diff options
Diffstat (limited to 'llama.cpp')
| -rw-r--r-- | llama.cpp | 1347 |
1 files changed, 1003 insertions, 344 deletions
@@ -56,8 +56,21 @@ #pragma warning(disable: 4244 4267) // possible loss of data #endif +static void llama_log_internal(llama_log_level level, const char* format, ...); +static void llama_log_callback_default(llama_log_level level, const char * text, void * user_data); +#define LLAMA_LOG_INFO(...) llama_log_internal(LLAMA_LOG_LEVEL_INFO , __VA_ARGS__) +#define LLAMA_LOG_WARN(...) llama_log_internal(LLAMA_LOG_LEVEL_WARN , __VA_ARGS__) +#define LLAMA_LOG_ERROR(...) llama_log_internal(LLAMA_LOG_LEVEL_ERROR, __VA_ARGS__) + + +#if !defined(GGML_USE_CUBLAS) && !defined(GGML_USE_METAL) +#include "ggml-alloc.h" +#define LLAMA_USE_ALLOCATOR +#else #define LLAMA_USE_SCRATCH #define LLAMA_MAX_SCRATCH_BUFFERS 16 +#endif + // available llama models enum e_model { @@ -67,6 +80,7 @@ enum e_model { MODEL_13B, MODEL_30B, MODEL_65B, + MODEL_70B, }; static const size_t kB = 1024; @@ -98,17 +112,18 @@ static void ggml_graph_compute_helper(std::vector<uint8_t> & buf, ggml_cgraph * } // -// memory sizes +// memory sizes (calculated for n_batch == 512) // -static const std::map<e_model, size_t> & MEM_REQ_SCRATCH0() +static const std::map<e_model, size_t> & MEM_REQ_SCRATCH0(int n_ctx) { static std::map<e_model, size_t> k_sizes = { - { MODEL_3B, 256ull * MB }, - { MODEL_7B, 512ull * MB }, - { MODEL_13B, 512ull * MB }, - { MODEL_30B, 512ull * MB }, - { MODEL_65B, 1024ull * MB }, + { MODEL_3B, ((size_t) n_ctx / 16ull + 92ull) * MB }, + { MODEL_7B, ((size_t) n_ctx / 16ull + 100ull) * MB }, + { MODEL_13B, ((size_t) n_ctx / 12ull + 120ull) * MB }, + { MODEL_30B, ((size_t) n_ctx / 9ull + 160ull) * MB }, + { MODEL_65B, ((size_t) n_ctx / 6ull + 256ull) * MB }, // guess + { MODEL_70B, ((size_t) n_ctx / 7ull + 164ull) * MB }, }; return k_sizes; } @@ -116,44 +131,32 @@ static const std::map<e_model, size_t> & MEM_REQ_SCRATCH0() static const std::map<e_model, size_t> & MEM_REQ_SCRATCH1() { static std::map<e_model, size_t> k_sizes = { - { MODEL_3B, 256ull * MB }, - { MODEL_7B, 512ull * MB }, - { MODEL_13B, 512ull * MB }, - { MODEL_30B, 512ull * MB }, - { MODEL_65B, 1024ull * MB }, + { MODEL_3B, 128ull * MB }, + { MODEL_7B, 160ull * MB }, + { MODEL_13B, 192ull * MB }, + { MODEL_30B, 256ull * MB }, + { MODEL_65B, 384ull * MB }, // guess + { MODEL_70B, 304ull * MB }, }; return k_sizes; } -// 2*n_embd*n_ctx*n_layer*sizeof(float16) -static const std::map<e_model, size_t> & MEM_REQ_KV_SELF() -{ - static std::map<e_model, size_t> k_sizes = { - { MODEL_3B, 682ull * MB }, - { MODEL_7B, 1026ull * MB }, - { MODEL_13B, 1608ull * MB }, - { MODEL_30B, 3124ull * MB }, - { MODEL_65B, 5120ull * MB }, - }; - return k_sizes; -} - -// this is mostly needed for temporary mul_mat buffers to dequantize the data -// not actually needed if BLAS is disabled +// used to store the compute graph tensors + non-scratch data static const std::map<e_model, size_t> & MEM_REQ_EVAL() { static std::map<e_model, size_t> k_sizes = { - { MODEL_3B, 512ull * MB }, - { MODEL_7B, 768ull * MB }, - { MODEL_13B, 1024ull * MB }, - { MODEL_30B, 1280ull * MB }, - { MODEL_65B, 1536ull * MB }, + { MODEL_3B, 8ull * MB }, + { MODEL_7B, 10ull * MB }, + { MODEL_13B, 12ull * MB }, + { MODEL_30B, 16ull * MB }, + { MODEL_65B, 24ull * MB }, // guess + { MODEL_70B, 24ull * MB }, }; return k_sizes; } // amount of VRAM needed per batch size to hold temporary results -// the values for 3b and 65b are not derived from testing but instead chosen conservatively +// the values for 3b are not derived from testing but instead chosen conservatively static const std::map<e_model, size_t> & VRAM_REQ_SCRATCH_BASE() { static std::map<e_model, size_t> k_sizes = { @@ -161,13 +164,14 @@ static const std::map<e_model, size_t> & VRAM_REQ_SCRATCH_BASE() { MODEL_7B, 512ull * kB }, { MODEL_13B, 640ull * kB }, { MODEL_30B, 768ull * kB }, - { MODEL_65B, 1536ull * kB }, + { MODEL_65B, 1280ull * kB }, + { MODEL_70B, 1280ull * kB }, }; return k_sizes; } // amount of VRAM needed per batch size and context to hold temporary results -// the values for 3b and 65b are not derived from testing but instead chosen conservatively +// the values for 3b are not derived from testing but instead chosen conservatively static const std::map<e_model, size_t> & VRAM_REQ_SCRATCH_PER_CONTEXT() { static std::map<e_model, size_t> k_sizes = { @@ -175,24 +179,56 @@ static const std::map<e_model, size_t> & VRAM_REQ_SCRATCH_PER_CONTEXT() { MODEL_7B, 128ull }, { MODEL_13B, 160ull }, { MODEL_30B, 208ull }, - { MODEL_65B, 416ull }, + { MODEL_65B, 256ull }, + { MODEL_70B, 256ull }, }; return k_sizes; } // default hparams (LLaMA 7B) struct llama_hparams { - uint32_t n_vocab = 32000; - uint32_t n_ctx = 512; // this is provided as user input? - uint32_t n_embd = 4096; - uint32_t n_mult = 256; - uint32_t n_head = 32; - uint32_t n_layer = 32; - uint32_t n_rot = 64; + uint32_t n_vocab = 32000; + uint32_t n_ctx = 512; // this is provided as user input? + uint32_t n_embd = 4096; + uint32_t n_mult = 256; + uint32_t n_head = 32; + uint32_t n_head_kv = 32; + uint32_t n_layer = 32; + uint32_t n_rot = 64; + + // LLaMAv2 + // TODO: load from model data hparams + float f_ffn_mult = 1.0f; + float f_rms_norm_eps = LLAMA_DEFAULT_RMS_EPS; + + float rope_freq_base = 10000.0f; + float rope_freq_scale = 1.0f; + enum llama_ftype ftype = LLAMA_FTYPE_MOSTLY_F16; bool operator!=(const llama_hparams & other) const { - return static_cast<bool>(memcmp(this, &other, sizeof(llama_hparams))); + return static_cast<bool>(memcmp(this, &other, sizeof(llama_hparams))); // NOLINT + } + + uint32_t n_gqa() const { + return n_head/n_head_kv; + } + + uint32_t n_embd_head() const { + return n_embd/n_head; + } + + uint32_t n_embd_gqa() const { + return n_embd/n_gqa(); + } + + size_t kv_size() const { + size_t result = 2ull; + result *= (size_t) n_embd_gqa(); + result *= (size_t) n_ctx; + result *= (size_t) n_layer; + result *= sizeof(ggml_fp16_t); + return result; } }; @@ -303,14 +339,23 @@ struct llama_model { }; struct llama_context { - llama_context(const llama_model & model, const llama_vocab & vocab) : model(model), vocab(vocab), t_load_us(model.t_load_us), t_start_us(model.t_start_us) {} -#ifdef GGML_USE_METAL + llama_context(const llama_model & model) : model(model), t_load_us(model.t_load_us), t_start_us(model.t_start_us) {} ~llama_context() { + if (model_owner) { + delete &model; + } +#ifdef GGML_USE_METAL if (ctx_metal) { ggml_metal_free(ctx_metal); } - } #endif +#ifdef LLAMA_USE_ALLOCATOR + if (alloc) { + ggml_allocr_free(alloc); + } +#endif + } + std::mt19937 rng; bool has_evaluated_once = false; @@ -324,7 +369,6 @@ struct llama_context { int32_t n_p_eval = 0; // number of tokens in eval calls for the prompt (with batch size > 1) const llama_model & model; - const llama_vocab & vocab; bool model_owner = false; @@ -349,7 +393,17 @@ struct llama_context { // memory buffers used to evaluate the model // TODO: move in llama_state llama_ctx_buffer buf_compute; + +#ifdef LLAMA_USE_ALLOCATOR + llama_ctx_buffer buf_alloc; + ggml_allocr * alloc = NULL; +#endif + +#ifdef LLAMA_USE_SCRATCH llama_ctx_buffer buf_scratch[LLAMA_MAX_SCRATCH_BUFFERS]; + int buf_last = 0; + size_t buf_max_size[LLAMA_MAX_SCRATCH_BUFFERS] = { 0 }; +#endif #ifdef GGML_USE_METAL ggml_metal_context * ctx_metal = NULL; @@ -359,9 +413,6 @@ struct llama_context { ggml_mpi_context * ctx_mpi = NULL; #endif - int buf_last = 0; - size_t buf_max_size[LLAMA_MAX_SCRATCH_BUFFERS] = { 0 }; - void use_buf(struct ggml_context * ctx, int i) { #if defined(LLAMA_USE_SCRATCH) size_t last_size = 0; @@ -394,6 +445,14 @@ struct llama_context { } }; +struct llama_state { + // We save the log callback globally + llama_log_callback log_callback = llama_log_callback_default; + void * log_callback_user_data = nullptr; +}; +// global state +static llama_state g_state; + template <typename T> static T checked_mul(T a, T b) { T ret = a * b; @@ -460,7 +519,7 @@ struct llama_file_loader { llama_file_loader(const char * fname, llama_load_tensors_map & tensors_map) : file(fname, "rb") { - fprintf(stderr, "llama.cpp: loading model from %s\n", fname); + LLAMA_LOG_INFO("llama.cpp: loading model from %s\n", fname); read_magic(); read_hparams(); read_vocab(); @@ -495,12 +554,16 @@ struct llama_file_loader { } void read_hparams() { hparams.n_vocab = file.read_u32(); - hparams.n_embd = file.read_u32(); - hparams.n_mult = file.read_u32(); - hparams.n_head = file.read_u32(); + hparams.n_embd = file.read_u32(); + hparams.n_mult = file.read_u32(); + hparams.n_head = file.read_u32(); hparams.n_layer = file.read_u32(); - hparams.n_rot = file.read_u32(); - hparams.ftype = (enum llama_ftype) file.read_u32(); + hparams.n_rot = file.read_u32(); + hparams.ftype = (enum llama_ftype) file.read_u32(); + + // LLaMAv2 + // TODO: read from header + hparams.n_head_kv = hparams.n_head; } void read_vocab() { vocab.id_to_token.resize(hparams.n_vocab); @@ -551,7 +614,9 @@ struct llama_file_loader { } // skip to the next multiple of 32 bytes - file.seek(-static_cast<ptrdiff_t>(file.tell()) & 31, SEEK_CUR); + if (file_version >= LLAMA_FILE_VERSION_GGJT_V1) { + file.seek(-static_cast<ptrdiff_t>(file.tell()) & 31, SEEK_CUR); + } tensor.file_off = file.tell(); tensor.name = name; @@ -569,7 +634,7 @@ struct llama_file_saver { llama_file_loader * any_file_loader; llama_file_saver(const char * fname, llama_file_loader * any_file_loader, enum llama_ftype new_ftype) : file(fname, "wb"), any_file_loader(any_file_loader) { - fprintf(stderr, "llama.cpp: saving model to %s\n", fname); + LLAMA_LOG_INFO("llama.cpp: saving model to %s\n", fname); write_magic(); write_hparams(new_ftype); write_vocab(); @@ -590,7 +655,7 @@ struct llama_file_saver { } void write_vocab() { if (any_file_loader->file_version == LLAMA_FILE_VERSION_GGML) { - fprintf(stderr, "llama.cpp: WARNING: input is an old file that doesn't have scores; will add dummy scores\n"); + LLAMA_LOG_WARN("llama.cpp: WARNING: input is an old file that doesn't have scores; will add dummy scores\n"); } uint32_t n_vocab = any_file_loader->hparams.n_vocab; for (uint32_t i = 0; i < n_vocab; i++) { @@ -648,7 +713,7 @@ struct llama_model_loader { *ctx_size_p = *mmapped_size_p = 0; for (const llama_load_tensor & lt : tensors_map.tensors) { *ctx_size_p += sizeof(struct ggml_tensor) + GGML_OBJECT_SIZE; - *(use_mmap ? mmapped_size_p : ctx_size_p) += lt.size; + *(use_mmap ? mmapped_size_p : ctx_size_p) += lt.size + 16; } } @@ -697,12 +762,12 @@ struct llama_model_loader { void load_all_data(llama_progress_callback progress_callback, void * progress_callback_user_data, llama_mlock * lmlock) { size_t data_size = 0; - size_t prefetch_size = 0; + size_t prefetch_size = file_loader->file.size; size_t lock_size = 0; for (const llama_load_tensor & lt : tensors_map.tensors) { data_size += lt.size; - if (lt.ggml_tensor->backend == GGML_BACKEND_CPU) { - prefetch_size += lt.size; + if (lt.ggml_tensor->backend != GGML_BACKEND_CPU) { + prefetch_size -= lt.size; } } @@ -781,7 +846,7 @@ struct llama_model_loader { uint8_t byte = lt.data[i]; sum = byte + (sum << 6) + (sum << 16) - sum; // sdbm hash } - fprintf(stderr, "%s checksum: %#08x (%s, size %zu)\n", lt.name.c_str(), sum, + LLAMA_LOG_INFO("%s checksum: %#08x (%s, size %zu)\n", lt.name.c_str(), sum, llama_format_tensor_shape(lt.ne).c_str(), lt.size); } @@ -797,7 +862,7 @@ static bool kv_cache_init( ggml_type wtype, int n_ctx, int n_gpu_layers) { - const int n_embd = hparams.n_embd; + const int n_embd = hparams.n_embd_gqa(); const int n_layer = hparams.n_layer; const int64_t n_mem = n_layer*n_ctx; @@ -814,7 +879,7 @@ static bool kv_cache_init( cache.ctx = ggml_init(params); if (!cache.ctx) { - fprintf(stderr, "%s: failed to allocate memory for kv cache\n", __func__); + LLAMA_LOG_ERROR("%s: failed to allocate memory for kv cache\n", __func__); return false; } @@ -841,12 +906,17 @@ struct llama_context_params llama_context_default_params() { /*.seed =*/ LLAMA_DEFAULT_SEED, /*.n_ctx =*/ 512, /*.n_batch =*/ 512, + /*.n_gqa =*/ 1, + /*.rms_norm_eps =*/ LLAMA_DEFAULT_RMS_EPS, /*.gpu_layers =*/ 0, /*.main_gpu =*/ 0, - /*.tensor_split =*/ {0}, + /*.tensor_split =*/ nullptr, + /*.rope_freq_base =*/ 10000.0f, + /*.rope_freq_scale =*/ 1.0f, /*.progress_callback =*/ nullptr, /*.progress_callback_user_data =*/ nullptr, /*.low_vram =*/ false, + /*.mul_mat_q =*/ false, /*.f16_kv =*/ true, /*.logits_all =*/ false, /*.vocab_only =*/ false, @@ -869,6 +939,10 @@ struct llama_model_quantize_params llama_model_quantize_default_params() { return result; } +int llama_max_devices() { + return LLAMA_MAX_DEVICES; +} + bool llama_mmap_supported() { return llama_mmap::SUPPORTED; } @@ -954,6 +1028,7 @@ static const char *llama_model_type_name(e_model type) { case MODEL_13B: return "13B"; case MODEL_30B: return "30B"; case MODEL_65B: return "65B"; + case MODEL_70B: return "70B"; default: LLAMA_ASSERT(false); } } @@ -964,9 +1039,14 @@ static void llama_model_load_internal( llama_vocab & vocab, int n_ctx, int n_batch, + int n_gqa, + float rms_norm_eps, int n_gpu_layers, int main_gpu, const float * tensor_split, + const bool mul_mat_q, + float rope_freq_base, + float rope_freq_scale, bool low_vram, ggml_type memory_type, bool use_mmap, @@ -983,8 +1063,12 @@ static void llama_model_load_internal( model.hparams = ml->file_loader->hparams; model.n_gpu_layers = n_gpu_layers; llama_file_version file_version = ml->file_loader->file_version; + auto & hparams = model.hparams; + // TODO: read from file + hparams.f_rms_norm_eps = rms_norm_eps; + { switch (hparams.n_layer) { case 26: model.type = e_model::MODEL_3B; break; @@ -1001,22 +1085,44 @@ static void llama_model_load_internal( } hparams.n_ctx = n_ctx; + + // LLaMAv2 + // TODO: temporary until GGUF + LLAMA_ASSERT(hparams.n_head % n_gqa == 0); + hparams.n_head_kv = hparams.n_head / n_gqa; + if (model.type == e_model::MODEL_65B && n_gqa == 8) { + LLAMA_LOG_WARN("%s: warning: assuming 70B model based on GQA == %d\n", __func__, n_gqa); + model.type = e_model::MODEL_70B; + hparams.f_ffn_mult = 1.3f; // from the params.json of the 70B model + } + + hparams.rope_freq_base = rope_freq_base; + hparams.rope_freq_scale = rope_freq_scale; } - const uint32_t n_ff = ((2*(4*hparams.n_embd)/3 + hparams.n_mult - 1)/hparams.n_mult)*hparams.n_mult; + // ref: https://github.com/facebookresearch/llama/blob/6c7fe276574e78057f917549435a2554000a876d/llama/model.py#L194-L199 + const uint32_t n_ff_raw = 2*(4*hparams.n_embd)/3; + const uint32_t n_ff_mult = hparams.f_ffn_mult*n_ff_raw; + const uint32_t n_ff = ((n_ff_mult + hparams.n_mult - 1)/hparams.n_mult)*hparams.n_mult; + //const uint32_t n_ff = 28672; { - fprintf(stderr, "%s: format = %s\n", __func__, llama_file_version_name(file_version)); - fprintf(stderr, "%s: n_vocab = %u\n", __func__, hparams.n_vocab); - fprintf(stderr, "%s: n_ctx = %u\n", __func__, hparams.n_ctx); - fprintf(stderr, "%s: n_embd = %u\n", __func__, hparams.n_embd); - fprintf(stderr, "%s: n_mult = %u\n", __func__, hparams.n_mult); - fprintf(stderr, "%s: n_head = %u\n", __func__, hparams.n_head); - fprintf(stderr, "%s: n_layer = %u\n", __func__, hparams.n_layer); - fprintf(stderr, "%s: n_rot = %u\n", __func__, hparams.n_rot); - fprintf(stderr, "%s: ftype = %u (%s)\n", __func__, hparams.ftype, llama_ftype_name(hparams.ftype)); - fprintf(stderr, "%s: n_ff = %u\n", __func__, n_ff); - fprintf(stderr, "%s: model size = %s\n", __func__, llama_model_type_name(model.type)); + LLAMA_LOG_INFO("%s: format = %s\n", __func__, llama_file_version_name(file_version)); + LLAMA_LOG_INFO("%s: n_vocab = %u\n", __func__, hparams.n_vocab); + LLAMA_LOG_INFO("%s: n_ctx = %u\n", __func__, hparams.n_ctx); + LLAMA_LOG_INFO("%s: n_embd = %u\n", __func__, hparams.n_embd); + LLAMA_LOG_INFO("%s: n_mult = %u\n", __func__, hparams.n_mult); + LLAMA_LOG_INFO("%s: n_head = %u\n", __func__, hparams.n_head); + LLAMA_LOG_INFO("%s: n_head_kv = %u\n", __func__, hparams.n_head_kv); + LLAMA_LOG_INFO("%s: n_layer = %u\n", __func__, hparams.n_layer); + LLAMA_LOG_INFO("%s: n_rot = %u\n", __func__, hparams.n_rot); // a.k.a. n_embd_head, n_head_dim + LLAMA_LOG_INFO("%s: n_gqa = %u\n", __func__, hparams.n_gqa()); + LLAMA_LOG_INFO("%s: rnorm_eps = %.1e\n", __func__, hparams.f_rms_norm_eps); + LLAMA_LOG_INFO("%s: n_ff = %u\n", __func__, n_ff); + LLAMA_LOG_INFO("%s: freq_base = %.1f\n", __func__, hparams.rope_freq_base); + LLAMA_LOG_INFO("%s: freq_scale = %g\n", __func__, hparams.rope_freq_scale); + LLAMA_LOG_INFO("%s: ftype = %u (%s)\n", __func__, hparams.ftype, llama_ftype_name(hparams.ftype)); + LLAMA_LOG_INFO("%s: model size = %s\n", __func__, llama_model_type_name(model.type)); } if (file_version < LLAMA_FILE_VERSION_GGJT_V2) { @@ -1044,13 +1150,13 @@ static void llama_model_load_internal( size_t ctx_size; size_t mmapped_size; ml->calc_sizes(&ctx_size, &mmapped_size); - fprintf(stderr, "%s: ggml ctx size = %7.2f MB\n", __func__, ctx_size/1024.0/1024.0); + LLAMA_LOG_INFO("%s: ggml ctx size = %7.2f MB\n", __func__, ctx_size/1024.0/1024.0); // create the ggml context { model.buf.resize(ctx_size); if (use_mlock) { - model.mlock_buf.init(model.buf.addr); + model.mlock_buf.init (model.buf.addr); model.mlock_buf.grow_to(model.buf.size); } @@ -1067,13 +1173,15 @@ static void llama_model_load_internal( } (void) main_gpu; + (void) mul_mat_q; #if defined(GGML_USE_CUBLAS) - fprintf(stderr, "%s: using CUDA for GPU acceleration\n", __func__); + LLAMA_LOG_INFO("%s: using CUDA for GPU acceleration\n", __func__); ggml_cuda_set_main_device(main_gpu); + ggml_cuda_set_mul_mat_q(mul_mat_q); #define LLAMA_BACKEND_OFFLOAD GGML_BACKEND_GPU #define LLAMA_BACKEND_OFFLOAD_SPLIT GGML_BACKEND_GPU_SPLIT #elif defined(GGML_USE_CLBLAST) - fprintf(stderr, "%s: using OpenCL for GPU acceleration\n", __func__); + LLAMA_LOG_INFO("%s: using OpenCL for GPU acceleration\n", __func__); #define LLAMA_BACKEND_OFFLOAD GGML_BACKEND_GPU #define LLAMA_BACKEND_OFFLOAD_SPLIT GGML_BACKEND_GPU #else @@ -1085,9 +1193,10 @@ static void llama_model_load_internal( size_t vram_weights = 0; size_t vram_scratch = 0; { - const uint32_t n_embd = hparams.n_embd; - const uint32_t n_layer = hparams.n_layer; - const uint32_t n_vocab = hparams.n_vocab; + const uint32_t n_embd = hparams.n_embd; + const uint32_t n_embd_gqa = hparams.n_embd_gqa(); + const uint32_t n_layer = hparams.n_layer; + const uint32_t n_vocab = hparams.n_vocab; ml->ggml_ctx = ctx; @@ -1135,16 +1244,16 @@ static void llama_model_load_internal( layer.attention_norm = ml->get_tensor(layers_i + ".attention_norm.weight", {n_embd}, backend); - layer.wq = ml->get_tensor(layers_i + ".attention.wq.weight", {n_embd, n_embd}, backend_split); - layer.wk = ml->get_tensor(layers_i + ".attention.wk.weight", {n_embd, n_embd}, backend_split); - layer.wv = ml->get_tensor(layers_i + ".attention.wv.weight", {n_embd, n_embd}, backend_split); - layer.wo = ml->get_tensor(layers_i + ".attention.wo.weight", {n_embd, n_embd}, backend_split); + layer.wq = ml->get_tensor(layers_i + ".attention.wq.weight", {n_embd, n_embd}, backend_split); + layer.wk = ml->get_tensor(layers_i + ".attention.wk.weight", {n_embd, n_embd_gqa}, backend_split); + layer.wv = ml->get_tensor(layers_i + ".attention.wv.weight", {n_embd, n_embd_gqa}, backend_split); + layer.wo = ml->get_tensor(layers_i + ".attention.wo.weight", {n_embd, n_embd}, backend_split); layer.ffn_norm = ml->get_tensor(layers_i + ".ffn_norm.weight", {n_embd}, backend); - layer.w1 = ml->get_tensor(layers_i + ".feed_forward.w1.weight", {n_embd, n_ff}, backend_split); - layer.w2 = ml->get_tensor(layers_i + ".feed_forward.w2.weight", { n_ff, n_embd}, backend_split); - layer.w3 = ml->get_tensor(layers_i + ".feed_forward.w3.weight", {n_embd, n_ff}, backend_split); + layer.w1 = ml->get_tensor(layers_i + ".feed_forward.w1.weight", {n_embd, n_ff}, backend_split); + layer.w2 = ml->get_tensor(layers_i + ".feed_forward.w2.weight", { n_ff, n_embd}, backend_split); + layer.w3 = ml->get_tensor(layers_i + ".feed_forward.w3.weight", {n_embd, n_ff}, backend_split); if (backend == GGML_BACKEND_GPU) { vram_weights += @@ -1162,25 +1271,29 @@ static void llama_model_load_internal( const size_t scale = memory_type == GGML_TYPE_F32 ? 