server-rocm.Dockerfile @@ -0,0 +1,45 @@ +ARG UBUNTU_VERSION=22.04 + +# This needs to generally match the container host's environment. +ARG ROCM_VERSION=5.6 + +# Target the CUDA build image +ARG BASE_ROCM_DEV_CONTAINER=rocm/dev-ubuntu-${UBUNTU_VERSION}:${ROCM_VERSION}-complete + +FROM ${BASE_ROCM_DEV_CONTAINER} as build + +# Unless otherwise specified, we make a fat build. +# List from https://github.com/ggerganov/llama.cpp/pull/1087#issuecomment-1682807878 +# This is mostly tied to rocBLAS supported archs. +ARG ROCM_DOCKER_ARCH=\ + gfx803 \ + gfx900 \ + gfx906 \ + gfx908 \ + gfx90a \ + gfx1010 \ + gfx1030 \ + gfx1100 \ + gfx1101 \ + gfx1102 + +COPY requirements.txt requirements.txt +COPY requirements requirements + +RUN pip install --upgrade pip setuptools wheel \ + && pip install -r requirements.txt + +WORKDIR /app + +COPY . . + +# Set nvcc architecture +ENV GPU_TARGETS=${ROCM_DOCKER_ARCH} +# Enable ROCm +ENV LLAMA_HIPBLAS=1 +ENV CC=/opt/rocm/llvm/bin/clang +ENV CXX=/opt/rocm/llvm/bin/clang++ + +RUN make + +ENTRYPOINT [ "/app/server" ] diff --git a/.devops/server.Dockerfile b/.devops/server.Dockerfile new file mode 100644 index 000000000..134588fe2 --- /dev/null +++ b/.devops/server.Dockerfile @@ -0,0 +1,20 @@ +ARG UBUNTU_VERSION=22.04 + +FROM ubuntu:$UBUNTU_VERSION as build + +RUN apt-get update && \ + apt-get install -y build-essential git + +WORKDIR /app + +COPY . . + +RUN make + +FROM ubuntu:$UBUNTU_VERSION as runtime + +COPY --from=build /app/server /server + +ENV LC_ALL=C.utf8 + +ENTRYPOINT [ "/server" ] diff --git a/.github/workflows/docker.yml b/.github/workflows/docker.yml index 825b8f503..94f9161fc 100644 --- a/.github/workflows/docker.yml +++ b/.github/workflows/docker.yml @@ -28,14 +28,18 @@ jobs: config: - { tag: "light", dockerfile: ".devops/main.Dockerfile", platforms: "linux/amd64,linux/arm64" } - { tag: "full", dockerfile: ".devops/full.Dockerfile", platforms: "linux/amd64,linux/arm64" } + - { tag: "server", dockerfile: ".devops/server.Dockerfile", platforms: "linux/amd64,linux/arm64" } # NOTE(canardletter): The CUDA builds on arm64 are very slow, so I # have disabled them for now until the reason why # is understood. - { tag: "light-cuda", dockerfile: ".devops/main-cuda.Dockerfile", platforms: "linux/amd64" } - { tag: "full-cuda", dockerfile: ".devops/full-cuda.Dockerfile", platforms: "linux/amd64" } + - { tag: "server-cuda", dockerfile: ".devops/server-cuda.Dockerfile", platforms: "linux/amd64" } - { tag: "light-rocm", dockerfile: ".devops/main-rocm.Dockerfile", platforms: "linux/amd64,linux/arm64" } - { tag: "full-rocm", dockerfile: ".devops/full-rocm.Dockerfile", platforms: "linux/amd64,linux/arm64" } + - { tag: "server-rocm", dockerfile: ".devops/server-rocm.Dockerfile", platforms: "linux/amd64,linux/arm64" } - { tag: "light-intel", dockerfile: ".devops/main-intel.Dockerfile", platforms: "linux/amd64" } + - { tag: "server-intel", dockerfile: ".devops/server-intel.Dockerfile", platforms: "linux/amd64" } steps: - name: Check out the repo uses: actions/checkout@v3 diff --git a/README.md b/README.md index 76e48ce8a..cd95f8144 100644 --- a/README.md +++ b/README.md @@ -931,17 +931,20 @@ Place your desired model into the `~/llama.cpp/models/` directory and execute th * Create a folder to store big models & intermediate files (ex. /llama/models) #### Images -We have two Docker images available for this project: +We have three Docker images available for this project: 1. `ghcr.io/ggerganov/llama.cpp:full`: This image includes both the main executable file and the tools to convert LLaMA models into ggml and convert into 4-bit quantization. (platforms: `linux/amd64`, `linux/arm64`) 2. `ghcr.io/ggerganov/llama.cpp:light`: This image only includes the main executable file. (platforms: `linux/amd64`, `linux/arm64`) +3. `ghcr.io/ggerganov/llama.cpp:server`: This image only includes the server executabhle file. (platforms: `linux/amd64`, `linux/arm64`) Additionally, there the following images, similar to the above: - `ghcr.io/ggerganov/llama.cpp:full-cuda`: Same as `full` but compiled with CUDA support. (platforms: `linux/amd64`) - `ghcr.io/ggerganov/llama.cpp:light-cuda`: