From f2c680f893cfb393654170f4c362fc72712acdb4 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Thu, 5 Dec 2024 14:30:33 +0200 Subject: [PATCH] talk-llama : sync llama.cpp --- examples/talk-llama/llama.cpp | 590 ++++++++++++++++++-------------- examples/talk-llama/llama.h | 9 +- examples/talk-llama/unicode.cpp | 11 + 3 files changed, 358 insertions(+), 252 deletions(-) diff --git a/examples/talk-llama/llama.cpp b/examples/talk-llama/llama.cpp index c51b36e6..00f78639 100644 --- a/examples/talk-llama/llama.cpp +++ b/examples/talk-llama/llama.cpp @@ -179,7 +179,7 @@ enum llm_arch { LLM_ARCH_COMMAND_R, LLM_ARCH_DBRX, LLM_ARCH_OLMO, - LLM_ARCH_OLMO_1124, + LLM_ARCH_OLMO2, LLM_ARCH_OLMOE, LLM_ARCH_OPENELM, LLM_ARCH_ARCTIC, @@ -233,7 +233,7 @@ static const std::map LLM_ARCH_NAMES = { { LLM_ARCH_COMMAND_R, "command-r" }, { LLM_ARCH_DBRX, "dbrx" }, { LLM_ARCH_OLMO, "olmo" }, - { LLM_ARCH_OLMO_1124, "olmo_1124" }, + { LLM_ARCH_OLMO2, "olmo2" }, { LLM_ARCH_OLMOE, "olmoe" }, { LLM_ARCH_OPENELM, "openelm" }, { LLM_ARCH_ARCTIC, "arctic" }, @@ -1036,6 +1036,8 @@ static const std::map> LLM_TENSOR_N { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, { LLM_TENSOR_OUTPUT, "output" }, { LLM_TENSOR_ROPE_FREQS, "rope_freqs" }, + { LLM_TENSOR_ROPE_FACTORS_LONG, "rope_factors_long" }, + { LLM_TENSOR_ROPE_FACTORS_SHORT, "rope_factors_short" }, { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" }, { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" }, @@ -1210,7 +1212,7 @@ static const std::map> LLM_TENSOR_N }, }, { - LLM_ARCH_OLMO_1124, + LLM_ARCH_OLMO2, { { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, @@ -1549,6 +1551,67 @@ static const std::map> LLM_TENSOR_N }, }; +enum llm_chat_template { + LLM_CHAT_TEMPLATE_CHATML, + LLM_CHAT_TEMPLATE_LLAMA_2, + LLM_CHAT_TEMPLATE_LLAMA_2_SYS, + LLM_CHAT_TEMPLATE_LLAMA_2_SYS_BOS, + LLM_CHAT_TEMPLATE_LLAMA_2_SYS_STRIP, + LLM_CHAT_TEMPLATE_MISTRAL_V1, + LLM_CHAT_TEMPLATE_MISTRAL_V3, + LLM_CHAT_TEMPLATE_MISTRAL_V3_TEKKEN, + LLM_CHAT_TEMPLATE_MISTRAL_V7, + LLM_CHAT_TEMPLATE_PHI_3, + LLM_CHAT_TEMPLATE_ZEPHYR, + LLM_CHAT_TEMPLATE_MONARCH, + LLM_CHAT_TEMPLATE_GEMMA, + LLM_CHAT_TEMPLATE_ORION, + LLM_CHAT_TEMPLATE_OPENCHAT, + LLM_CHAT_TEMPLATE_VICUNA, + LLM_CHAT_TEMPLATE_VICUNA_ORCA, + LLM_CHAT_TEMPLATE_DEEPSEEK, + LLM_CHAT_TEMPLATE_DEEPSEEK_2, + LLM_CHAT_TEMPLATE_COMMAND_R, + LLM_CHAT_TEMPLATE_LLAMA_3, + LLM_CHAT_TEMPLATE_CHATGML_3, + LLM_CHAT_TEMPLATE_CHATGML_4, + LLM_CHAT_TEMPLATE_MINICPM, + LLM_CHAT_TEMPLATE_EXAONE_3, + LLM_CHAT_TEMPLATE_RWKV_WORLD, + LLM_CHAT_TEMPLATE_GRANITE, + LLM_CHAT_TEMPLATE_UNKNOWN, +}; + +static const std::map LLM_CHAT_TEMPLATES = { + { "chatml", LLM_CHAT_TEMPLATE_CHATML }, + { "llama2", LLM_CHAT_TEMPLATE_LLAMA_2 }, + { "llama2-sys", LLM_CHAT_TEMPLATE_LLAMA_2_SYS }, + { "llama2-sys-bos", LLM_CHAT_TEMPLATE_LLAMA_2_SYS_BOS }, + { "llama2-sys-strip", LLM_CHAT_TEMPLATE_LLAMA_2_SYS_STRIP }, + { "mistral-v1", LLM_CHAT_TEMPLATE_MISTRAL_V1 }, + { "mistral-v3", LLM_CHAT_TEMPLATE_MISTRAL_V3 }, + { "mistral-v3-tekken", LLM_CHAT_TEMPLATE_MISTRAL_V3_TEKKEN }, + { "mistral-v7", LLM_CHAT_TEMPLATE_MISTRAL_V7 }, + { "phi3", LLM_CHAT_TEMPLATE_PHI_3 }, + { "zephyr", LLM_CHAT_TEMPLATE_ZEPHYR }, + { "monarch", LLM_CHAT_TEMPLATE_MONARCH }, + { "gemma", LLM_CHAT_TEMPLATE_GEMMA }, + { "orion", LLM_CHAT_TEMPLATE_ORION }, + { "openchat", LLM_CHAT_TEMPLATE_OPENCHAT }, + { "vicuna", LLM_CHAT_TEMPLATE_VICUNA }, + { "vicuna-orca", LLM_CHAT_TEMPLATE_VICUNA_ORCA }, + { "deepseek", LLM_CHAT_TEMPLATE_DEEPSEEK }, + { "deepseek2", LLM_CHAT_TEMPLATE_DEEPSEEK_2 }, + { "command-r", LLM_CHAT_TEMPLATE_COMMAND_R }, + { "llama3", LLM_CHAT_TEMPLATE_LLAMA_3 }, + { "chatglm3", LLM_CHAT_TEMPLATE_CHATGML_3 }, + { "chatglm4", LLM_CHAT_TEMPLATE_CHATGML_4 }, + { "minicpm", LLM_CHAT_TEMPLATE_MINICPM }, + { "exaone3", LLM_CHAT_TEMPLATE_EXAONE_3 }, + { "rwkv-world", LLM_CHAT_TEMPLATE_RWKV_WORLD }, + { "granite", LLM_CHAT_TEMPLATE_GRANITE }, +}; + static llm_arch llm_arch_from_string(const std::string & name) { for (const auto & kv : LLM_ARCH_NAMES) { // NOLINT if (kv.second == name) { @@ -1622,9 +1685,10 @@ struct LLM_TN { // static const std::map LLAMA_ROPE_SCALING_TYPES = { - { LLAMA_ROPE_SCALING_TYPE_NONE, "none" }, - { LLAMA_ROPE_SCALING_TYPE_LINEAR, "linear" }, - { LLAMA_ROPE_SCALING_TYPE_YARN, "yarn" }, + { LLAMA_ROPE_SCALING_TYPE_NONE, "none" }, + { LLAMA_ROPE_SCALING_TYPE_LINEAR, "linear" }, + { LLAMA_ROPE_SCALING_TYPE_YARN, "yarn" }, + { LLAMA_ROPE_SCALING_TYPE_LONGROPE, "longrope" }, }; static llama_rope_scaling_type llama_rope_scaling_type_from_string(const std::string & name) { @@ -2341,6 +2405,7 @@ enum e_model { MODEL_16B, MODEL_20B, MODEL_30B, + MODEL_32B, MODEL_34B, MODEL_35B, MODEL_40B, @@ -4866,7 +4931,9 @@ struct llama_model_loader { mappings.reserve(files.size()); mmaps_used.reserve(files.size()); for (const auto & file : files) { - std::unique_ptr mapping(new llama_mmap(file.get(), prefetch ? -1 : 0, ggml_is_numa())); + auto * reg = ggml_backend_dev_backend_reg(ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU)); + auto * is_numa_fn = (decltype(ggml_is_numa) *) ggml_backend_reg_get_proc_address(reg, "ggml_backend_cpu_is_numa"); + std::unique_ptr mapping(new llama_mmap(file.get(), prefetch ? -1 : 0, is_numa_fn())); mmaps_used.emplace_back(mapping->size, 0); if (mlock_mmaps) { std::unique_ptr mlock_mmap(new llama_mlock()); @@ -5328,6 +5395,7 @@ static const char * llama_model_type_name(e_model type) { case MODEL_16B: return "16B"; case MODEL_20B: return "20B"; case MODEL_30B: return "30B"; + case MODEL_32B: return "32B"; case MODEL_34B: return "34B"; case MODEL_35B: return "35B"; case MODEL_40B: return "40B"; @@ -5515,8 +5583,12 @@ static void llm_load_hparams( case LLM_ARCH_MINICPM: { ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); + ml.get_key(LLM_KV_EMBEDDING_SCALE, hparams.f_embedding_scale); + ml.get_key(LLM_KV_RESIDUAL_SCALE, hparams.f_residual_scale); + ml.get_key(LLM_KV_LOGIT_SCALE, hparams.f_logit_scale); switch (hparams.n_layer) { + case 52: model.type = e_model::MODEL_1B; break; case 40: model.type = e_model::MODEL_2B; break; default: model.type = e_model::MODEL_UNKNOWN; } @@ -5688,7 +5760,10 @@ static void llm_load_hparams( case 24: model.type = hparams.n_embd == 1024 ? e_model::MODEL_0_5B : e_model::MODEL_1B; break; case 28: model.type = hparams.n_embd == 1536 ? e_model::MODEL_1_5B : e_model::MODEL_7B; break; case 32: model.type = e_model::MODEL_7B; break; + case 36: model.type = e_model::MODEL_3B; break; case 40: model.type = hparams.n_head() == 20 ? e_model::MODEL_4B : e_model::MODEL_13B; break; + case 48: model.type = e_model::MODEL_14B; break; + case 64: model.type = e_model::MODEL_32B; break; case 80: model.type = e_model::MODEL_70B; break; default: model.type = e_model::MODEL_UNKNOWN; } @@ -5898,7 +5973,7 @@ static void llm_load_hparams( default: model.type = e_model::MODEL_UNKNOWN; } } break; - case LLM_ARCH_OLMO_1124: + case LLM_ARCH_OLMO2: { ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); @@ -6997,7 +7072,7 @@ static void llm_load_print_meta(llama_model_loader & ml, llama_model & model) { LLAMA_LOG_INFO("%s: n_ff_shexp = %d\n", __func__, hparams.n_ff_shexp); } - if (model.arch == LLM_ARCH_GRANITE || model.arch == LLM_ARCH_GRANITE_MOE) { + if (model.arch == LLM_ARCH_MINICPM || model.arch == LLM_ARCH_GRANITE || model.arch == LLM_ARCH_GRANITE_MOE) { LLAMA_LOG_INFO("%s: f_embedding_scale = %f\n", __func__, hparams.f_embedding_scale); LLAMA_LOG_INFO("%s: f_residual_scale = %f\n", __func__, hparams.f_residual_scale); LLAMA_LOG_INFO("%s: f_attention_scale = %f\n", __func__, hparams.f_attention_scale); @@ -7181,12 +7256,12 @@ static bool weight_buft_supported(const llama_hparams & hparams, ggml_tensor * w } break; case GGML_OP_ADD: { - ggml_tensor * a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, w->ne[0], 512); + ggml_tensor * a = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, w->ne[0], w->ne[1], w->ne[2], w->ne[3]); op_tensor = ggml_add(ctx, a, w); } break; case GGML_OP_MUL: { - ggml_tensor * a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, w->ne[0], 512); + ggml_tensor * a = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, w->ne[0], w->ne[1], w->ne[2], w->ne[3]); op_tensor = ggml_mul(ctx, a, w); } break; case GGML_OP_DIV: @@ -7622,7 +7697,13 @@ static bool llm_load_tensors( layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0); - layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot/2}, llama_model_loader::TENSOR_NOT_REQUIRED | (i != 0 ? llama_model_loader::TENSOR_DUPLICATED : 0)); + if (hparams.rope_scaling_type_train == LLAMA_ROPE_SCALING_TYPE_LONGROPE) { + layer.rope_long = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_LONG, "weight", i), {n_rot/2}, llama_model_loader::TENSOR_NOT_REQUIRED | (i != 0 ? llama_model_loader::TENSOR_DUPLICATED : 0)); + layer.rope_short = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_SHORT, "weight", i), {n_rot/2}, llama_model_loader::TENSOR_NOT_REQUIRED | (i != 0 ? llama_model_loader::TENSOR_DUPLICATED : 0)); + } + else { + layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot/2}, llama_model_loader::TENSOR_NOT_REQUIRED | (i != 0 ? llama_model_loader::TENSOR_DUPLICATED : 0)); + } if (n_expert == 0) { layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0); @@ -8591,7 +8672,7 @@ static bool llm_load_tensors( layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, 0); } } break; - case LLM_ARCH_OLMO_1124: + case LLM_ARCH_OLMO2: { model.tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0); @@ -9190,7 +9271,7 @@ static bool llm_load_tensors( ggml_backend_dev_t dev = ggml_backend_buft_get_device(buft); if (!dev) { // FIXME: workaround for CPU backend buft having a NULL device - dev = ggml_backend_reg_dev_get(ggml_backend_cpu_reg(), 0); + dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU); } ggml_backend_dev_props props; ggml_backend_dev_get_props(dev, &props); @@ -13429,153 +13510,6 @@ struct llm_build_context { return gf; } - // ref: https://arxiv.org/abs/2203.03466 - // https://github.com/ggerganov/llama.cpp/issues/5276#issuecomment-1925774738 - // based on the original build_llama() function - struct ggml_cgraph * build_minicpm() { - struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), 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); - - const int64_t n_embd = hparams.n_embd; - //TODO: if the model varies, these parameters need to be read from the model - const int64_t n_embd_base = 256; - const float scale_embd = 12.0f; - const float scale_depth = 1.4f; - - struct ggml_tensor * cur; - struct ggml_tensor * inpL; - - inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb); - - // scale the input embeddings - inpL = ggml_scale(ctx0, inpL, scale_embd); - cb(inpL, "inp_scaled", -1); - - // inp_pos - contains the positions - struct ggml_tensor * inp_pos = build_inp_pos(); - - // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = build_inp_KQ_mask(); - - 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, NULL, - LLM_NORM_RMS, cb, il); - cb(cur, "attn_norm", il); - - // self-attention - { - // compute Q and K and RoPE them - struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, 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 = llm_build_lora_mm(lctx, 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 = llm_build_lora_mm(lctx, 