2 : 1; // this is the total memory required to run the inference - const size_t mem_required = + size_t mem_required = ctx_size + - mmapped_size - vram_weights + // weights in VRAM not in memory - MEM_REQ_SCRATCH0().at(model.type) + + mmapped_size - vram_weights; // weights in VRAM not in memory + +#ifndef LLAMA_USE_ALLOCATOR + mem_required += + MEM_REQ_SCRATCH0(hparams.n_ctx).at(model.type) + MEM_REQ_SCRATCH1().at(model.type) + - MEM_REQ_EVAL().at (model.type); + MEM_REQ_EVAL().at(model.type); +#endif // this is the memory required by one llama_state const size_t mem_required_state = - scale*MEM_REQ_KV_SELF().at(model.type); + scale*hparams.kv_size(); - fprintf(stderr, "%s: mem required = %7.2f MB (+ %7.2f MB per state)\n", __func__, + LLAMA_LOG_INFO("%s: mem required = %7.2f MB (+ %7.2f MB per state)\n", __func__, mem_required / 1024.0 / 1024.0, mem_required_state / 1024.0 / 1024.0); (void) vram_scratch; (void) n_batch; #ifdef GGML_USE_CUBLAS if (low_vram) { - fprintf(stderr, "%s: not allocating a VRAM scratch buffer due to low VRAM option\n", __func__); + LLAMA_LOG_INFO("%s: not allocating a VRAM scratch buffer due to low VRAM option\n", __func__); ggml_cuda_set_scratch_size(0); // disable scratch } else { const size_t vram_scratch_base = VRAM_REQ_SCRATCH_BASE().at(model.type); @@ -1188,7 +1301,7 @@ static void llama_model_load_internal( vram_scratch = n_batch * (vram_scratch_base + n_ctx * vram_scratch_per_context); ggml_cuda_set_scratch_size(vram_scratch); if (n_gpu_layers > 0) { - fprintf(stderr, "%s: allocating batch_size x (%zd kB + n_ctx x %zd B) = %zd MB VRAM for the scratch buffer\n", + LLAMA_LOG_INFO("%s: allocating batch_size x (%zd kB + n_ctx x %zd B) = %zd MB VRAM for the scratch buffer\n", __func__, vram_scratch_base / kB, vram_scratch_per_context, (vram_scratch + MB - 1) / MB); // round up } @@ -1198,9 +1311,9 @@ static void llama_model_load_internal( #if defined(GGML_USE_CUBLAS) || defined(GGML_USE_CLBLAST) const int n_gpu = std::min(n_gpu_layers, int(hparams.n_layer)); - fprintf(stderr, "%s: offloading %d repeating layers to GPU\n", __func__, n_gpu); + LLAMA_LOG_INFO("%s: offloading %d repeating layers to GPU\n", __func__, n_gpu); if (n_gpu_layers > (int) hparams.n_layer) { - fprintf(stderr, "%s: offloading non-repeating layers to GPU\n", __func__); + LLAMA_LOG_INFO("%s: offloading non-repeating layers to GPU\n", __func__); } size_t vram_kv_cache = 0; @@ -1209,18 +1322,18 @@ static void llama_model_load_internal( const int max_offloadable_layers = low_vram ? hparams.n_layer + 1 : hparams.n_layer + 3; if (n_gpu_layers > (int) hparams.n_layer + 1) { if (low_vram) { - fprintf(stderr, "%s: cannot offload v cache to GPU due to low VRAM option\n", __func__); + LLAMA_LOG_INFO("%s: cannot offload v cache to GPU due to low VRAM option\n", __func__); } else { - fprintf(stderr, "%s: offloading v cache to GPU\n", __func__); - vram_kv_cache += MEM_REQ_KV_SELF().at(model.type) / 2; + LLAMA_LOG_INFO("%s: offloading v cache to GPU\n", __func__); + vram_kv_cache += hparams.kv_size() / 2; } } if (n_gpu_layers > (int) hparams.n_layer + 2) { if (low_vram) { - fprintf(stderr, "%s: cannot offload k cache to GPU due to low VRAM option\n", __func__); + LLAMA_LOG_WARN("%s: cannot offload k cache to GPU due to low VRAM option\n", __func__); } else { - fprintf(stderr, "%s: offloading k cache to GPU\n", __func__); - vram_kv_cache += MEM_REQ_KV_SELF().at(model.type) / 2; + LLAMA_LOG_INFO("%s: offloading k cache to GPU\n", __func__); + vram_kv_cache += hparams.kv_size() / 2; } } #elif defined(GGML_USE_CLBLAST) @@ -1228,9 +1341,9 @@ static void llama_model_load_internal( const int max_offloadable_layers = hparams.n_layer + 1; #endif // GGML_USE_CUBLAS - fprintf(stderr, "%s: offloaded %d/%d layers to GPU\n", + LLAMA_LOG_INFO("%s: offloaded %d/%d layers to GPU\n", __func__, std::min(n_gpu_layers, max_offloadable_layers), max_backend_supported_layers); - fprintf(stderr, "%s: total VRAM used: %zu MB\n", + LLAMA_LOG_INFO("%s: total VRAM used: %zu MB\n", __func__, (vram_weights + vram_scratch + vram_kv_cache + MB - 1) / MB); // round up #else (void) n_gpu_layers; @@ -1268,9 +1381,14 @@ static bool llama_model_load( llama_vocab & vocab, int n_ctx, int n_batch, + int n_gqa, + float rms_norm_eps, int n_gpu_layers, int main_gpu, - float * tensor_split, + const float * tensor_split, + const bool mul_mat_q, + float rope_freq_base, + float rope_freq_scale, bool low_vram, ggml_type memory_type, bool use_mmap, @@ -1279,41 +1397,25 @@ static bool llama_model_load( llama_progress_callback progress_callback, void *progress_callback_user_data) { try { - llama_model_load_internal(fname, model, vocab, n_ctx, n_batch, n_gpu_layers, main_gpu, tensor_split, low_vram, memory_type, + llama_model_load_internal(fname, model, vocab, n_ctx, n_batch, n_gqa, rms_norm_eps, n_gpu_layers, + main_gpu, tensor_split, mul_mat_q, rope_freq_base, rope_freq_scale, low_vram, memory_type, use_mmap, use_mlock, vocab_only, progress_callback, progress_callback_user_data); return true; } catch (const std::exception & err) { - fprintf(stderr, "error loading model: %s\n", err.what()); + LLAMA_LOG_ERROR("error loading model: %s\n", err.what()); return false; } } -// evaluate the transformer -// -// - lctx: llama context -// - tokens: new batch of tokens to process -// - embd embeddings input -// - n_tokens number of tokens -// - n_past: the context size so far -// - n_threads: number of threads to use -// -static bool llama_eval_internal( +static struct ggml_cgraph * llama_build_graph( llama_context & lctx, const llama_token * tokens, const float * embd, int n_tokens, - int n_past, - int n_threads, - const char * cgraph_fname) { + int n_past) { LLAMA_ASSERT((!tokens && embd) || (tokens && !embd)); -#ifdef GGML_USE_MPI - ggml_mpi_eval_init(lctx.ctx_mpi, &n_tokens, &n_past, &n_threads); -#endif - - const int64_t t_start_us = ggml_time_us(); - const int N = n_tokens; const auto & model = lctx.model; @@ -1323,37 +1425,54 @@ static bool llama_eval_internal( LLAMA_ASSERT(!!kv_self.ctx); - const int n_embd = hparams.n_embd; - const int n_layer = hparams.n_layer; - const int n_ctx = hparams.n_ctx; - const int n_head = hparams.n_head; - const int n_vocab = hparams.n_vocab; - const int n_rot = hparams.n_embd/hparams.n_head; + const int64_t n_embd = hparams.n_embd; + const int64_t n_layer = hparams.n_layer; + const int64_t n_ctx = hparams.n_ctx; + const int64_t n_head = hparams.n_head; + const int64_t n_head_kv = hparams.n_head_kv; + const int64_t n_embd_head = hparams.n_embd_head(); + const int64_t n_embd_gqa = hparams.n_embd_gqa(); + + LLAMA_ASSERT(n_embd_head == hparams.n_rot); + + const float freq_base = hparams.rope_freq_base; + const float freq_scale = hparams.rope_freq_scale; + const float rms_norm_eps = hparams.f_rms_norm_eps; + const int n_gpu_layers = model.n_gpu_layers; auto & mem_per_token = lctx.mem_per_token; auto & buf_compute = lctx.buf_compute; + struct ggml_init_params params = { /*.mem_size =*/ buf_compute.size, /*.mem_buffer =*/ buf_compute.addr, /*.no_alloc =*/ false, }; - struct ggml_context * ctx0 = ggml_init(params); +#ifdef LLAMA_USE_ALLOCATOR + params.no_alloc = true; +#endif - ggml_cgraph gf = {}; + struct ggml_context * ctx0 = ggml_init(params); - // for big prompts, if BLAS is enabled, it is better to use only one thread - // otherwise, the threads are spin-lock waiting for the BLAS calls and are degrading the performance - n_threads = N >= 32 && ggml_cpu_has_blas() && !ggml_cpu_has_gpublas() ? 1 : n_threads; + ggml_cgraph * gf = ggml_new_graph(ctx0); struct ggml_tensor * cur; struct ggml_tensor * inpL; if (tokens) { struct ggml_tensor * inp_tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N); + +#ifdef LLAMA_USE_ALLOCATOR + ggml_allocr_alloc(lctx.alloc, inp_tokens); + if (!ggml_allocr_is_measure(lctx.alloc)) { + memcpy(inp_tokens->data, tokens, N*ggml_element_size(inp_tokens)); + } +#else memcpy(inp_tokens->data, tokens, N*ggml_element_size(inp_tokens)); +#endif ggml_set_name(inp_tokens, "inp_tokens"); inpL = ggml_get_rows(ctx0, model.tok_embeddings, inp_tokens); @@ -1363,7 +1482,15 @@ static bool llama_eval_internal( #endif inpL = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, N); + +#ifdef LLAMA_USE_ALLOCATOR + ggml_allocr_alloc(lctx.alloc, inpL); + if (!ggml_allocr_is_measure(lctx.alloc)) { + memcpy(inpL->data, embd, N * n_embd * ggml_element_size(inpL)); + } +#else memcpy(inpL->data, embd, N * n_embd * ggml_element_size(inpL)); +#endif } const int i_gpu_start = n_layer - n_gpu_layers; @@ -1390,6 +1517,17 @@ static bool llama_eval_internal( } #endif // GGML_USE_CUBLAS + struct ggml_tensor * KQ_scale = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, 1); +#ifdef LLAMA_USE_ALLOCATOR + ggml_allocr_alloc(lctx.alloc, KQ_scale); + if (!ggml_allocr_is_measure(lctx.alloc)) { + ggml_set_f32(KQ_scale, 1.0f/sqrtf(float(n_embd)/n_head)); + } +#else + ggml_set_f32(KQ_scale, 1.0f/sqrtf(float(n_embd)/n_head)); +#endif + ggml_set_name(KQ_scale, "1/sqrt(n_embd_head)"); + for (int il = 0; il < n_layer; ++il) { ggml_format_name(inpL, "layer_inp_%d", il); @@ -1407,7 +1545,7 @@ static bool llama_eval_internal( // norm { - cur = ggml_rms_norm(ctx0, inpL); + cur = ggml_rms_norm(ctx0, inpL, rms_norm_eps); offload_func(cur); ggml_set_name(cur, "rms_norm_0"); @@ -1428,11 +1566,11 @@ static bool llama_eval_internal( offload_func_kq(tmpq); ggml_set_name(tmpq, "tmpq"); - struct ggml_tensor * Kcur = ggml_rope_inplace(ctx0, ggml_reshape_3d(ctx0, tmpk, n_embd/n_head, n_head, N), n_past, n_rot, 0, 0); + struct ggml_tensor * Kcur = ggml_rope_custom_inplace(ctx0, ggml_reshape_3d(ctx0, tmpk, n_embd_head, n_head_kv, N), n_past, n_embd_head, 0, 0, freq_base, freq_scale); offload_func_kq(Kcur); ggml_set_name(Kcur, "Kcur"); - struct ggml_tensor * Qcur = ggml_rope_inplace(ctx0, ggml_reshape_3d(ctx0, tmpq, n_embd/n_head, n_head, N), n_past, n_rot, 0, 0); + struct ggml_tensor * Qcur = ggml_rope_custom_inplace(ctx0, ggml_reshape_3d(ctx0, tmpq, n_embd_head, n_head, N), n_past, n_embd_head, 0, 0, freq_base, freq_scale); offload_func_kq(Qcur); ggml_set_name(Qcur, "Qcur"); @@ -1444,23 +1582,23 @@ static bool llama_eval_internal( offload_func_v(tmpv); ggml_set_name(tmpv, "tmpv"); - struct ggml_tensor * Vcur = ggml_transpose(ctx0, ggml_reshape_2d(ctx0, tmpv, n_embd, N)); + struct ggml_tensor * Vcur = ggml_transpose(ctx0, ggml_reshape_2d(ctx0, tmpv, n_embd_gqa, N)); offload_func_v(Vcur); ggml_set_name(Vcur, "Vcur"); - struct ggml_tensor * k = ggml_view_1d(ctx0, kv_self.k, N*n_embd, (ggml_element_size(kv_self.k)*n_embd)*(il*n_ctx + n_past)); + struct ggml_tensor * k = ggml_view_1d(ctx0, kv_self.k, N*n_embd_gqa, (ggml_element_size(kv_self.k)*n_embd_gqa)*(il*n_ctx + n_past)); offload_func_kq(k); ggml_set_name(k, "k"); - struct ggml_tensor * v = ggml_view_2d(ctx0, kv_self.v, N, n_embd, + struct ggml_tensor * v = ggml_view_2d(ctx0, kv_self.v, N, n_embd_gqa, ( n_ctx)*ggml_element_size(kv_self.v), - (il*n_ctx)*ggml_element_size(kv_self.v)*n_embd + n_past*ggml_element_size(kv_self.v)); + (il*n_ctx)*ggml_element_size(kv_self.v)*n_embd_gqa + n_past*ggml_element_size(kv_self.v)); offload_func_v(v); ggml_set_name(v, "v"); // important: storing RoPE-ed version of K in the KV cache! - ggml_build_forward_expand(&gf, ggml_cpy(ctx0, Kcur, k)); - ggml_build_forward_expand(&gf, ggml_cpy(ctx0, Vcur, v)); + ggml_build_forward_expand(gf, ggml_cpy(ctx0, Kcur, k)); + ggml_build_forward_expand(gf, ggml_cpy(ctx0, Vcur, v)); } struct ggml_tensor * Q = @@ -1473,8 +1611,8 @@ static bool llama_eval_internal( struct ggml_tensor * K = ggml_permute(ctx0, ggml_reshape_3d(ctx0, - ggml_view_1d(ctx0, kv_self.k, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(kv_self.k)*n_embd), - n_embd/n_head, n_head, n_past + N), + ggml_view_1d(ctx0, kv_self.k, (n_past + N)*n_embd_gqa, il*n_ctx*ggml_element_size(kv_self.k)*n_embd_gqa), + n_embd_head, n_head_kv, n_past + N), 0, 2, 1, 3); offload_func_kq(K); ggml_set_name(K, "K"); @@ -1484,10 +1622,7 @@ static bool llama_eval_internal( offload_func_kq(KQ); ggml_set_name(KQ, "KQ"); - // KQ_scaled = KQ / sqrt(n_embd/n_head) - struct ggml_tensor * KQ_scale = ggml_new_f32(ctx0, 1.0f/sqrtf(float(n_embd)/n_head)); - ggml_set_name(KQ_scale, "1/sqrt(n_embd/n_head)"); - + // KQ_scaled = KQ / sqrt(n_embd_head) // KQ_scaled shape [n_past + N, N, n_head, 1] struct ggml_tensor * KQ_scaled = ggml_scale_inplace(ctx0, KQ, KQ_scale); offload_func_kq(KQ_scaled); @@ -1506,10 +1641,10 @@ static bool llama_eval_internal( // split cached V into n_head heads struct ggml_tensor * V = ggml_view_3d(ctx0, kv_self.v, - n_past + N, n_embd/n_head, n_head, + n_past + N, n_embd_head, n_head_kv, n_ctx*ggml_element_size(kv_self.v), - n_ctx*ggml_element_size(kv_self.v)*n_embd/n_head, - il*n_ctx*ggml_element_size(kv_self.v)*n_embd); + n_ctx*ggml_element_size(kv_self.v)*n_embd_head, + n_ctx*ggml_element_size(kv_self.v)*n_embd_gqa*il); offload_func_v(V); ggml_set_name(V, "V"); @@ -1521,7 +1656,7 @@ static bool llama_eval_internal( // make V contiguous in memory to speed up the matmul, however we waste time on the copy // on M1 this is faster for the perplexity computation, but ~5% slower for the single-token generation // is there a better way? - struct ggml_tensor * V_cont = ggml_cpy(ctx0, V, ggml_new_tensor_3d(ctx0, kv_self.v->type, n_past + N, n_embd/n_head, n_head)); + struct ggml_tensor * V_cont = ggml_cpy(ctx0, V, ggml_new_tensor_3d(ctx0, kv_self.v->type, n_past + N, n_embd_head, n_head)); struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V_cont, KQ_soft_max); #endif @@ -1555,7 +1690,7 @@ static bool llama_eval_internal( { // norm { - cur = ggml_rms_norm(ctx0, inpFF); + cur = ggml_rms_norm(ctx0, inpFF, rms_norm_eps); offload_func(cur); ggml_set_name(cur, "rms_norm_1"); @@ -1603,12 +1738,9 @@ static bool llama_eval_internal( lctx.use_buf(ctx0, 0); - // used at the end to optionally extract the embeddings - struct ggml_tensor * embeddings = NULL; - // norm { - cur = ggml_rms_norm(ctx0, inpL); + cur = ggml_rms_norm(ctx0, inpL, rms_norm_eps); offload_func_nr(cur); ggml_set_name(cur, "rms_norm_2"); @@ -1616,8 +1748,6 @@ static bool llama_eval_internal( cur = ggml_mul(ctx0, cur, model.norm); // offload_func_nr(cur); // TODO CPU + GPU mirrored backend ggml_set_name(cur, "result_norm"); - - embeddings = cur; } // lm_head @@ -1629,18 +1759,103 @@ static bool llama_eval_internal( // logits -> probs //cur = ggml_soft_max_inplace(ctx0, cur); - // run the computation - ggml_build_forward_expand(&gf, cur); + ggml_build_forward_expand(gf, cur); + + if (mem_per_token == 0) { + mem_per_token = ggml_used_mem(ctx0)/N; + } + +#if 0 + LLAMA_LOG_INFO("\n%s: used_mem: eval ctx %.3f MB, scratch %.3f MB %.3f MB, work buf %.3f MB, n_past = %d, N = %d\n", __func__, + ggml_used_mem(ctx0)/1024.0/1024.0, + lctx.get_buf_max_mem(0)/1024.0/1024.0, + lctx.get_buf_max_mem(1)/1024.0/1024.0, + lctx.work_buffer.size()/1024.0/1024.0, + n_past, N); +#endif + + ggml_free(ctx0); + + return gf; +} + +// evaluate the transformer +// +// - lctx: llama context +// - tokens: new batch of tokens to process +// - embd embeddings input +// - n_tokens number of tokens +// - n_past: the context size so far +// - n_threads: number of threads to use +// +static bool llama_eval_internal( + llama_context & lctx, + const llama_token * tokens, + const float * embd, + int n_tokens, + int n_past, + int n_threads, + const char * cgraph_fname) { + + LLAMA_ASSERT((!tokens && embd) || (tokens && !embd)); + + const int64_t t_start_us = ggml_time_us(); + +#ifdef GGML_USE_MPI + ggml_mpi_eval_init(lctx.ctx_mpi, &n_tokens, &n_past, &n_threads); +#endif + + const int N = n_tokens; + + const auto & model = lctx.model; + const auto & hparams = model.hparams; + + const auto & kv_self = lctx.kv_self; + + LLAMA_ASSERT(!!kv_self.ctx); + + const int64_t n_embd = hparams.n_embd; + const int64_t n_vocab = hparams.n_vocab; + +#ifdef LLAMA_USE_ALLOCATOR + ggml_allocr_reset(lctx.alloc); +#endif + + ggml_cgraph * gf = llama_build_graph(lctx, tokens, embd, n_tokens, n_past); + +#ifdef LLAMA_USE_ALLOCATOR + ggml_allocr_alloc_graph(lctx.alloc, gf); +#endif + + // LLAMA_LOG_INFO("graph build time: %.3f ms (%d nodes, %d leafs)\n", (ggml_time_us() - t_start_us)/1000.0, gf->n_nodes, gf->n_leafs); + + // for big prompts, if BLAS is enabled, it is better to use only one thread + // otherwise, the threads are spin-lock waiting for the BLAS calls and are degrading the performance + n_threads = N >= 32 && ggml_cpu_has_blas() && !ggml_cpu_has_gpublas() ? 1 : n_threads; + + struct ggml_tensor * res = gf->nodes[gf->n_nodes - 1]; + struct ggml_tensor * embeddings = gf->nodes[gf->n_nodes - 2]; + + LLAMA_ASSERT(strcmp(res->name, "result_output") == 0); + LLAMA_ASSERT(strcmp(embeddings->name, "result_norm") == 0); #if GGML_USE_MPI - ggml_mpi_graph_compute_pre(lctx.ctx_mpi, &gf, n_layer); + const int64_t n_layer = hparams.n_layer; + ggml_mpi_graph_compute_pre(lctx.ctx_mpi, gf, n_layer); #endif #ifdef GGML_USE_METAL if (lctx.ctx_metal && N == 1) { + // TODO: disabled until #2413 is resolved + //if (!ggml_metal_if_optimized(lctx.ctx_metal)) { + // ggml_metal_graph_find_concurrency(lctx.ctx_metal, gf); + //} ggml_metal_set_n_cb (lctx.ctx_metal, n_threads); - ggml_metal_graph_compute(lctx.ctx_metal, &gf); - ggml_metal_get_tensor (lctx.ctx_metal, cur); + ggml_metal_graph_compute(lctx.ctx_metal, gf); + ggml_metal_get_tensor (lctx.ctx_metal, res); + if (!lctx.embedding.empty()) { + ggml_metal_get_tensor(lctx.ctx_metal, embeddings); + } } else { // IMPORTANT: // Since we don't have efficient Matrix x Matrix Metal multiplication yet, we fallback to vanilla @@ -1658,34 +1873,32 @@ static bool llama_eval_internal( ggml_metal_get_tensor(lctx.ctx_metal, kv_self.v); } - ggml_graph_compute_helper(lctx.work_buffer, &gf, n_threads); + ggml_graph_compute_helper(lctx.work_buffer, gf, n_threads); } #else - ggml_graph_compute_helper(lctx.work_buffer, &gf, n_threads); + ggml_graph_compute_helper(lctx.work_buffer, gf, n_threads); #endif #if GGML_USE_MPI - ggml_mpi_graph_compute_post(lctx.ctx_mpi, &gf, n_layer); + ggml_mpi_graph_compute_post(lctx.ctx_mpi, gf, n_layer); #endif // update kv token count lctx.kv_self.n = n_past + N; - struct ggml_tensor * res = gf.nodes[gf.n_nodes - 1]; - if (cgraph_fname) { - ggml_graph_export(&gf, cgraph_fname); + ggml_graph_export(gf, cgraph_fname); } #ifdef GGML_PERF // print timing information per ggml operation (for debugging purposes) // requires GGML_PERF to be defined - ggml_graph_print(&gf); + ggml_graph_print(gf); #endif // plot the computation graph in dot format (for debugging purposes) //if (n_past%100 == 0) { - // ggml_graph_dump_dot(&gf, NULL, "llama.dot"); + // ggml_graph_dump_dot(gf, NULL, "llama.dot"); //} // extract logits @@ -1710,19 +1923,6 @@ static bool llama_eval_internal( memcpy(embedding_out.data(), (float *) ggml_get_data(embeddings) + (n_embd*(N - 1)), sizeof(float)*n_embd); } - if (mem_per_token == 0) { - mem_per_token = ggml_used_mem(ctx0)/N; - } - -#if 0 - printf("\n%s: used_mem = %.3f MB, scratch -- %.3f MB %.3f MB\n", __func__, - ggml_used_mem(ctx0)/1024.0/1024.0, - lctx.get_buf_max_mem(0)/1024.0/1024.0, - lctx.get_buf_max_mem(1)/1024.0/1024.0); -#endif - - ggml_free(ctx0); - // measure the performance only for the single-token evals if (N == 1) { lctx.t_eval_us += ggml_time_us() - t_start_us; @@ -1814,7 +2014,7 @@ struct llama_tokenizer { left_sym.n += right_sym.n; right_sym.n = 0; - //printf("left = '%*s' size = %zu\n", (int) left_sym.n, left_sym.text, bigram.size); + //LLAMA_LOG_INFO("left = '%*s' size = %zu\n", (int) left_sym.n, left_sym.text, bigram.size); // remove the right sym from the chain left_sym.next = right_sym.next; @@ -1834,7 +2034,9 @@ struct llama_tokenizer { if (token == vocab_.token_to_id.end()) { // output any symbols that did not form tokens as bytes. for (int j = 0; j < (int) symbol.n; ++j) { - llama_vocab::id token_id = static_cast<uint8_t>(symbol.text[j]) + 3; + // NOTE: old version, before #2420 - not sure what are the implications of this + //llama_vocab::id token_id = static_cast<uint8_t>(symbol.text[j]) + 3; + llama_vocab::id token_id = vocab_.token_to_id.at(std::string(1, symbol.text[j])); output.push_back(token_id); } } else { @@ -1892,6 +2094,279 @@ static std::vector<llama_vocab::id> llama_tokenize(const llama_vocab & vocab, co } // +// grammar - internal +// + +struct llama_grammar { + const std::vector<std::vector<llama_grammar_element>> rules; + std::vector<std::vector<const llama_grammar_element *>> stacks; +}; + +struct llama_grammar_candidate { + size_t index; + const uint32_t * code_points; +}; + +// NOTE: assumes valid utf8 (but checks for overrun) +// adds a terminating 0 for use as pointer +std::vector<uint32_t> decode_utf8(const char * src) { + static const int lookup[] = { 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 2, 3, 4 }; + const char * pos = src; + std::vector<uint32_t> code_points; + while (*pos != 0) { + uint8_t first_byte = static_cast<uint8_t>(*pos); + uint8_t highbits = first_byte >> 4; + int len = lookup[highbits]; + uint8_t mask = (1 << (8 - len)) - 1; + uint32_t value = first_byte & mask; + const char * end = pos + len; // may overrun! + ++pos; + for ( ; pos < end && *pos != 0; ++pos) { + value = (value << 6) + (static_cast<uint8_t>(*pos) & 0x3F); + } + code_points.push_back(value); + } + code_points.push_back(0); + return code_points; +} + +// returns true iff pos points to the end of one of the definitions of a rule +static bool llama_grammar_is_end_of_sequence(const llama_grammar_element * pos) { + switch (pos->type) { + case LLAMA_GRETYPE_END: return true; + case LLAMA_GRETYPE_ALT: return true; + default: return false; + } +} + +// returns true iff chr satisfies the char range at pos (regular or inverse range) +// asserts that pos is pointing to a char range element +static std::pair<bool, const llama_grammar_element *> llama_grammar_match_char( + const llama_grammar_element * pos, + const uint32_t chr) { + + bool found = false; + bool is_positive_char = pos->type == LLAMA_GRETYPE_CHAR; + LLAMA_ASSERT(is_positive_char || pos->type == LLAMA_GRETYPE_CHAR_NOT); + + do { + if (pos[1].type == LLAMA_GRETYPE_CHAR_RNG_UPPER) { + // inclusive range, e.g. [a-z] + found = found || (pos->value <= chr && chr <= pos[1].value); + pos += 2; + } else { + // exact char match, e.g. [a] or "a" + found = found || pos->value == chr; + pos += 1; + } + } while (pos->type == LLAMA_GRETYPE_CHAR_ALT); + + return std::make_pair(found == is_positive_char, pos); +} + +// transforms a grammar pushdown stack into N possible stacks, all ending +// at a character range (terminal element) +static void llama_grammar_advance_stack( + const std::vector<std::vector<llama_grammar_element>> & rules, + const std::vector<const llama_grammar_element *> & stack, + std::vector<std::vector<const llama_grammar_element *>> & new_stacks) { + + if (stack.empty()) { + new_stacks.push_back(stack); + return; + } + + const llama_grammar_element * pos = stack.back(); + + switch (pos->type) { + case LLAMA_GRETYPE_RULE_REF: { + const size_t rule_id = static_cast<size_t>(pos->value); + const llama_grammar_element * subpos = rules[rule_id].data(); + do { + // init new stack without the top (pos) + std::vector<const llama_grammar_element *> new_stack(stack.begin(), stack.end() - 1); + if (!llama_grammar_is_end_of_sequence(pos + 1)) { + // if this rule ref is followed by another element, add that to stack + new_stack.push_back(pos + 1); + } + if (!llama_grammar_is_end_of_sequence(subpos)) { + // if alternate is nonempty, add to stack + new_stack.push_back(subpos); + } + llama_grammar_advance_stack(rules, new_stack, new_stacks); + while (!llama_grammar_is_end_of_sequence(subpos)) { + // scan to end of alternate def + subpos++; + } + if (subpos->type == LLAMA_GRETYPE_ALT) { + // there's another alternate def of this rule to process + subpos++; + } else { + break; + } + } while (true); + break; + } + case LLAMA_GRETYPE_CHAR: + case LLAMA_GRETYPE_CHAR_NOT: + new_stacks.push_back(stack); + break; + default: + // end of alternate (LLAMA_GRETYPE_END, LLAMA_GRETYPE_ALT) or middle of char range + // (LLAMA_GRETYPE_CHAR_ALT, LLAMA_GRETYPE_CHAR_RNG_UPPER); stack should never be left on + // those + LLAMA_ASSERT(false); + } +} + +// takes a set of possible pushdown stacks on a grammar, which are required to +// be positioned at a character range (see `llama_grammar_advance_stack`), and +// produces the N possible stacks if the given char is accepted at those +// positions +static std::vector<std::vector<const llama_grammar_element *>> llama_grammar_accept( + const std::vector<std::vector<llama_grammar_element>> & rules, + const std::vector<std::vector<const llama_grammar_element *>> & stacks, + const uint32_t chr) { + + std::vector<std::vector<const llama_grammar_element *>> new_stacks; + + for (const auto & stack : stacks) { + if (stack.empty()) { + continue; + } + + auto match = llama_grammar_match_char(stack.back(), chr); + if (match.first) { + const llama_grammar_element * pos = match.second; + + // update top of stack to next element, if any + std::vector<const llama_grammar_element *> new_stack(stack.begin(), stack.end() - 1); + if (!llama_grammar_is_end_of_sequence(pos)) { + new_stack.push_back(pos); + } + llama_grammar_advance_stack(rules, new_stack, new_stacks); + } + } + + return new_stacks; +} + +static std::vector<llama_grammar_candidate> llama_grammar_reject_candidates( + const std::vector<std::vector<llama_grammar_element>> & rules, + const std::vector<std::vector<const llama_grammar_element *>> & stacks, + const std::vector<llama_grammar_candidate> & candidates); + +static std::vector<llama_grammar_candidate> llama_grammar_reject_candidates_for_stack( + const std::vector<std::vector<llama_grammar_element>> & rules, + const std::vector<const llama_grammar_element *> & stack, + const std::vector<llama_grammar_candidate> & candidates) { + + std::vector<llama_grammar_candidate> rejects; + + if (stack.empty()) { + // accept nothing; EOS is handled elsewhere + rejects.insert(rejects.end(), candidates.begin(), candidates.end()); + return rejects; + } + + const llama_grammar_element * stack_pos = stack.back(); + + std::vector<llama_grammar_candidate> next_candidates; + for (auto tok : candidates) { + if (llama_grammar_match_char(stack_pos, tok.code_points[0]).first) { + if (tok.code_points[1] != 0) { + next_candidates.push_back({ tok.index, tok.code_points + 1 }); + } + } else { + rejects.push_back(tok); + } + } + + auto stack_pos_after = llama_grammar_match_char(stack_pos, 0).second; + + // update top of stack to next element, if any + std::vector<const llama_grammar_element *> stack_after(stack.begin(), stack.end() - 1); + if (!llama_grammar_is_end_of_sequence(stack_pos_after)) { + stack_after.push_back(stack_pos_after); + } + std::vector<std::vector<const llama_grammar_element *>> next_stacks; + llama_grammar_advance_stack(rules, stack_after, next_stacks); + + auto next_rejects = llama_grammar_reject_candidates(rules, next_stacks, next_candidates); + for (auto tok : next_rejects) { + rejects.push_back({ tok.index, tok.code_points - 1 }); + } + + return rejects; +} + +static std::vector<llama_grammar_candidate> llama_grammar_reject_candidates( + const std::vector<std::vector<llama_grammar_element>> & rules, + const std::vector<std::vector<const llama_grammar_element *>> & stacks, + const std::vector<llama_grammar_candidate> & candidates) { + LLAMA_ASSERT(!stacks.empty()); // REVIEW + + if (candidates.empty()) { + return std::vector<llama_grammar_candidate>(); + } + + auto rejects = llama_grammar_reject_candidates_for_stack(rules, stacks.front(), candidates); + + for (size_t i = 1, size = stacks.size(); i < size; ++i) { + rejects = llama_grammar_reject_candidates_for_stack(rules, stacks[i], rejects); + } + return rejects; +} + +// +// grammar - external +// + +struct llama_grammar * llama_grammar_init( + const llama_grammar_element ** rules, + size_t n_rules, + size_t start_rule_index) { + const llama_grammar_element * pos; + + // copy rule definitions into vectors + std::vector<std::vector<llama_grammar_element>> vec_rules(n_rules); + for (size_t i = 0; i < n_rules; i++) { + for (pos = rules[i]; pos->type != LLAMA_GRETYPE_END; pos++) { + vec_rules[i].push_back(*pos); + } + vec_rules[i].push_back({LLAMA_GRETYPE_END, 0}); + } + + // loop over alternates of start rule to build initial stacks + std::vector<std::vector<const llama_grammar_element *>> stacks; + pos = rules[start_rule_index]; + do { + std::vector<const llama_grammar_element *> stack; + if (!llama_grammar_is_end_of_sequence(pos)) { + // if alternate is nonempty, add to stack + stack.push_back(pos); + } + llama_grammar_advance_stack(vec_rules, stack, stacks); + while (!llama_grammar_is_end_of_sequence(pos)) { + // scan to end of alternate def + pos++; + } + if (pos->type == LLAMA_GRETYPE_ALT) { + // there's another alternate def of this rule to process + pos++; + } else { + break; + } + } while (true); + + return new llama_grammar{ std::move(vec_rules), std::move(stacks) }; +} + +void llama_grammar_free(struct llama_grammar * grammar) { + delete grammar; +} + +// // sampling // @@ -2006,9 +2481,18 @@ void llama_sample_tail_free(struct llama_context * ctx, llama_token_data_array * } // Normalize the second derivatives - float second_derivatives_sum = std::accumulate(second_derivatives.begin(), second_derivatives.end(), 0.0f); - for (float & value : second_derivatives) { - value /= second_derivatives_sum; + { + const float second_derivatives_sum = std::accumulate(second_derivatives.begin(), second_derivatives.end(), 0.0f); + + if (second_derivatives_sum > 1e-6f) { + for (float & value : second_derivatives) { + value /= second_derivatives_sum; + } + } else { + for (float & value : second_derivatives) { + value = 1.0f / second_derivatives.size(); + } + } } float cum_sum = 0.0f; @@ -2167,6 +2651,47 @@ void llama_sample_frequency_and_presence_penalties(struct llama_context * ctx, l } } +void llama_sample_grammar(struct llama_context * ctx, llama_token_data_array * candidates, const struct llama_grammar * grammar) { + assert(ctx); + const int64_t t_start_sample_us = ggml_time_us(); + + bool allow_eos = false; + for (const auto & stack : grammar->stacks) { + if (stack.empty()) { + allow_eos = true; + break; + } + } + + const llama_token eos = llama_token_eos(); + + std::vector<std::vector<uint32_t>> candidates_decoded; + std::vector<llama_grammar_candidate> candidates_grammar; + + for (size_t i = 0; i < candidates->size; ++i) { + const llama_token id = candidates->data[i].id; + const char * str = llama_token_to_str(ctx, id); + if (id == eos) { + if (!allow_eos) { + candidates->data[i].logit = -INFINITY; + } + } else if (*str == 0) { + candidates->data[i].logit = -INFINITY; + } else { + candidates_decoded.push_back(decode_utf8(str)); + candidates_grammar.push_back({ i, candidates_decoded.back().data() }); + } + } + + const auto rejects = + llama_grammar_reject_candidates(grammar->rules, grammar->stacks, candidates_grammar); + for (auto & reject : rejects) { + candidates->data[reject.index].logit = -INFINITY; + } + + ctx->t_sample_us += ggml_time_us() - t_start_sample_us; +} + static void llama_log_softmax(float * array, size_t size) { float max_l = *std::max_element(array, array + size); float sum = 0.f; @@ -2185,9 +2710,8 @@ void llama_sample_classifier_free_guidance( struct llama_context * ctx, llama_token_data_array * candidates, struct llama_context * guidance_ctx, - float scale, - float smooth_factor) { - int64_t t_start_sample_us = t_start_sample_us = ggml_time_us(); + float scale) { + int64_t t_start_sample_us = ggml_time_us(); assert(ctx); auto n_vocab = llama_n_vocab(ctx); @@ -2207,16 +2731,7 @@ void llama_sample_classifier_free_guidance( for (int i = 0; i < n_vocab; ++i) { float logit_guidance = logits_guidance[i]; float logit_base = logits_base[i]; - logits_guidance[i] = scale * (logit_base - logit_guidance) + logit_guidance; - } - - llama_log_softmax(logits_guidance, n_vocab); - - for (int i = 0; i < n_vocab; ++i) { - float logit_base = logits_base[i]; - float logit_guidance = logits_guidance[i]; - - candidates->data[i].logit = smooth_factor * logit_guidance + (1.f - smooth_factor) * logit_base; + candidates->data[i].logit = scale * (logit_base - logit_guidance) + logit_guidance; } if (ctx) { @@ -2352,6 +2867,29 @@ llama_token llama_sample_token(struct llama_context * ctx, llama_token_data_arra return result; } +void llama_grammar_accept_token(struct llama_context * ctx, struct llama_grammar * grammar, llama_token token) { + const int64_t t_start_sample_us = ggml_time_us(); + + if (token == llama_token_eos()) { + for (const auto & stack : grammar->stacks) { + if (stack.empty()) { + return; + } + } + LLAMA_ASSERT(false); + } + + const char * str = llama_token_to_str(ctx, token); + // Note terminating 0 in decoded string + auto code_points = decode_utf8(str); + for (auto it = code_points.begin(), end = code_points.end() - 1; it != end; ++it) { + grammar->stacks = llama_grammar_accept(grammar->rules, grammar->stacks, *it); + } + LLAMA_ASSERT(!grammar->stacks.empty()); + + ctx->t_sample_us += ggml_time_us() - t_start_sample_us; +} + // // quantization // @@ -2425,8 +2963,8 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s case LLAMA_FTYPE_MOSTLY_Q5_0: quantized_type = GGML_TYPE_Q5_0; break; case LLAMA_FTYPE_MOSTLY_Q5_1: quantized_type = GGML_TYPE_Q5_1; break; case LLAMA_FTYPE_MOSTLY_Q8_0: quantized_type = GGML_TYPE_Q8_0; break; - case LLAMA_FTYPE_MOSTLY_F16: quantized_type = GGML_TYPE_F16; break; - case LLAMA_FTYPE_ALL_F32: quantized_type = GGML_TYPE_F32; break; + case LLAMA_FTYPE_MOSTLY_F16: quantized_type = GGML_TYPE_F16; break; + case LLAMA_FTYPE_ALL_F32: quantized_type = GGML_TYPE_F32; break; #ifdef GGML_USE_K_QUANTS // K-quants @@ -2484,7 +3022,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s tensor.data = read_data.addr; model_loader->load_data_for(tensor); - printf("[%4zu/%4zu] %36s - %16s, type = %6s, ", + LLAMA_LOG_INFO("[%4zu/%4zu] %36s - %16s, type = %6s, ", ++idx, model_loader->tensors_map.tensors.size(), tensor.name.c_str(), llama_format_tensor_shape(tensor.ne).c_str(), ggml_type_name(tensor.type)); @@ -2506,20 +3044,10 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s new_type = tensor.type; new_data = tensor.data; new_size = tensor.size; - printf("size = %8.3f MB\n", tensor.size/1024.0/1024.0); + LLAMA_LOG_INFO("size = %8.3f MB\n", tensor.size/1024.0/1024.0); } else { new_type = quantized_type; #ifdef GGML_USE_K_QUANTS - bool convert_incompatible_tensor = false; - if (quantized_type == GGML_TYPE_Q2_K || quantized_type == GGML_TYPE_Q3_K || quantized_type == GGML_TYPE_Q4_K || - quantized_type == GGML_TYPE_Q5_K || quantized_type == GGML_TYPE_Q6_K) { - int nx = tensor.ne.at(0); - int ny = tensor.ne.at(1); - if (nx % QK_K != 0 || ny % QK_K != 0) { - fprintf(stderr, "\n\nTensor sizes %d x %d are not divisible by %d, required for k-quants.