Same as `light` but compiled with CUDA support. (platforms: `linux/amd64`) +- `ghcr.io/ggerganov/llama.cpp:server-cuda`: Same as `server` but compiled with CUDA support. (platforms: `linux/amd64`) - `ghcr.io/ggerganov/llama.cpp:full-rocm`: Same as `full` but compiled with ROCm support. (platforms: `linux/amd64`, `linux/arm64`) - `ghcr.io/ggerganov/llama.cpp:light-rocm`: Same as `light` but compiled with ROCm support. (platforms: `linux/amd64`, `linux/arm64`) +- `ghcr.io/ggerganov/llama.cpp:server-rocm`: Same as `server` but compiled with ROCm support. (platforms: `linux/amd64`, `linux/arm64`) The GPU enabled images are not currently tested by CI beyond being built. They are not built with any variation from the ones in the Dockerfiles defined in [.devops/](.devops/) and the GitHub Action defined in [.github/workflows/docker.yml](.github/workflows/docker.yml). If you need different settings (for example, a different CUDA or ROCm library, you'll need to build the images locally for now). @@ -967,6 +970,12 @@ or with a light image: docker run -v /path/to/models:/models ghcr.io/ggerganov/llama.cpp:light -m /models/7B/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 512 ``` +or with a server image: + +```bash +docker run -v /path/to/models:/models -p 8000:8000 ghcr.io/ggerganov/llama.cpp:server -m /models/7B/ggml-model-q4_0.gguf --port 8000 --host 0.0.0.0 -n 512 +``` + ### Docker With CUDA Assuming one has the [nvidia-container-toolkit](https://github.com/NVIDIA/nvidia-container-toolkit) properly installed on Linux, or is using a GPU enabled cloud, `cuBLAS` should be accessible inside the container. @@ -976,6 +985,7 @@ Assuming one has the [nvidia-container-toolkit](https://github.com/NVIDIA/nvidia ```bash docker build -t local/llama.cpp:full-cuda -f .devops/full-cuda.Dockerfile . docker build -t local/llama.cpp:light-cuda -f .devops/main-cuda.Dockerfile . +docker build -t local/llama.cpp:server-cuda -f .devops/server-cuda.Dockerfile . ``` You may want to pass in some different `ARGS`, depending on the CUDA environment supported by your container host, as well as the GPU architecture. @@ -989,6 +999,7 @@ The resulting images, are essentially the same as the non-CUDA images: 1. `local/llama.cpp:full-cuda`: This image includes both the main executable file and the tools to convert LLaMA models into ggml and convert into 4-bit quantization. 2. `local/llama.cpp:light-cuda`: This image only includes the main executable file. +3. `local/llama.cpp:server-cuda`: This image only includes the server executable file. #### Usage @@ -997,6 +1008,7 @@ After building locally, Usage is similar to the non-CUDA examples, but you'll ne ```bash docker run --gpus all -v /path/to/models:/models local/llama.cpp:full-cuda --run -m /models/7B/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 512 --n-gpu-layers 1 docker run --gpus all -v /path/to/models:/models local/llama.cpp:light-cuda -m /models/7B/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 512 --n-gpu-layers 1 +docker run --gpus all -v /path/to/models:/models local/llama.cpp:server-cuda -m /models/7B/ggml-model-q4_0.gguf --port 8000 --host 0.0.0.0 -n 512 --n-gpu-layers 1 ``` ### Contributing diff --git a/examples/server/README.md b/examples/server/README.md index 1c92a2041..dce4ec47c 100644 --- a/examples/server/README.md +++ b/examples/server/README.md @@ -66,6 +66,14 @@ server.exe -m models\7B\ggml-model.gguf -c 2048 The above command will start a server that by default listens on `127.0.0.1:8080`. You can consume the endpoints with Postman or NodeJS with axios library. You can visit the web front end at the same url. +### Docker: +```bash +docker run -p 8080:8080 -v /path/to/models:/models ggerganov/llama.cpp:server -m models/7B/ggml-model.gguf -c 512 --host 0.0.0.0 --port 8080 + +# or, with CUDA: +docker run -p 8080:8080 -v /path/to/models:/models --gpus all ggerganov/llama.cpp:server-cuda -m models/7B/ggml-model.gguf -c 512 --host 0.0.0.0 --port 8080 --n-gpu-layers 99 +``` + ## Testing with CURL Using [curl](https://curl.se/). On Windows `curl.exe` should be available in the base OS. From