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_ext( - ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - cb(Qcur, "Qcur", il); - - Kcur = ggml_rope_ext( - ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - cb(Kcur, "Kcur", il); - - cur = llm_build_kv(ctx0, lctx, kv_self, gf, - model.layers[il].wo, model.layers[il].bo, - Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); - } - - if (il == n_layer - 1) { - // skip computing output for unused tokens - struct ggml_tensor * inp_out_ids = build_inp_out_ids(); - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - // scale_res - scale the hidden states for residual connection - const float scale_res = scale_depth/sqrtf(float(n_layer)); - cur = ggml_scale(ctx0, cur, scale_res); - cb(cur, "hidden_scaled", -1); - - 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, NULL, - LLM_NORM_RMS, cb, il); - cb(cur, "ffn_norm", il); - - cur = llm_build_ffn(ctx0, lctx, cur, - model.layers[il].ffn_up, NULL, NULL, - model.layers[il].ffn_gate, NULL, NULL, - model.layers[il].ffn_down, NULL, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, cb, il); - cb(cur, "ffn_out", il); - } - - // scale the hidden states for residual connection - cur = ggml_scale(ctx0, cur, scale_res); - cb(cur, "hidden_scaled_ffn", -1); - - cur = ggml_add(ctx0, cur, ffn_inp); - cur = lctx.cvec.apply_to(ctx0, cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = llm_build_norm(ctx0, cur, hparams, - model.output_norm, NULL, - LLM_NORM_RMS, cb, -1); - cb(cur, "result_norm", -1); - - // lm_head scaling - const float scale_lmhead = float(n_embd_base)/float(n_embd); - cur = ggml_scale(ctx0, cur, scale_lmhead); - cb(cur, "lmhead_scaling", -1); - - // lm_head - cur = llm_build_lora_mm(lctx, ctx0, model.output, cur); - cb(cur, "result_output", -1); - - ggml_build_forward_expand(gf, cur); - - return gf; - } - struct ggml_cgraph * build_minicpm3() { struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false); @@ -14481,7 +14415,7 @@ struct llm_build_context { return gf; } - struct ggml_cgraph * build_olmo_1124() { + struct ggml_cgraph * build_olmo2() { struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false); // mutable variable, needed during the last layer of the computation to skip unused tokens @@ -16674,6 +16608,7 @@ static struct ggml_cgraph * llama_build_graph( switch (model.arch) { case LLM_ARCH_LLAMA: + case LLM_ARCH_MINICPM: case LLM_ARCH_GRANITE: case LLM_ARCH_GRANITE_MOE: { @@ -16757,10 +16692,6 @@ static struct ggml_cgraph * llama_build_graph( { result = llm.build_internlm2(); } break; - case LLM_ARCH_MINICPM: - { - result = llm.build_minicpm(); - } break; case LLM_ARCH_MINICPM3: { result = llm.build_minicpm3(); @@ -16797,9 +16728,9 @@ static struct ggml_cgraph * llama_build_graph( { result = llm.build_olmo(); } break; - case LLM_ARCH_OLMO_1124: + case LLM_ARCH_OLMO2: { - result = llm.build_olmo_1124(); + result = llm.build_olmo2(); } break; case LLM_ARCH_OLMOE: { @@ -17443,8 +17374,9 @@ static enum ggml_status llama_graph_compute( int n_threads, ggml_threadpool * threadpool) { if (lctx.backend_cpu != nullptr) { - ggml_backend_cpu_set_threadpool(lctx.backend_cpu, threadpool); - ggml_backend_cpu_set_abort_callback(lctx.backend_cpu, lctx.abort_callback, lctx.abort_callback_data); + auto * reg = ggml_backend_dev_backend_reg(ggml_backend_get_device(lctx.backend_cpu)); + auto * set_threadpool_fn = (decltype(ggml_backend_cpu_set_threadpool) *) ggml_backend_reg_get_proc_address(reg, "ggml_backend_cpu_set_threadpool"); + set_threadpool_fn(lctx.backend_cpu, threadpool); } // set the number of threads for all the backends @@ -18211,13 +18143,13 @@ static void llama_kv_cache_defrag_internal(struct llama_context & lctx) { static void llama_kv_cache_update_internal(struct