\n",nx,ny,QK_K); - convert_incompatible_tensor = true; - } - } if (tensor.name == "output.weight") { int nx = tensor.ne.at(0); int ny = tensor.ne.at(1); @@ -2545,13 +3073,23 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q4_K; else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) new_type = GGML_TYPE_Q5_K; } + bool convert_incompatible_tensor = false; + if (new_type == GGML_TYPE_Q2_K || new_type == GGML_TYPE_Q3_K || new_type == GGML_TYPE_Q4_K || + new_type == GGML_TYPE_Q5_K || new_type == GGML_TYPE_Q6_K) { + int nx = tensor.ne.at(0); + int ny = tensor.ne.at(1); + if (nx % QK_K != 0 || ny % QK_K != 0) { + LLAMA_LOG_INFO("\n\nTensor sizes %d x %d are not divisible by %d, required for k-quants.\n",nx,ny,QK_K); + convert_incompatible_tensor = true; + } + } if (convert_incompatible_tensor) { if (tensor.name == "output.weight") { new_type = GGML_TYPE_F16; //fall back to F16 instead of just failing. - fprintf(stderr, "F16 will be used for this tensor instead.\n"); + LLAMA_LOG_WARN("F16 will be used for this tensor instead.\n"); } else if (tensor.name == "tok_embeddings.weight") { new_type = GGML_TYPE_Q4_0; //fall back to Q4_0 instead of just failing. - fprintf(stderr, "Q4_0 will be used for this tensor instead.\n"); + LLAMA_LOG_WARN("Q4_0 will be used for this tensor instead.\n"); } else { throw std::runtime_error("Unsupported tensor size encountered\n"); } @@ -2571,7 +3109,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s f32_data = (float *) f32_conv_buf.addr; } - printf("quantizing .. "); + LLAMA_LOG_INFO("quantizing to %s .. ", ggml_type_name(new_type)); fflush(stdout); work.resize(nelements * 4); // upper bound on size @@ -2621,7 +3159,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s } } - printf("size = %8.2f MB -> %8.2f MB | hist: ", tensor.size/1024.0/1024.0, new_size/1024.0/1024.0); + LLAMA_LOG_INFO("size = %8.2f MB -> %8.2f MB | hist: ", tensor.size/1024.0/1024.0, new_size/1024.0/1024.0); int64_t tot_count = 0; for (size_t i = 0; i < hist_cur.size(); i++) { hist_all[i] += hist_cur[i]; @@ -2630,18 +3168,18 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s if (tot_count > 0) { for (size_t i = 0; i < hist_cur.size(); i++) { - printf("%5.3f ", hist_cur[i] / float(nelements)); + LLAMA_LOG_INFO("%5.3f ", hist_cur[i] / float(nelements)); } } - printf("\n"); + LLAMA_LOG_INFO("\n"); } total_size_org += tensor.size; total_size_new += new_size; file_saver.write_tensor(tensor, new_type, new_data, new_size); } - printf("%s: model size = %8.2f MB\n", __func__, total_size_org/1024.0/1024.0); - printf("%s: quant size = %8.2f MB\n", __func__, total_size_new/1024.0/1024.0); + LLAMA_LOG_INFO("%s: model size = %8.2f MB\n", __func__, total_size_org/1024.0/1024.0); + LLAMA_LOG_INFO("%s: quant size = %8.2f MB\n", __func__, total_size_new/1024.0/1024.0); { int64_t sum_all = 0; @@ -2650,11 +3188,11 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s } if (sum_all > 0) { - printf("%s: hist: ", __func__); + LLAMA_LOG_INFO("%s: hist: ", __func__); for (size_t i = 0; i < hist_all.size(); i++) { - printf("%5.3f ", hist_all[i] / float(sum_all)); + LLAMA_LOG_INFO("%5.3f ", hist_all[i] / float(sum_all)); } - printf("\n"); + LLAMA_LOG_INFO("\n"); } } } @@ -2674,11 +3212,12 @@ struct llama_model * llama_load_model_from_file( ggml_type memory_type = params.f16_kv ? GGML_TYPE_F16 : GGML_TYPE_F32; - if (!llama_model_load(path_model, *model, model->vocab, params.n_ctx, params.n_batch, params.n_gpu_layers, - params.main_gpu, params.tensor_split, params.low_vram, memory_type, params.use_mmap, params.use_mlock, - params.vocab_only, params.progress_callback, params.progress_callback_user_data)) { + if (!llama_model_load(path_model, *model, model->vocab, params.n_ctx, params.n_batch, params.n_gqa, params.rms_norm_eps, params.n_gpu_layers, + params.main_gpu, params.tensor_split, params.mul_mat_q, params.rope_freq_base, params.rope_freq_scale,params.low_vram, + memory_type, params.use_mmap, params.use_mlock, params.vocab_only, params.progress_callback, + params.progress_callback_user_data)) { + LLAMA_LOG_ERROR("%s: failed to load model\n", __func__); delete model; - fprintf(stderr, "%s: failed to load model\n", __func__); return nullptr; } @@ -2697,7 +3236,7 @@ struct llama_context * llama_new_context_with_model( return nullptr; } - llama_context * ctx = new llama_context(*model, model->vocab); + llama_context * ctx = new llama_context(*model); if (params.seed == LLAMA_DEFAULT_SEED) { params.seed = time(NULL); @@ -2711,10 +3250,9 @@ struct llama_context * llama_new_context_with_model( unsigned percentage = (unsigned) (100 * progress); while (percentage > *cur_percentage_p) { *cur_percentage_p = percentage; - fprintf(stderr, "."); - fflush(stderr); + LLAMA_LOG_INFO("."); if (percentage >= 100) { - fprintf(stderr, "\n"); + LLAMA_LOG_INFO("\n"); } } }; @@ -2728,14 +3266,14 @@ struct llama_context * llama_new_context_with_model( // reserve memory for context buffers if (!params.vocab_only) { if (!kv_cache_init(ctx->model.hparams, ctx->kv_self, memory_type, ctx->model.hparams.n_ctx, params.n_gpu_layers)) { - fprintf(stderr, "%s: kv_cache_init() failed for self-attention cache\n", __func__); + LLAMA_LOG_ERROR("%s: kv_cache_init() failed for self-attention cache\n", __func__); llama_free(ctx); return nullptr; } { const size_t memory_size = ggml_nbytes(ctx->kv_self.k) + ggml_nbytes(ctx->kv_self.v); - fprintf(stderr, "%s: kv self size = %7.2f MB\n", __func__, memory_size / 1024.0 / 1024.0); + LLAMA_LOG_INFO("%s: kv self size = %7.2f MB\n", __func__, memory_size / 1024.0 / 1024.0); } const auto & hparams = ctx->model.hparams; @@ -2751,10 +3289,47 @@ struct llama_context * llama_new_context_with_model( ctx->embedding.resize(hparams.n_embd); } - ctx->buf_compute.resize(MEM_REQ_EVAL().at(ctx->model.type)); +#ifdef LLAMA_USE_ALLOCATOR + { + static const size_t tensor_alignment = 32; + // the compute buffer is used to store the tensor and graph structs, while the allocator buffer is used for the tensor data + ctx->buf_compute.resize(ggml_tensor_overhead()*GGML_MAX_NODES + ggml_graph_overhead()); + + // create measure allocator + ctx->alloc = ggml_allocr_new_measure(tensor_alignment); + + // build worst-case graph + int n_tokens = std::min((int)hparams.n_ctx, params.n_batch); + int n_past = hparams.n_ctx - n_tokens; + llama_token token = llama_token_bos(); // not actually used by llama_build_graph, but required to choose between token and embedding inputs graph + ggml_cgraph * gf = llama_build_graph(*ctx, &token, NULL, n_tokens, n_past); + + // measure memory requirements for the graph + size_t alloc_size = ggml_allocr_alloc_graph(ctx->alloc, gf) + tensor_alignment; + + LLAMA_LOG_INFO("%s: compute buffer total size = %7.2f MB\n", __func__, (ctx->buf_compute.size + alloc_size) / 1024.0 / 1024.0); + + // debug - for comparison with scratch buffer + //size_t prev_req = + // MEM_REQ_SCRATCH0(hparams.n_ctx).at(ctx->model.type) + + // MEM_REQ_SCRATCH1().at(ctx->model.type) + + // MEM_REQ_EVAL().at(ctx->model.type); + //LLAMA_LOG_INFO("%s: (debug) equivalent with scratch buffer = %7.2f MB\n", __func__, prev_req / 1024.0 / 1024.0); + + // recreate allocator with exact memory requirements + ggml_allocr_free(ctx->alloc); + + ctx->buf_alloc.resize(alloc_size); + ctx->alloc = ggml_allocr_new(ctx->buf_alloc.addr, ctx->buf_alloc.size, tensor_alignment); + } +#else + ctx->buf_compute.resize(MEM_REQ_EVAL().at(ctx->model.type) + ggml_graph_overhead()); +#endif - ctx->buf_scratch[0].resize(MEM_REQ_SCRATCH0().at(ctx->model.type)); +#ifdef LLAMA_USE_SCRATCH + ctx->buf_scratch[0].resize(MEM_REQ_SCRATCH0(hparams.n_ctx).at(ctx->model.type)); ctx->buf_scratch[1].resize(MEM_REQ_SCRATCH1().at(ctx->model.type)); +#endif } #ifdef GGML_USE_METAL @@ -2775,13 +3350,13 @@ struct llama_context * llama_new_context_with_model( const size_t max_size = ggml_get_max_tensor_size(ctx->model.ctx); - printf("%s: max tensor size = %8.2f MB\n", __func__, max_size/1024.0/1024.0); + LLAMA_LOG_INFO("%s: max tensor size = %8.2f MB\n", __func__, max_size/1024.0/1024.0); -#define LLAMA_METAL_CHECK_BUF(result) \ - if (!(result)) { \ - fprintf(stderr, "%s: failed to add buffer\n", __func__); \ - llama_free(ctx); \ - return NULL; \ +#define LLAMA_METAL_CHECK_BUF(result) \ + if (!(result)) { \ + LLAMA_LOG_ERROR("%s: failed to add buffer\n", __func__); \ + llama_free(ctx); \ + return NULL; \ } LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "data", data_ptr, data_size, max_size)); @@ -2824,9 +3399,6 @@ struct llama_context * llama_init_from_file( } void llama_free(struct llama_context * ctx) { - if (ctx->model_owner) { - delete &ctx->model; - } delete ctx; } @@ -2838,19 +3410,19 @@ int llama_model_quantize( llama_model_quantize_internal(fname_inp, fname_out, params); return 0; } catch (const std::exception & err) { - fprintf(stderr, "%s: failed to quantize: %s\n", __func__, err.what()); + LLAMA_LOG_ERROR("%s: failed to quantize: %s\n", __func__, err.what()); return 1; } } int llama_apply_lora_from_file_internal(const struct llama_model & model, const char * path_lora, const char * path_base_model, int n_threads) { - fprintf(stderr, "%s: applying lora adapter from '%s' - please wait ...\n", __func__, path_lora); + LLAMA_LOG_INFO("%s: applying lora adapter from '%s' - please wait ...