f2e69d28c01303ca9dc79907f89ef120a6ac4a92 Mon Sep 17 00:00:00 2001 From: sharpHL <132747147+sharpHL@users.noreply.github.com> Date: Sun, 28 Jan 2024 16:00:30 +0800 Subject: [PATCH 64/66] llama : add support for Orion-14B (#5118) * add support for Orion-14B(https://huggingface.co/OrionStarAI/Orion-14B-Chat) * flake8 support * Update llama.cpp Co-authored-by: Georgi Gerganov * Update llama.cpp Co-authored-by: Georgi Gerganov * Update llama.cpp Co-authored-by: Georgi Gerganov * Update llama.cpp Co-authored-by: Georgi Gerganov * Update llama.cpp Co-authored-by: slaren * Update llama.cpp * Update llama.cpp --------- Co-authored-by: lixiaopu Co-authored-by: Georgi Gerganov Co-authored-by: slaren --- convert-hf-to-gguf.py | 81 +++++++++++++++++ gguf-py/gguf/constants.py | 24 ++++- llama.cpp | 187 ++++++++++++++++++++++++++++++++++++++ 3 files changed, 291 insertions(+), 1 deletion(-) diff --git a/convert-hf-to-gguf.py b/convert-hf-to-gguf.py index 7a0a8c3db..6ab7f486e 100755 --- a/convert-hf-to-gguf.py +++ b/convert-hf-to-gguf.py @@ -201,6 +201,8 @@ class Model: return PlamoModel if model_architecture == "CodeShellForCausalLM": return CodeShellModel + if model_architecture == "OrionForCausalLM": + return OrionModel return Model def _is_model_safetensors(self) -> bool: @@ -250,6 +252,8 @@ class Model: return gguf.MODEL_ARCH.PLAMO if arch == "CodeShellForCausalLM": return gguf.MODEL_ARCH.CODESHELL + if arch == "OrionForCausalLM": + return gguf.MODEL_ARCH.ORION raise NotImplementedError(f'Architecture "{arch}" not supported!') @@ -572,6 +576,83 @@ class MPTModel(Model): self.gguf_writer.add_tensor("output.weight", data) +class OrionModel(Model): + def set_vocab(self): + self._set_vocab_sentencepiece() + + def set_gguf_parameters(self): + block_count = self.hparams["num_hidden_layers"] + head_count = self.hparams["num_attention_heads"] + head_count_kv = self.hparams.get("num_key_value_heads", head_count) + hf_repo = self.hparams.get("_name_or_path", "") + + ctx_length = 0 + if "max_sequence_length" in self.hparams: + ctx_length = self.hparams["max_sequence_length"] + elif "max_position_embeddings" in self.hparams: + ctx_length = self.hparams["max_position_embeddings"] + elif "model_max_length" in self.hparams: + ctx_length = self.hparams["model_max_length"] + else: + print("gguf: can not find ctx length parameter.") + sys.exit() + + self.gguf_writer.add_file_type(self.ftype) + self.gguf_writer.add_name(self.dir_model.name) + self.gguf_writer.add_source_hf_repo(hf_repo) + self.gguf_writer.add_tensor_data_layout("Meta AI original pth") + self.gguf_writer.add_context_length(ctx_length) + self.gguf_writer.add_embedding_length(self.hparams["hidden_size"]) + self.gguf_writer.add_block_count(block_count) + self.gguf_writer.add_feed_forward_length(self.hparams["intermediate_size"]) + self.gguf_writer.add_head_count(head_count) + self.gguf_writer.add_head_count_kv(head_count_kv) + self.gguf_writer.add_layer_norm_eps(self.hparams["rms_norm_eps"]) + + def write_tensors(self): + # Collect tensors from generator object + model_kv = dict(self.get_tensors()) + block_count = self.hparams["num_hidden_layers"] + tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count) + + for name, data_torch in model_kv.items(): + # we don't need these + if name.endswith(".rotary_emb.inv_freq"): + continue + + old_dtype = data_torch.dtype + + # convert any unsupported data types to float32 + if data_torch.dtype not in (torch.float16, torch.float32): + data_torch = data_torch.to(torch.float32) + + data = data_torch.squeeze().numpy() + + # map tensor names + new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias")) + if new_name is None: + print(f"Can not map tensor {name!r}") + sys.exit() + + n_dims = len(data.shape) + data_dtype = data.dtype + + # if f32 desired, convert any float16 to float32 + if self.ftype == 0 and data_dtype == np.float16: + data = data.astype(np.float32) + + # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32 + if self.ftype == 1 and data_dtype == np.float16 and n_dims == 1: + data = data.astype(np.float32) + + # if f16 desired, convert any float32 2-dim weight tensors to float16 + if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2: + data = data.astype(np.float16) + + print(f"{name} -> {new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}") + self.gguf_writer.add_tensor(new_name, data) + + class BaichuanModel(Model): def set_vocab(self): self._set_vocab_sentencepiece() diff --git a/gguf-py/gguf/constants.py b/gguf-py/gguf/constants.py index 2d9c33c7d..f5c933a41 100644 --- a/gguf-py/gguf/constants.py +++ b/gguf-py/gguf/constants.py @@ -101,6 +101,7 @@ class MODEL_ARCH(IntEnum): PHI2 = auto() PLAMO = auto() CODESHELL = auto() + ORION = auto() class MODEL_TENSOR(IntEnum): @@ -151,6 +152,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = { MODEL_ARCH.PHI2: "phi2", MODEL_ARCH.PLAMO: "plamo", MODEL_ARCH.CODESHELL: "codeshell", + MODEL_ARCH.ORION: "orion", } TENSOR_NAMES: dict[MODEL_TENSOR, str] = { @@ -427,7 +429,23 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = { MODEL_TENSOR.FFN_NORM, MODEL_TENSOR.FFN_DOWN, MODEL_TENSOR.FFN_UP, - ] + ], + MODEL_ARCH.ORION: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ROPE_FREQS, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_Q, + MODEL_TENSOR.ATTN_K, + MODEL_TENSOR.ATTN_V, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.ATTN_ROT_EMBD, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_GATE, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + ], # TODO } @@ -452,6 +470,10 @@ MODEL_TENSOR_SKIP: dict[MODEL_ARCH, list[MODEL_TENSOR]] = { MODEL_TENSOR.ROPE_FREQS, MODEL_TENSOR.ATTN_ROT_EMBD, ], + MODEL_ARCH.ORION: [ + MODEL_TENSOR.ROPE_FREQS, + MODEL_TENSOR.ATTN_ROT_EMBD, + ], } # diff --git a/llama.cpp b/llama.cpp index b03b67e16..4cd0f16eb 100644 --- a/llama.cpp +++ b/llama.cpp @@ -196,6 +196,7 @@ enum llm_arch { LLM_ARCH_PHI2, LLM_ARCH_PLAMO, LLM_ARCH_CODESHELL, + LLM_ARCH_ORION, LLM_ARCH_UNKNOWN, }; @@ -217,6 +218,7 @@ static std::map LLM_ARCH_NAMES = { { LLM_ARCH_PHI2, "phi2" }, { LLM_ARCH_PLAMO, "plamo" }, { LLM_ARCH_CODESHELL, "codeshell" }, + { LLM_ARCH_ORION, "orion" }, }; enum llm_kv { @@ -641,6 +643,25 @@ static std::map> LLM_TENSOR_NAMES = { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, }, }, + { + LLM_ARCH_ORION, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ROPE_FREQS, "rope_freqs" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" }, + { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" }, + { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_ATTN_ROT_EMBD, "blk.%d.attn_rot_embd" }, + { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, + { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + }, + }, { LLM_ARCH_UNKNOWN, @@ -1332,6 +1353,7 @@ enum e_model { MODEL_7B, MODEL_8B, MODEL_13B, + MODEL_14B, MODEL_15B, MODEL_30B, MODEL_34B, @@ -2683,6 +2705,7 @@ static const char * llama_model_type_name(e_model type) { case MODEL_7B: return "7B"; case MODEL_8B: return "8B"; case MODEL_13B: return "13B"; + case MODEL_14B: return "14B"; case MODEL_15B: return "15B"; case MODEL_30B: return "30B"; case MODEL_34B: return "34B"; @@ -2950,7 +2973,15 @@ static void llm_load_hparams( default: model.type = e_model::MODEL_UNKNOWN; } } break; + case LLM_ARCH_ORION: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps); + switch (hparams.n_layer) { + case 40: model.type = e_model::MODEL_14B; break; + default: model.type = e_model::MODEL_UNKNOWN; + } + } break; default: (void)0; } @@ -3933,6 +3964,38 @@ static bool llm_load_tensors( layer.ffn_up_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_UP, "bias", i), {n_ff}); } } break; + case LLM_ARCH_ORION: + { + model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}); + { + model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}); + model.output_norm_b = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "bias"), {n_embd}); + model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}); + } + for (int i = 0; i < n_layer; ++i) { + ggml_context * ctx_layer = ctx_for_layer(i); + ggml_context * ctx_split = ctx_for_layer_split(i); + + auto & layer = model.layers[i]; + + layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}); + layer.attn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "bias", i), {n_embd}); + + layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd}); + layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_gqa}); + layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_gqa}); + layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}); + + layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}); + layer.ffn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "bias", i), {n_embd}); + + layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}); + layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}); + layer.ffn_up = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}); + } + } break; + + default: throw std::runtime_error("unknown architecture"); } @@ -4563,6 +4626,126 @@ struct llm_build_context { ctx0 = nullptr; } } + struct ggml_cgraph * build_orion() { + struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false); + + const int64_t n_embd_head = hparams.n_embd_head_v; + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + struct ggml_tensor * cur; + struct ggml_tensor * inpL; + + inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, lctx.inp_tokens, lctx.inp_embd, cb); + cb(inpL, "inp_embd", -1); + + // inp_pos - contains the positions + struct ggml_tensor * inp_pos = ggml_view_1d(ctx0, lctx.inp_pos, n_tokens, 0); + cb(inp_pos, "inp_pos", -1); + + // KQ_mask (mask for 1 head, it will be broadcasted to all heads) + struct ggml_tensor * KQ_mask = ggml_view_2d(ctx0, lctx.inp_KQ_mask, n_kv, n_tokens, n_kv*ggml_type_size(lctx.inp_KQ_mask->type), 0); + cb(KQ_mask, "KQ_mask", -1); + + // shift the entire K-cache if needed + if (do_rope_shift) { + llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, lctx.inp_K_shift, LLM_ROPE, n_ctx, freq_base, freq_scale, cb); + } + + for (int il = 0; il < n_layer; ++il) { + struct ggml_tensor * inpSA = inpL; + + // norm + cur = llm_build_norm(ctx0, inpL, hparams, + model.layers[il].attn_norm, model.layers[il].attn_norm_b, + LLM_NORM, cb, il); + cb(cur, "attn_norm", il); + + // self-attention + { + // compute Q and K and RoPE them + struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + // if (model.layers[il].bq) { + // Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + // cb(Qcur, "Qcur", il); + // } + + struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + // if (model.layers[il].bk) { + // Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + // cb(Kcur, "Kcur", il); + // } + + struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + // if (model.layers[il].bv) { + // Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + // cb(Vcur, "Vcur", il); + // } + + Qcur = ggml_rope_custom( + ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, + hparams.n_rot, 2, 0, n_orig_ctx, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Qcur, "Qcur", il); + + Kcur = ggml_rope_custom( + ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, + hparams.n_rot, 2, 0, n_orig_ctx, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Kcur, "Kcur", il); + + cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, + model.layers[il].wo, NULL, + Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il); + cb(cur, "kqv_out", il); + } + + struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + cur = llm_build_norm(ctx0, ffn_inp, hparams, + model.layers[il].ffn_norm, model.layers[il].ffn_norm_b, + LLM_NORM, cb, il); + cb(cur, "ffn_norm", il); + + cur = llm_build_ffn(ctx0, cur, + model.layers[il].ffn_up, NULL, + model.layers[il].ffn_gate, NULL, + model.layers[il].ffn_down, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, cb, il); + cb(cur, "ffn_out", il); + + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = llm_build_norm(ctx0, cur, hparams, + model.output_norm, model.output_norm_b, + LLM_NORM, cb, -1); + cb(cur, "result_norm", -1); + + // lm_head + cur = ggml_mul_mat(ctx0, model.output, cur); + cb(cur, "result_output", -1); + + ggml_build_forward_expand(gf, cur); + + return gf; + } + + struct ggml_cgraph * build_llama() { struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false); @@ -6520,6 +6703,10 @@ static struct ggml_cgraph * llama_build_graph( { result = llm.build_codeshell(); } break; + case LLM_ARCH_ORION: + { + result = llm.build_orion(); + } break; default: GGML_ASSERT(false); } From af4980bfedfd8df43b9e4cd1442895e85fee37bc Mon Sep 17 00:00:00 2001 From: Marcus Dunn <51931484+MarcusDunn@users.noreply.github.com> Date: Sun, 28 Jan 2024 00:30:44 -0800 Subject: [PATCH 65/66] readme : add link to rust bindings (#5148) * added link to another set of rust bindings with brief note on differences. * fixed link name --- README.md | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/README.md b/README.md index cd95f8144..44898d2f2 100644 --- a/README.md +++ b/README.md @@ -122,7 +122,8 @@ as the main playground for developing new features for the [ggml](https://github - Node.js: [withcatai/node-llama-cpp](https://github.com/withcatai/node-llama-cpp) - JS/TS (llama.cpp server client): [lgrammel/modelfusion](https://modelfusion.dev/integration/model-provider/llamacpp) - Ruby: [yoshoku/llama_cpp.rb](https://github.com/yoshoku/llama_cpp.rb) -- Rust: [mdrokz/rust-llama.cpp](https://github.com/mdrokz/rust-llama.cpp) +- Rust (nicer API): [mdrokz/rust-llama.cpp](https://github.com/mdrokz/rust-llama.cpp) +- Rust (more direct bindings): [utilityai/llama-cpp-rs](https://github.com/utilityai/llama-cpp-rs) - C#/.NET: [SciSharp/LLamaSharp](https://github.com/SciSharp/LLamaSharp) - Scala 3: [donderom/llm4s](https://github.com/donderom/llm4s) - Clojure: [phronmophobic/llama.clj](https://github.com/phronmophobic/llama.clj) From b2b2bf988c098851b4f3831f0cf38394bff75121 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Johannes=20G=C3=A4=C3=9Fler?= Date: Sun, 28 Jan 2024 09:35:14 +0100 Subject: [PATCH 66/66] Tests for min_p, sampling queue (#5147) --- llama.cpp | 5 ++ tests/test-sampling.cpp | 169 ++++++++++++++++++++++++++++++++++++---- 2 files changed, 159 insertions(+), 15 deletions(-) diff --git a/llama.cpp b/llama.cpp index 4cd0f16eb..391c956ec 100644 --- a/llama.cpp +++ b/llama.cpp @@ -8133,6 +8133,11 @@ void llama_sample_softmax(struct llama_context * ctx, llama_token_data_array * c } void llama_sample_top_k(struct llama_context * ctx, llama_token_data_array * candidates, int32_t k, size_t min_keep) { + // TODO: move bucket sort to separate function so that top_p/tail_free/typical/softmax first is equally fast + // if (k >= (int32_t)candidates->size) { + // return; + // } + const int64_t t_start_sample_us = ggml_time_us(); k = std::max(k, (int) min_keep); diff --git a/tests/test-sampling.cpp b/tests/test-sampling.cpp index 32e58941c..c3b3d6629 100644 --- a/tests/test-sampling.cpp +++ b/tests/test-sampling.cpp @@ -5,11 +5,10 @@ #undef NDEBUG #endif -#include -#include -#include -#include #include +#include +#include +#include static void dump(const llama_token_data_array * candidates) { for (size_t i = 0; i < candidates->size; i++) { @@ -20,11 +19,11 @@ static void dump(const llama_token_data_array * candidates) { #define DUMP(__candidates) do { printf("%s:%d (%s)\n", __FILE__, __LINE__, __func__); dump((__candidates)); printf("-\n"); } while(0) static void test_top_k(const std::vector & probs, const std::vector & expected_probs, int k) { - size_t n_vocab = probs.size(); + const size_t n_vocab = probs.size(); std::vector candidates; candidates.reserve(n_vocab); for (llama_token token_id = 0; token_id < (llama_token)n_vocab; token_id++) { - float logit = log(probs[token_id]); + const float logit = logf(probs[token_id]); candidates.emplace_back(llama_token_data{token_id, logit, 0.0f}); } @@ -41,11 +40,11 @@ static void test_top_k(const std::vector & probs, const std::vector & probs, const std::vector & expected_probs, float p) { - size_t n_vocab = probs.size(); + const size_t n_vocab = probs.size(); std::vector candidates; candidates.reserve(n_vocab); for (llama_token token_id = 0; token_id < (llama_token)n_vocab; token_id++) { - float logit = log(probs[token_id]); + const float logit = logf(probs[token_id]); candidates.emplace_back(llama_token_data{token_id, logit, 0.0f}); } @@ -62,11 +61,11 @@ static void test_top_p(const std::vector & probs, const std::vector & probs, const std::vector & expected_probs, float z) { - size_t n_vocab = probs.size(); + const