\n", __func__, path_lora); const int64_t t_start_lora_us = ggml_time_us(); auto fin = std::ifstream(path_lora, std::ios::binary); if (!fin) { - fprintf(stderr, "%s: failed to open '%s'\n", __func__, path_lora); + LLAMA_LOG_ERROR("%s: failed to open '%s'\n", __func__, path_lora); return 1; } @@ -2859,14 +3431,14 @@ int llama_apply_lora_from_file_internal(const struct llama_model & model, const uint32_t magic; fin.read((char *) &magic, sizeof(magic)); if (magic != LLAMA_FILE_MAGIC_GGLA) { - fprintf(stderr, "%s: bad file magic\n", __func__); + LLAMA_LOG_ERROR("%s: bad file magic\n", __func__); return 1; } uint32_t format_version; fin.read((char *) &format_version, sizeof(format_version)); if (format_version != 1) { - fprintf(stderr, "%s: unsupported file version\n", __func__ ); + LLAMA_LOG_ERROR("%s: unsupported file version\n", __func__ ); return 1; } } @@ -2877,7 +3449,7 @@ int llama_apply_lora_from_file_internal(const struct llama_model & model, const fin.read((char *) &lora_alpha, sizeof(lora_alpha)); float scaling = (float)lora_alpha / (float)lora_r; - fprintf(stderr, "%s: r = %d, alpha = %d, scaling = %.2f\n", __func__, lora_r, lora_alpha, scaling); + LLAMA_LOG_INFO("%s: r = %d, alpha = %d, scaling = %.2f\n", __func__, lora_r, lora_alpha, scaling); // create a temporary ggml context to store the lora tensors @@ -2903,7 +3475,7 @@ int llama_apply_lora_from_file_internal(const struct llama_model & model, const ggml_context * base_ctx = NULL; llama_buffer base_buf; if (path_base_model) { - fprintf(stderr, "%s: loading base model from '%s'\n", __func__, path_base_model); + LLAMA_LOG_INFO("%s: loading base model from '%s'\n", __func__, path_base_model); model_loader.reset(new llama_model_loader(path_base_model, /*use_mmap*/ true)); size_t ctx_size; @@ -2960,17 +3532,17 @@ int llama_apply_lora_from_file_internal(const struct llama_model & model, const const std::string lora_suffix = ".lora"; size_t pos = name.rfind(lora_suffix); if (pos == std::string::npos) { - fprintf(stderr, "%s: error: '%s' is not a lora tensor\n", __func__, name.c_str()); + LLAMA_LOG_ERROR("%s: error: '%s' is not a lora tensor\n", __func__, name.c_str()); return 1; } std::string lora_type = name.substr(pos + lora_suffix.length()); std::string base_name = name; base_name.erase(pos); - // fprintf(stderr, "%s: %s => %s (lora type %s) ", __func__, name.c_str(),base_name.c_str(), lora_type.c_str()); + // LLAMA_LOG_INFO("%s: %s => %s (lora type %s) \n", __func__, name.c_str(),base_name.c_str(), lora_type.c_str()); if (model_tensors.find(base_name) == model_tensors.end()) { - fprintf(stderr, "%s: unknown tensor '%s' in lora adapter\n", __func__, name.data()); + LLAMA_LOG_ERROR("%s: unknown tensor '%s' in lora adapter\n", __func__, name.data()); return 1; } @@ -2981,7 +3553,7 @@ int llama_apply_lora_from_file_internal(const struct llama_model & model, const case 1: wtype = GGML_TYPE_F16; break; default: { - fprintf(stderr, "%s: invalid tensor data type '%d'\n", + LLAMA_LOG_ERROR("%s: invalid tensor data type '%d'\n", __func__, ftype); return false; } @@ -2991,7 +3563,7 @@ int llama_apply_lora_from_file_internal(const struct llama_model & model, const lora_tensor = ggml_new_tensor_2d(lora_ctx, wtype, ne[0], ne[1]); } else { - fprintf(stderr, "%s: unsupported tensor dimension %d\n", __func__, n_dims); + LLAMA_LOG_ERROR("%s: unsupported tensor dimension %d\n", __func__, n_dims); return 1; } ggml_set_name(lora_tensor, "lora_tensor"); @@ -3029,7 +3601,7 @@ int llama_apply_lora_from_file_internal(const struct llama_model & model, const if (model_loader) { // load from base model if (model_loader->tensors_map.name_to_idx.find(base_name) == model_loader->tensors_map.name_to_idx.end()) { - fprintf(stderr, "%s: error: tensor '%s' not found in base model\n", __func__, base_name.c_str()); + LLAMA_LOG_ERROR("%s: error: tensor '%s' not found in base model\n", __func__, base_name.c_str()); return 1; } size_t idx = model_loader->tensors_map.name_to_idx[base_name]; @@ -3045,8 +3617,8 @@ int llama_apply_lora_from_file_internal(const struct llama_model & model, const if (ggml_is_quantized(base_t->type)) { if (!warned) { - fprintf(stderr, "%s: warning: using a lora adapter with a quantized model may result in poor quality, " - "use a f16 or f32 base model with --lora-base\n", __func__); + LLAMA_LOG_WARN("%s: warning: using a lora adapter with a quantized model may result in poor quality, " + "use a f16 or f32 base model with --lora-base\n", __func__); warned = true; } } @@ -3060,8 +3632,8 @@ int llama_apply_lora_from_file_internal(const struct llama_model & model, const ggml_set_name(loraB, "loraB"); if (base_t->ne[0] != loraA->ne[1] || base_t->ne[1] != loraB->ne[1]) { - fprintf(stderr, "%s: incompatible tensor dimensions (%" PRId64 " and %" PRId64 ");" - " are you sure that this adapter is for this model?\n", __func__, base_t->ne[0], loraA->ne[1]); + LLAMA_LOG_ERROR("%s: incompatible tensor dimensions (%" PRId64 " and %" PRId64 ");" + " are you sure that this adapter is for this model?\n", __func__, base_t->ne[0], loraA->ne[1]); return 1; } @@ -3106,7 +3678,7 @@ int llama_apply_lora_from_file_internal(const struct llama_model & model, const n_tensors++; if (n_tensors % 4 == 0) { - fprintf(stderr, "."); + LLAMA_LOG_INFO("."); } } } @@ -3118,7 +3690,7 @@ int llama_apply_lora_from_file_internal(const struct llama_model & model, const } const int64_t t_lora_us = ggml_time_us() - t_start_lora_us; - fprintf(stderr, " done (%.2f ms)\n", t_lora_us / 1000.0); + LLAMA_LOG_INFO(" done (%.2f ms)\n", t_lora_us / 1000.0); return 0; } @@ -3127,7 +3699,7 @@ int llama_apply_lora_from_file(struct llama_context * ctx, const char * path_lor try { return llama_apply_lora_from_file_internal(ctx->model, path_lora, path_base_model, n_threads); } catch (const std::exception & err) { - fprintf(stderr, "%s: failed to apply lora adapter: %s\n", __func__, err.what()); + LLAMA_LOG_ERROR("%s: failed to apply lora adapter: %s\n", __func__, err.what()); return 1; } } @@ -3136,7 +3708,7 @@ int llama_model_apply_lora_from_file(const struct llama_model * model, const cha try { return llama_apply_lora_from_file_internal(*model, path_lora, path_base_model, n_threads); } catch (const std::exception & err) { - fprintf(stderr, "%s: failed to apply lora adapter: %s\n", __func__, err.what()); + LLAMA_LOG_ERROR("%s: failed to apply lora adapter: %s\n", __func__, err.what()); return 1; } } @@ -3185,10 +3757,20 @@ size_t llama_get_state_size(const struct llama_context * ctx) { return s_total; } -// Copies the state to the specified destination address -size_t llama_copy_state_data(struct llama_context * ctx, uint8_t * dst) { - uint8_t * out = dst; - +/** copy state data into either a buffer or file depending on the passed in context + * + * file context: + * llama_file file("/path", "wb"); + * llama_data_file_context data_ctx(&file); + * llama_copy_state_data(ctx, &data_ctx); + * + * buffer context: + * std::vector<uint8_t> buf(max_size, 0); + * llama_data_buffer_context data_ctx(&buf.data()); + * llama_copy_state_data(ctx, &data_ctx); + * +*/ +void llama_copy_state_data_internal(struct llama_context * ctx, llama_data_context * data_ctx) { // copy rng { std::stringstream rng_ss; @@ -3200,8 +3782,8 @@ size_t llama_copy_state_data(struct llama_context * ctx, uint8_t * dst) { memset(&rng_buf[0], 0, LLAMA_MAX_RNG_STATE); memcpy(&rng_buf[0], rng_ss.str().data(), rng_ss.str().size()); - memcpy(out, &rng_size, sizeof(rng_size)); out += sizeof(rng_size); - memcpy(out, &rng_buf[0], LLAMA_MAX_RNG_STATE); out += LLAMA_MAX_RNG_STATE; + data_ctx->write(&rng_size, sizeof(rng_size)); + data_ctx->write(&rng_buf[0], LLAMA_MAX_RNG_STATE); } // copy logits @@ -3209,25 +3791,29 @@ size_t llama_copy_state_data(struct llama_context * ctx, uint8_t * dst) { const size_t logits_cap = ctx->logits.capacity(); const size_t logits_size = ctx->logits.size(); - memcpy(out, &logits_cap, sizeof(logits_cap)); out += sizeof(logits_cap); - memcpy(out, &logits_size, sizeof(logits_size)); out += sizeof(logits_size); + data_ctx->write(&logits_cap, sizeof(logits_cap)); + data_ctx->write(&logits_size, sizeof(logits_size)); if (logits_size) { - memcpy(out, ctx->logits.data(), logits_size * sizeof(float)); + data_ctx->write(ctx->logits.data(), logits_size * sizeof(float)); } - out += logits_cap * sizeof(float); + // If there is a gap between the size and the capacity, write padding + size_t padding_size = (logits_cap - logits_size) * sizeof(float); + if (padding_size > 0) { + std::vector<uint8_t> padding(padding_size, 0); // Create a buffer filled with zeros + data_ctx->write(padding.data(), padding_size); + } } // copy embeddings { const size_t embedding_size = ctx->embedding.size(); - memcpy(out, &embedding_size, sizeof(embedding_size)); out += sizeof(embedding_size); + data_ctx->write(&embedding_size, sizeof(embedding_size)); if (embedding_size) { - memcpy(out, ctx->embedding.data(), embedding_size * sizeof(float)); - out += embedding_size * sizeof(float); + data_ctx->write(ctx->embedding.data(), embedding_size * sizeof(float)); } } @@ -3236,14 +3822,14 @@ size_t llama_copy_state_data(struct llama_context * ctx, uint8_t * dst) { const auto & kv_self = ctx->kv_self; const auto & hparams = ctx->model.hparams; const int n_layer = hparams.n_layer; - const int n_embd = hparams.n_embd; + const int n_embd = hparams.n_embd_gqa(); const int n_ctx = hparams.n_ctx; const size_t kv_size = kv_self.buf.size; const int kv_ntok = llama_get_kv_cache_token_count(ctx); - memcpy(out, &kv_size, sizeof(kv_size)); out += sizeof(kv_size); - memcpy(out, &kv_ntok, sizeof(kv_ntok)); out += sizeof(kv_ntok); + data_ctx->write(&kv_size, sizeof(kv_size)); + data_ctx->write(&kv_ntok, sizeof(kv_ntok)); if (kv_size) { const size_t elt_size = ggml_element_size(kv_self.k); @@ -3252,12 +3838,12 @@ size_t llama_copy_state_data(struct llama_context * ctx, uint8_t * dst) { ggml_cgraph gf{}; ggml_tensor * kout3d = ggml_new_tensor_3d(cpy_ctx, kv_self.k->type, n_embd, kv_ntok, n_layer); - kout3d->data = out; - out += ggml_nbytes(kout3d); + std::vector<uint8_t> kout3d_data(ggml_nbytes(kout3d), 0); + kout3d->data = kout3d_data.data(); ggml_tensor * vout3d = ggml_new_tensor_3d(cpy_ctx, kv_self.v->type, kv_ntok, n_embd, n_layer); - vout3d->data = out; - out += ggml_nbytes(vout3d); + std::vector<uint8_t> vout3d_data(ggml_nbytes(vout3d), 0); + vout3d->data = vout3d_data.data(); ggml_tensor * k3d = ggml_view_3d(cpy_ctx, kv_self.k, n_embd, kv_ntok, n_layer, @@ -3272,15 +3858,20 @@ size_t llama_copy_state_data(struct llama_context * ctx, uint8_t * dst) { ggml_graph_compute_helper(ctx->work_buffer, &gf, /*n_threads*/ 1); ggml_free(cpy_ctx); + + // our data is now in the kout3d_data and vout3d_data buffers + // write them to file + data_ctx->write(kout3d_data.data(), kout3d_data.size()); + data_ctx->write(vout3d_data.data(), vout3d_data.size()); } } +} - const size_t written = out - dst; - const size_t max_size = llama_get_state_size(ctx); - - LLAMA_ASSERT(written <= max_size); +size_t llama_copy_state_data(struct llama_context * ctx, uint8_t * dst) { + llama_data_buffer_context data_ctx(dst); + llama_copy_state_data_internal(ctx, &data_ctx); - return written; + return data_ctx.get_size_written(); } // Sets the state reading from the specified source address @@ -3339,7 +3930,7 @@ size_t llama_set_state_data(struct llama_context * ctx, uint8_t * src) { const auto & kv_self = ctx->kv_self; const auto & hparams = ctx->model.hparams; const int n_layer = hparams.n_layer; - const int n_embd = hparams.n_embd; + const int n_embd = hparams.n_embd_gqa(); const int n_ctx = hparams.n_ctx; size_t kv_size; @@ -3399,7 +3990,7 @@ static bool llama_load_session_file_internal(struct llama_context * ctx, const c const uint32_t version = file.read_u32(); if (magic != LLAMA_SESSION_MAGIC || version != LLAMA_SESSION_VERSION) { - fprintf(stderr, "%s : unknown (magic, version) for session file: %08x, %08x\n", __func__, magic, version); + LLAMA_LOG_ERROR("%s : unknown (magic, version) for session file: %08x, %08x\n", __func__, magic, version); return false; } @@ -3407,7 +3998,7 @@ static bool llama_load_session_file_internal(struct llama_context * ctx, const c file.read_raw(&session_hparams, sizeof(llama_hparams)); if (session_hparams != ctx->model.hparams) { - fprintf(stderr, "%s : model hparams didn't match from session file!\n", __func__); + LLAMA_LOG_INFO("%s : model hparams didn't match from session file!\n", __func__); return false; } } @@ -3417,7 +4008,7 @@ static bool llama_load_session_file_internal(struct llama_context * ctx, const c const uint32_t n_token_count = file.read_u32(); if (n_token_count > n_token_capacity) { - fprintf(stderr, "%s : token count in session file exceeded capacity! %u > %zu\n", __func__, n_token_count, n_token_capacity); + LLAMA_LOG_ERROR("%s : token count in session file exceeded capacity! %u > %zu\n", __func__, n_token_count, n_token_capacity); return false; } @@ -3431,7 +4022,7 @@ static bool llama_load_session_file_internal(struct llama_context * ctx, const c const size_t n_state_size_max = llama_get_state_size(ctx); if (n_state_size_cur > n_state_size_max) { - fprintf(stderr, "%s : the state size in session file is too big! max %zu, got %zu\n", __func__, n_state_size_max, n_state_size_cur); + LLAMA_LOG_ERROR("%s : the state size in session file is too big! max %zu, got %zu\n", __func__, n_state_size_max, n_state_size_cur); return false; } @@ -3448,7 +4039,7 @@ bool llama_load_session_file(struct llama_context * ctx, const char * path_sessi try { return llama_load_session_file_internal(ctx, path_session, tokens_out, n_token_capacity, n_token_count_out); } catch (const std::exception & err) { - fprintf(stderr, "error loading session file: %s\n", err.what()); + LLAMA_LOG_ERROR("error loading session file: %s\n", err.what()); return false; } } @@ -3465,15 +4056,9 @@ bool llama_save_session_file(struct llama_context * ctx, const char * path_sessi file.write_u32((uint32_t) n_token_count); file.write_raw(tokens, sizeof(llama_token) * n_token_count); - // save the context state - { - const size_t n_state_size_max = llama_get_state_size(ctx); - - std::vector<uint8_t> state_data(n_state_size_max); - const size_t n_state_size_cur = llama_copy_state_data(ctx, state_data.data()); - - file.write_raw(state_data.data(), n_state_size_cur); - } + // save the context state using stream saving + llama_data_file_context data_ctx(&file); + llama_copy_state_data_internal(ctx, &data_ctx); return true; } @@ -3485,7 +4070,7 @@ int llama_eval( int n_past, int n_threads) { if (!llama_eval_internal(*ctx, tokens, nullptr, n_tokens, n_past, n_threads, nullptr)) { - fprintf(stderr, "%s: failed to eval\n", __func__); + LLAMA_LOG_ERROR("%s: failed to eval\n", __func__); return 1; } @@ -3507,7 +4092,7 @@ int llama_eval_embd( int n_past, int n_threads) { if (!llama_eval_internal(*ctx, nullptr, embd, n_tokens, n_past, n_threads, nullptr)) { - fprintf(stderr, "%s: failed to eval\n", __func__); + LLAMA_LOG_ERROR("%s: failed to eval\n", __func__); return 1; } @@ -3528,23 +4113,23 @@ int llama_eval_export(struct llama_context * ctx, const char * fname) { const std::vector<llama_token> tmp(n_batch, llama_token_bos()); if (!llama_eval_internal(*ctx, tmp.data(), nullptr, tmp.size(), n_ctx, 1, fname)) { - fprintf(stderr, "%s: failed to eval\n", __func__); + LLAMA_LOG_ERROR("%s: failed to eval\n", __func__); return 1; } return 0; } -int llama_tokenize( - struct llama_context * ctx, +int llama_tokenize_with_model( + const struct llama_model * model, const char * text, llama_token * tokens, int n_max_tokens, bool add_bos) { - auto res = llama_tokenize(ctx->vocab, text, add_bos); + auto res = llama_tokenize(model->vocab, text, add_bos); if (n_max_tokens < (int) res.size()) { - fprintf(stderr, "%s: too many tokens\n", __func__); + LLAMA_LOG_ERROR("%s: too many tokens\n", __func__); return -((int) res.size()); } @@ -3555,8 +4140,29 @@ int llama_tokenize( return res.size(); } +int llama_tokenize( + struct llama_context * ctx, + const char * text, + llama_token * tokens, + int n_max_tokens, + bool add_bos) { + return llama_tokenize_with_model(&ctx->model, text, tokens, n_max_tokens, add_bos); +} + +int llama_n_vocab_from_model(const struct llama_model * model) { + return model->vocab.id_to_token.size(); +} + +int llama_n_ctx_from_model(const struct llama_model * model) { + return model->hparams.n_ctx; +} + +int llama_n_embd_from_model(const struct llama_model * model) { + return model->hparams.n_embd; +} + int llama_n_vocab(const struct llama_context * ctx) { - return ctx->vocab.id_to_token.size(); + return ctx->model.vocab.id_to_token.size(); } int llama_n_ctx(const struct llama_context * ctx) { @@ -3567,19 +4173,27 @@ int llama_n_embd(const struct llama_context * ctx) { return ctx->model.hparams.n_embd; } -int llama_get_vocab( - const struct llama_context * ctx, +int llama_get_vocab_from_model( + const struct llama_model * model, const char * * strings, float * scores, int capacity) { - int n = std::min(capacity, (int) ctx->vocab.id_to_token.size()); + int n = std::min(capacity, (int) model->vocab.id_to_token.size()); for (int i = 0; i<n; ++i) { - strings[i] = ctx->vocab.id_to_token[i].tok.c_str(); - scores[i] = ctx->vocab.id_to_token[i].score; + strings[i] = model->vocab.id_to_token[i].tok.c_str(); + scores[i] = model->vocab.id_to_token[i].score; } return n; } +int llama_get_vocab( + const struct llama_context * ctx, + const char * * strings, + float * scores, + int capacity) { + return llama_get_vocab_from_model(&ctx->model, strings, scores, capacity); +} + float * llama_get_logits(struct llama_context * ctx) { return ctx->logits.data(); } @@ -3588,12 +4202,16 @@ float * llama_get_embeddings(struct llama_context * ctx) { return ctx->embedding.data(); } -const char * llama_token_to_str(const struct llama_context * ctx, llama_token token) { - if (token >= llama_n_vocab(ctx)) { +const char * llama_token_to_str_with_model(const struct llama_model * model, llama_token token) { + if (token >= llama_n_vocab_from_model(model)) { return nullptr; } - return ctx->vocab.id_to_token[token].tok.c_str(); + return model->vocab.id_to_token[token].tok.c_str(); +} + +const char * llama_token_to_str(const struct llama_context * ctx, llama_token token) { + return llama_token_to_str_with_model(&ctx->model, token); } llama_token llama_token_bos() { @@ -3628,15 +4246,15 @@ struct llama_timings llama_get_timings(struct llama_context * ctx) { void llama_print_timings(struct llama_context * ctx) { const llama_timings timings = llama_get_timings(ctx); - fprintf(stderr, "\n"); - fprintf(stderr, "%s: load time = %8.2f ms\n", __func__, timings.t_load_ms); - fprintf(stderr, "%s: sample time = %8.2f ms / %5d runs (%8.2f ms per token, %8.2f tokens per second)\n", + LLAMA_LOG_INFO("\n"); + LLAMA_LOG_INFO("%s: load time = %8.2f ms\n", __func__, timings.t_load_ms); + LLAMA_LOG_INFO("%s: sample time = %8.2f ms / %5d runs (%8.2f ms per token, %8.2f tokens per second)\n", __func__, timings.t_sample_ms, timings.n_sample, timings.t_sample_ms / timings.n_sample, 1e3 / timings.t_sample_ms * timings.n_sample); - fprintf(stderr, "%s: prompt eval time = %8.2f ms / %5d tokens (%8.2f ms per token, %8.2f tokens per second)\n", + LLAMA_LOG_INFO("%s: prompt eval time = %8.2f ms / %5d tokens (%8.2f ms per token, %8.2f tokens per second)\n", __func__, timings.t_p_eval_ms, timings.n_p_eval, timings.t_p_eval_ms / timings.n_p_eval, 1e3 / timings.t_p_eval_ms * timings.n_p_eval); - fprintf(stderr, "%s: eval time = %8.2f ms / %5d runs (%8.2f ms per token, %8.2f tokens per second)\n", + LLAMA_LOG_INFO("%s: eval time = %8.2f ms / %5d runs (%8.2f ms per token, %8.2f tokens per second)\n", __func__, timings.t_eval_ms, timings.n_eval, timings.t_eval_ms / timings.n_eval, 1e3 / timings.t_eval_ms * timings.n_eval); - fprintf(stderr, "%s: total time = %8.2f ms\n", __func__, (timings.t_end_ms - timings.t_start_ms)); + LLAMA_LOG_INFO("%s: total time = %8.2f ms\n", __func__, (timings.t_end_ms - timings.t_start_ms)); } void llama_reset_timings(struct llama_context * ctx) { @@ -3672,3 +4290,44 @@ const char * llama_print_system_info(void) { const std::vector<std::pair<std::string, struct ggml_tensor *>>& llama_internal_get_tensor_map(struct llama_context * ctx) { return ctx->model.tensors_by_name; } + + +void llama_log_set(llama_log_callback log_callback, void * user_data) { + g_state.log_callback = log_callback ? log_callback : llama_log_callback_default; + g_state.log_callback_user_data = user_data; +} + +#if defined(_MSC_VER) && !defined(vsnprintf) +#define vsnprintf _vsnprintf +#endif + +static void llama_log_internal_v(llama_log_level level, const char * format, va_list args) { + va_list args_copy; + va_copy(args_copy, args); + char buffer[128]; + int len = vsnprintf(buffer, 128, format, args); + if (len < 128) { + g_state.log_callback(level, buffer, g_state.log_callback_user_data); + } else { + char* buffer2 = new char[len+1]; + vsnprintf(buffer2, len+1, format, args_copy); + buffer2[len] = 0; + g_state.log_callback(level, buffer2, g_state.log_callback_user_data); + delete[] buffer2; + } + va_end(args_copy); +} + +static void llama_log_internal(llama_log_level level, const char * format, ...) { + va_list args; + va_start(args, format); + llama_log_internal_v(level, format, args); + va_end(args); +} + +static void llama_log_callback_default(llama_log_level level, const char * text, void * user_data) { + (void) level; + (void) user_data; + fputs(text, stderr); + fflush(stderr); +} |