size_t n_vocab = probs.size(); std::vector candidates; candidates.reserve(n_vocab); for (llama_token token_id = 0; token_id < (llama_token)n_vocab; token_id++) { - float logit = log(probs[token_id]); + const float logit = logf(probs[token_id]); candidates.emplace_back(llama_token_data{token_id, logit, 0.0f}); } @@ -81,12 +80,33 @@ static void test_tfs(const std::vector & probs, const std::vector } } -static void test_typical(const std::vector & probs, const std::vector & expected_probs, float p) { - size_t n_vocab = probs.size(); +static void test_min_p(const std::vector & probs, const std::vector & expected_probs, float p) { + const size_t n_vocab = probs.size(); std::vector candidates; candidates.reserve(n_vocab); for (llama_token token_id = 0; token_id < (llama_token)n_vocab; token_id++) { - float logit = log(probs[token_id]); + const float logit = logf(probs[token_id]); + candidates.emplace_back(llama_token_data{token_id, logit, 0.0f}); + } + + llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false }; + DUMP(&candidates_p); + llama_sample_min_p(nullptr, &candidates_p, p, 1); + DUMP(&candidates_p); + llama_sample_softmax(nullptr, &candidates_p); + + GGML_ASSERT(candidates_p.size == expected_probs.size()); + for (size_t i = 0; i < candidates_p.size; i++) { + GGML_ASSERT(fabs(candidates_p.data[i].p - expected_probs[i]) < 1e-3); + } +} + +static void test_typical(const std::vector & probs, const std::vector & expected_probs, float p) { + const size_t n_vocab = probs.size(); + std::vector candidates; + candidates.reserve(n_vocab); + for (llama_token token_id = 0; token_id < (llama_token)n_vocab; token_id++) { + const float logit = logf(probs[token_id]); candidates.emplace_back(llama_token_data{token_id, logit, 0.0f}); } @@ -107,11 +127,11 @@ static void test_repetition_penalties( ) { GGML_ASSERT(probs.size() == expected_probs.size()); - size_t n_vocab = probs.size(); + const size_t n_vocab = probs.size(); std::vector candidates; candidates.reserve(n_vocab); for (llama_token token_id = 0; token_id < (llama_token)n_vocab; token_id++) { - float logit = log(probs[token_id]); + const float logit = logf(probs[token_id]); candidates.emplace_back(llama_token_data{token_id, logit, 0.0f}); } @@ -128,6 +148,88 @@ static void test_repetition_penalties( } } +static void test_sampler_queue( + const size_t n_vocab, const std::string samplers_sequence, const int top_k, const float top_p, const float min_p +) { + std::vector candidates; + candidates.reserve(n_vocab); + for (llama_token token_id = 0; token_id < (llama_token)n_vocab; token_id++) { + const float logit = logf(token_id); + candidates.emplace_back(llama_token_data{token_id, logit, 0.0f}); + } + + llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false }; + + llama_token min_token_id = 0; + const llama_token max_token_id = n_vocab-1; + + for (auto s : samplers_sequence) { + switch (s){ + case 'k': llama_sample_top_k (nullptr, &candidates_p, top_k, 1); break; + case 'f': GGML_ASSERT(false && "tail_free test not implemented"); break; + case 'y': GGML_ASSERT(false && "typical test not implemented"); break; + case 'p': llama_sample_top_p (nullptr, &candidates_p, top_p, 1); break; + case 'm': llama_sample_min_p (nullptr, &candidates_p, min_p, 1); break; + case 't': GGML_ASSERT(false && "temperature test not implemented"); break; + default : GGML_ASSERT(false && "Unknown sampler"); break; + } + + llama_sample_softmax(nullptr, &candidates_p); // make sure tokens are sorted for tests + + const int size = candidates_p.size; + + if (s == 'k') { + const int expected_size = std::min(size, top_k); + min_token_id = std::max(min_token_id, (llama_token)(n_vocab - top_k)); + + GGML_ASSERT(size == expected_size); + GGML_ASSERT(candidates_p.data[0].id == max_token_id); + GGML_ASSERT(candidates_p.data[expected_size-1].id == min_token_id); + } else if (s == 'p') { + const int softmax_divisor = n_vocab * (n_vocab-1) / 2 - min_token_id * (min_token_id-1) / 2; + const int softmax_numerator_target = ceilf(top_p * softmax_divisor); + + min_token_id = n_vocab; + int expected_size = 0; + int cumsum = 0; + do { // do-while because always at least one token is sampled + min_token_id--; + expected_size++; + + cumsum += min_token_id; + } while (cumsum < softmax_numerator_target); + + // token 0 has p == 0, need special consideration for cumsum because top_p immediately returns + if (min_token_id == 1) { + min_token_id--; + expected_size += 1; + } + + GGML_ASSERT(size == expected_size); + GGML_ASSERT(candidates_p.data[0].id == max_token_id); + GGML_ASSERT(candidates_p.data[expected_size-1].id == min_token_id); + } else if (s == 'm') { + int expected_size = ceilf((1.0f-min_p) * n_vocab); + expected_size = std::max(expected_size, 1); + expected_size = std::min(expected_size, size); + + min_token_id = floorf(min_p * n_vocab); + min_token_id = std::max(min_token_id, 1); + min_token_id = std::max(min_token_id, (llama_token)(n_vocab - size)); + min_token_id = std::min(min_token_id, (llama_token)(n_vocab - 1)); + + GGML_ASSERT(size == expected_size); + GGML_ASSERT(candidates_p.data[0].id == max_token_id); + GGML_ASSERT(candidates_p.data[expected_size-1].id == min_token_id); + } else { + GGML_ASSERT(false); + } + } + + printf("Sampler queue %3s OK with n_vocab=%05ld top_k=%05d top_p=%f min_p=%f\n", + samplers_sequence.c_str(), n_vocab, top_k, top_p, min_p); +} + int main(void) { ggml_time_init(); @@ -139,6 +241,15 @@ int main(void) { test_top_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f, 0.3f, 0.2f}, 0.8f); test_top_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f, 0.3f, 0.2f, 0.1f}, 1); + test_min_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f/1.0f, 0.3f/1.0f, 0.2f/1.0f, 0.1f/1.0f}, 0.00f); + test_min_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f/1.0f, 0.3f/1.0f, 0.2f/1.0f, 0.1f/1.0f}, 0.24f); + test_min_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f/0.9f, 0.3f/0.9f, 0.2f/0.9f}, 0.26f); + test_min_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f/0.9f, 0.3f/0.9f, 0.2f/0.9f}, 0.49f); + test_min_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f/0.7f, 0.3f/0.7f}, 0.51f); + test_min_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f/0.7f, 0.3f/0.7f}, 0.74f); + test_min_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f/0.4f}, 0.76f); + test_min_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f/0.4f}, 1.00f); + test_tfs({0.1f, 0.15f, 0.2f, 0.25f, 0.3f}, {0.3f}, 0.25f); test_tfs({0.1f, 0.15f, 0.2f, 0.25f, 0.3f}, {0.3f, 0.25f}, 0.75f); test_tfs({0.1f, 0.15f, 0.2f, 0.25f, 0.3f}, {0.3f, 0.25f}, 0.99f); @@ -154,6 +265,34 @@ int main(void) { test_repetition_penalties({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 2}, {0.499966f, 0.499966f, 0.000023f, 0.000023f, 0.000023f}, 1.0f, 5.0f, 5.0f); test_repetition_penalties({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 2, 0, 0}, {0.499977f, 0.499977f, 0.000023f, 0.000023f, 0.000000f}, 1.0f, 5.0f, 5.0f); + test_sampler_queue(10000, "k", 10000, 1.0f, 1.0f); + test_sampler_queue(10000, "k", 1, 1.0f, 1.0f); + test_sampler_queue(10000, "p", 10000, 1.0f, 1.0f); + test_sampler_queue(10000, "p", 10000, 0.0f, 1.0f); + test_sampler_queue(10000, "m", 10000, 1.0f, 1.0f); + test_sampler_queue(10000, "m", 10000, 1.0f, 1e-12); + + test_sampler_queue(10000, "k", 100, 1.0000f, 1.0f); + test_sampler_queue(10000, "p", 10000, 0.0002f, 1.0f); + test_sampler_queue(10000, "p", 10000, 0.8000f, 1.0f); + test_sampler_queue(10000, "m", 10000, 1.0000f, 9997.9f/9999.0f); + test_sampler_queue(10000, "m", 10000, 1.0000f, 0.1f); + + test_sampler_queue(10000, "kp", 100, 0.8f, 0.1f); + test_sampler_queue(10000, "km", 100, 0.8f, 0.1f); + test_sampler_queue(10000, "pk", 100, 0.8f, 0.1f); + test_sampler_queue(10000, "pm", 100, 0.8f, 0.1f); + test_sampler_queue(10000, "mk", 100, 0.8f, 0.1f); + test_sampler_queue(10000, "mp", 100, 0.8f, 9997.9f/9999.0f); + test_sampler_queue(10000, "mp", 100, 0.8f, 0.1f); + + test_sampler_queue(10000, "kpm", 100, 0.8f, 0.1f); + test_sampler_queue(10000, "kmp", 100, 0.8f, 0.1f); + test_sampler_queue(10000, "pkm", 100, 0.8f, 0.1f); + test_sampler_queue(10000, "pmk", 100, 0.8f, 0.1f); + test_sampler_queue(10000, "mkp", 100, 0.8f, 0.1f); + test_sampler_queue(10000, "mpk", 100, 0.8f, 0.1f); + printf("OK\n"); return 0;