llama_context & lctx) { bool need_reserve = false; - // apply K-shift if needed - if (lctx.model.hparams.rope_type != LLAMA_ROPE_TYPE_NONE && lctx.kv_self.has_shift) { + if (lctx.kv_self.has_shift) { if (!llama_kv_cache_can_shift(&lctx)) { - GGML_ABORT("Deepseek2 does not support K-shift"); + GGML_ABORT("The current context does not support K-shift"); } - { + // apply K-shift if needed + if (lctx.model.hparams.rope_type != LLAMA_ROPE_TYPE_NONE) { ggml_backend_sched_reset(lctx.sched.get()); ggml_cgraph * gf = llama_build_graph_k_shift(lctx); @@ -19361,6 +19293,7 @@ void llama_lora_adapter_free(struct llama_lora_adapter * adapter) { // struct llama_model_params llama_model_default_params() { struct llama_model_params result = { + /*.devices =*/ nullptr, /*.n_gpu_layers =*/ 0, /*.split_mode =*/ LLAMA_SPLIT_MODE_LAYER, /*.main_gpu =*/ 0, @@ -19478,7 +19411,11 @@ void llama_backend_init(void) { void llama_numa_init(enum ggml_numa_strategy numa) { if (numa != GGML_NUMA_STRATEGY_DISABLED) { - ggml_numa_init(numa); + auto * dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU); + GGML_ASSERT(dev && "CPU backend is not loaded"); + auto * reg = ggml_backend_dev_backend_reg(dev); + auto * numa_init_fn = (decltype(ggml_numa_init) *) ggml_backend_reg_get_proc_address(reg, "ggml_backend_cpu_numa_init"); + numa_init_fn(numa); } } @@ -19569,19 +19506,24 @@ struct llama_model * llama_load_model_from_file( } // create list of devices to use with this model - // currently, we use all available devices - // TODO: rework API to give user more control over device selection - for (size_t i = 0; i < ggml_backend_dev_count(); ++i) { - ggml_backend_dev_t dev = ggml_backend_dev_get(i); - switch (ggml_backend_dev_type(dev)) { - case GGML_BACKEND_DEVICE_TYPE_CPU: - case GGML_BACKEND_DEVICE_TYPE_ACCEL: - // skip CPU backends since they are handled separately - break; + if (params.devices) { + for (ggml_backend_dev_t * dev = params.devices; *dev; ++dev) { + model->devices.push_back(*dev); + } + } else { + // use all available devices + for (size_t i = 0; i < ggml_backend_dev_count(); ++i) { + ggml_backend_dev_t dev = ggml_backend_dev_get(i); + switch (ggml_backend_dev_type(dev)) { + case GGML_BACKEND_DEVICE_TYPE_CPU: + case GGML_BACKEND_DEVICE_TYPE_ACCEL: + // skip CPU backends since they are handled separately + break; - case GGML_BACKEND_DEVICE_TYPE_GPU: - model->devices.push_back(dev); - break; + case GGML_BACKEND_DEVICE_TYPE_GPU: + model->devices.push_back(dev); + break; + } } } @@ -19752,9 +19694,6 @@ struct llama_context * llama_new_context_with_model( __func__, n_ctx_per_seq, hparams.n_ctx_train); } - ctx->abort_callback = params.abort_callback; - ctx->abort_callback_data = params.abort_callback_data; - ctx->logits_all = params.logits_all; // build worst-case graph for encoder if a model contains encoder @@ -19803,7 +19742,7 @@ struct llama_context * llama_new_context_with_model( } // add CPU backend - ctx->backend_cpu = ggml_backend_cpu_init(); + ctx->backend_cpu = ggml_backend_init_by_type(GGML_BACKEND_DEVICE_TYPE_CPU, nullptr); if (ctx->backend_cpu == nullptr) { LLAMA_LOG_ERROR("%s: failed to initialize CPU backend\n", __func__); llama_free(ctx); @@ -19823,6 +19762,8 @@ struct llama_context * llama_new_context_with_model( } } + llama_set_abort_callback(ctx, params.abort_callback, params.abort_callback_data); + if (!llama_kv_cache_init(ctx->kv_self, ctx, type_k, type_v, kv_size, cparams.offload_kqv)) { LLAMA_LOG_ERROR("%s: llama_kv_cache_init() failed for self-attention cache\n", __func__); llama_free(ctx); @@ -19868,7 +19809,8 @@ struct llama_context * llama_new_context_with_model( std::vector backend_ptrs; for (auto & backend : ctx->backends) { auto * buft = ggml_backend_get_default_buffer_type(backend.get()); - if (ggml_backend_is_cpu(backend.get()) && !model->devices.empty()) { + auto backend_type = ggml_backend_dev_type(ggml_backend_get_device(backend.get())); + if (backend_type == GGML_BACKEND_DEVICE_TYPE_CPU && !model->devices.empty()) { // use the host buffer of the first device CPU for faster transfer of the intermediate state auto * dev = model->devices[0]; auto * host_buft = ggml_backend_dev_host_buffer_type(dev); @@ -19896,7 +19838,8 @@ struct llama_context * llama_new_context_with_model( // pipeline parallelism requires support for async compute and events in all devices if (pipeline_parallel) { for (auto & backend : ctx->backends) { - if (ggml_backend_is_cpu(backend.get())) { + auto dev_type = ggml_backend_dev_type(ggml_backend_get_device(backend.get())); + if (dev_type == GGML_BACKEND_DEVICE_TYPE_CPU) { // ignore CPU backend continue; } @@ -20070,7 +20013,7 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) { case LLM_ARCH_QWEN: case LLM_ARCH_QWEN2: case LLM_ARCH_QWEN2MOE: - case LLM_ARCH_OLMO_1124: + case LLM_ARCH_OLMO2: case LLM_ARCH_OLMOE: case LLM_ARCH_PHI2: case LLM_ARCH_PHI3: @@ -20463,7 +20406,7 @@ void llama_kv_cache_update(struct llama_context * ctx) { } bool llama_kv_cache_can_shift(struct llama_context * ctx) { - return ctx->model.arch != LLM_ARCH_DEEPSEEK2; // not supported due to MLA + return !ctx->kv_self.recurrent && ctx->model.arch != LLM_ARCH_DEEPSEEK2; // not supported due to MLA } // deprecated @@ -21450,6 +21393,14 @@ int32_t llama_n_threads_batch(struct llama_context * ctx) { void llama_set_abort_callback(struct llama_context * ctx, bool (*abort_callback)(void * data), void * abort_callback_data) { ctx->abort_callback = abort_callback; ctx->abort_callback_data = abort_callback_data; + + for (auto & backend : ctx->backends) { + auto * reg = ggml_backend_dev_backend_reg(ggml_backend_get_device(backend.get())); + auto * set_abort_callback_fn = (ggml_backend_set_abort_callback_t) ggml_backend_reg_get_proc_address(reg, "ggml_backend_set_abort_callback"); + if (set_abort_callback_fn) { + set_abort_callback_fn(backend.get(), ctx->abort_callback, ctx->abort_callback_data); + } + } } void llama_set_embeddings(struct llama_context * ctx, bool embeddings) { @@ -21816,18 +21767,109 @@ int32_t llama_detokenize( // chat templates // +static llm_chat_template llama_chat_detect_template(const std::string & tmpl) { + if (LLM_CHAT_TEMPLATES.find(tmpl) != LLM_CHAT_TEMPLATES.end()) { + return LLM_CHAT_TEMPLATES.at(tmpl); + } + auto tmpl_contains = [&tmpl](const char * haystack) -> bool { + return tmpl.find(haystack) != std::string::npos; + }; + if (tmpl_contains("<|im_start|>")) { + return LLM_CHAT_TEMPLATE_CHATML; + } else if (tmpl.find("mistral") == 0 || tmpl_contains("[INST]")) { + if (tmpl_contains("[SYSTEM_PROMPT]")) { + return LLM_CHAT_TEMPLATE_MISTRAL_V7; + } else if ( + // catches official 'v1' template + tmpl_contains("' [INST] ' + system_message") + // catches official 'v3' and 'v3-tekken' templates + || tmpl_contains("[AVAILABLE_TOOLS]") + ) { + // Official mistral 'v1', 'v3' and 'v3-tekken' templates + // See: https://github.com/mistralai/cookbook/blob/main/concept-deep-dive/tokenization/chat_templates.md + // See: https://github.com/mistralai/cookbook/blob/main/concept-deep-dive/tokenization/templates.md + if (tmpl_contains(" [INST]")) { + return LLM_CHAT_TEMPLATE_MISTRAL_V1; + } else if (tmpl_contains("\"[INST]\"")) { + return LLM_CHAT_TEMPLATE_MISTRAL_V3_TEKKEN; + } + return LLM_CHAT_TEMPLATE_MISTRAL_V3; + } else { + // llama2 template and its variants + // [variant] support system message + // See: https://huggingface.co/blog/llama2#how-to-prompt-llama-2 + bool support_system_message = tmpl_contains("<>"); + bool add_bos_inside_history = tmpl_contains("bos_token + '[INST]"); + bool strip_message = tmpl_contains("content.strip()"); + if (strip_message) { + return LLM_CHAT_TEMPLATE_LLAMA_2_SYS_STRIP; + } else if (add_bos_inside_history) { + return LLM_CHAT_TEMPLATE_LLAMA_2_SYS_BOS; + } else if (support_system_message) { + return LLM_CHAT_TEMPLATE_LLAMA_2_SYS; + } else { + return LLM_CHAT_TEMPLATE_LLAMA_2; + } + } + } else if (tmpl_contains("<|assistant|>") && tmpl_contains("<|end|>")) { + return LLM_CHAT_TEMPLATE_PHI_3; + } else if (tmpl_contains("<|user|>") && tmpl_contains("<|endoftext|>")) { + return LLM_CHAT_TEMPLATE_ZEPHYR; + } else if (tmpl_contains("bos_token + message['role']")) { + return LLM_CHAT_TEMPLATE_MONARCH; + } else if (tmpl_contains("")) { + return LLM_CHAT_TEMPLATE_GEMMA; + } else if (tmpl_contains("'\\n\\nAssistant: ' + eos_token")) { + // OrionStarAI/Orion-14B-Chat + return LLM_CHAT_TEMPLATE_ORION; + } else if (tmpl_contains("GPT4 Correct ")) { + // openchat/openchat-3.5-0106 + return LLM_CHAT_TEMPLATE_OPENCHAT; + } else if (tmpl_contains("USER: ") && tmpl_contains("ASSISTANT: ")) { + // eachadea/vicuna-13b-1.1 (and Orca variant) + if (tmpl_contains("SYSTEM: ")) { + return LLM_CHAT_TEMPLATE_VICUNA_ORCA; + } + return LLM_CHAT_TEMPLATE_VICUNA; + } else if (tmpl_contains("### Instruction:") && tmpl_contains("<|EOT|>")) { + // deepseek-ai/deepseek-coder-33b-instruct + return LLM_CHAT_TEMPLATE_DEEPSEEK; + } else if (tmpl_contains("<|START_OF_TURN_TOKEN|>") && tmpl_contains("<|USER_TOKEN|>")) { + // CohereForAI/c4ai-command-r-plus + return LLM_CHAT_TEMPLATE_COMMAND_R; + } else if (tmpl_contains("<|start_header_id|>") && tmpl_contains("<|end_header_id|>")) { + return LLM_CHAT_TEMPLATE_LLAMA_3; + } else if (tmpl_contains("[gMASK]sop")) { + // chatglm3-6b + return LLM_CHAT_TEMPLATE_CHATGML_3; + } else if (tmpl_contains("[gMASK]")) { + return LLM_CHAT_TEMPLATE_CHATGML_4; + } else if (tmpl_contains(LU8("<用户>"))) { + // MiniCPM-3B-OpenHermes-2.5-v2-GGUF + return LLM_CHAT_TEMPLATE_MINICPM; + } else if (tmpl_contains("'Assistant: ' + message['content'] + eos_token")) { + return LLM_CHAT_TEMPLATE_DEEPSEEK_2; + } else if (tmpl_contains("[|system|]") && tmpl_contains("[|assistant|]") && tmpl_contains("[|endofturn|]")) { + // ref: https://huggingface.co/LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct/discussions/8#66bae61b1893d14ee8ed85bb + // EXAONE-3.0-7.8B-Instruct + return LLM_CHAT_TEMPLATE_EXAONE_3; + } else if (tmpl_contains("rwkv-world")) { + return LLM_CHAT_TEMPLATE_RWKV_WORLD; + } else if (tmpl_contains("<|start_of_role|>")) { + return LLM_CHAT_TEMPLATE_GRANITE; + } + return LLM_CHAT_TEMPLATE_UNKNOWN; +} + // Simple version of "llama_apply_chat_template" that only works with strings // This function uses heuristic checks to determine commonly used template. It is not a jinja parser. static int32_t llama_chat_apply_template_internal( - const std::string & tmpl, + const llm_chat_template tmpl, const std::vector & chat, std::string & dest, bool add_ass) { // Taken from the research: https://github.com/ggerganov/llama.cpp/issues/5527 std::stringstream ss; - auto tmpl_contains = [&tmpl](std::string haystack) -> bool { - return tmpl.find(haystack) != std::string::npos; - }; - if (tmpl == "chatml" || tmpl_contains("<|im_start|>")) { + if (tmpl == LLM_CHAT_TEMPLATE_CHATML) { // chatml template for (auto message : chat) { ss << "<|im_start|>" << message->role << "\n" << message->content << "<|im_end|>\n"; @@ -21835,16 +21877,59 @@ static int32_t llama_chat_apply_template_internal( if (add_ass) { ss << "<|im_start|>assistant\n"; } - } else if (tmpl == "llama2" || tmpl == "mistral" || tmpl_contains("[INST]")) { + } else if (tmpl == LLM_CHAT_TEMPLATE_MISTRAL_V7) { + // Official mistral 'v7' template + // See: https://huggingface.co/mistralai/Mistral-Large-Instruct-2411#basic-instruct-template-v7 + for (auto message : chat) { + std::string role(message->role); + std::string content(message->content); + if (role == "system") { + ss << "[SYSTEM_PROMPT] " << content << "[/SYSTEM_PROMPT]"; + } else if (role == "user") { + ss << "[INST] " << content << "[/INST]"; + } + else { + ss << " " << content << ""; + } + } + } else if (tmpl == LLM_CHAT_TEMPLATE_MISTRAL_V1 + || tmpl == LLM_CHAT_TEMPLATE_MISTRAL_V3 + || tmpl == LLM_CHAT_TEMPLATE_MISTRAL_V3_TEKKEN) { + // See: https://github.com/mistralai/cookbook/blob/main/concept-deep-dive/tokenization/chat_templates.md + // See: https://github.com/mistralai/cookbook/blob/main/concept-deep-dive/tokenization/templates.md + std::string leading_space = tmpl == LLM_CHAT_TEMPLATE_MISTRAL_V1 ? " " : ""; + std::string trailing_space = tmpl == LLM_CHAT_TEMPLATE_MISTRAL_V3_TEKKEN ? "" : " "; + bool trim_assistant_message = tmpl == LLM_CHAT_TEMPLATE_MISTRAL_V3; + bool is_inside_turn = false; + for (auto message : chat) { + if (!is_inside_turn) { + ss << leading_space << "[INST]" << trailing_space; + is_inside_turn = true; + } + std::string role(message->role); + std::string content(message->content); + if (role == "system") { + ss << content << "\n\n"; + } else if (role == "user") { + ss << content << leading_space << "[/INST]"; + } else { + ss << trailing_space << (trim_assistant_message ? trim(content) : content) << ""; + is_inside_turn = false; + } + } + } else if ( + tmpl == LLM_CHAT_TEMPLATE_LLAMA_2 + || tmpl == LLM_CHAT_TEMPLATE_LLAMA_2_SYS + || tmpl == LLM_CHAT_TEMPLATE_LLAMA_2_SYS_BOS + || tmpl == LLM_CHAT_TEMPLATE_LLAMA_2_SYS_STRIP) { // llama2 template and its variants // [variant] support system message - bool support_system_message = tmpl_contains("<>") || tmpl == "mistral"; - // [variant] space before + after response - bool space_around_response = tmpl_contains("' ' + eos_token"); + // See: https://huggingface.co/blog/llama2#how-to-prompt-llama-2 + bool support_system_message = tmpl != LLM_CHAT_TEMPLATE_LLAMA_2; // [variant] add BOS inside history - bool add_bos_inside_history = tmpl_contains("bos_token + '[INST]"); + bool add_bos_inside_history = tmpl == LLM_CHAT_TEMPLATE_LLAMA_2_SYS_BOS; // [variant] trim spaces from the input message - bool strip_message = tmpl_contains("content.strip()"); + bool strip_message = tmpl == LLM_CHAT_TEMPLATE_LLAMA_2_SYS_STRIP; // construct the prompt bool is_inside_turn = true; // skip BOS at the beginning ss << "[INST] "; @@ -21865,12 +21950,11 @@ static int32_t llama_chat_apply_template_internal( } else if (role == "user") { ss << content << " [/INST]"; } else { - ss << (space_around_response ? " " : "") << content << (space_around_response ? " " : "") << ""; + ss << content << ""; is_inside_turn = false; } } - // llama2 templates seem to not care about "add_generation_prompt" - } else if (tmpl == "phi3" || (tmpl_contains("<|assistant|>") && tmpl_contains("<|end|>"))) { + } else if (tmpl == LLM_CHAT_TEMPLATE_PHI_3) { // Phi 3 for (auto message : chat) { std::string role(message->role); @@ -21879,7 +21963,7 @@ static int32_t llama_chat_apply_template_internal( if (add_ass) { ss << "<|assistant|>\n"; } - } else if (tmpl == "zephyr" || tmpl_contains("<|user|>")) { + } else if (tmpl == LLM_CHAT_TEMPLATE_ZEPHYR) { // zephyr template for (auto message : chat) { ss << "<|" << message->role << "|>" << "\n" << message->content << "<|endoftext|>\n"; @@ -21887,7 +21971,7 @@ static int32_t llama_chat_apply_template_internal( if (add_ass) { ss << "<|assistant|>\n"; } - } else if (tmpl == "monarch" || tmpl_contains("bos_token + message['role']")) { + } else if (tmpl == LLM_CHAT_TEMPLATE_MONARCH) { // mlabonne/AlphaMonarch-7B template (the is included inside history) for (auto message : chat) { std::string bos = (message == chat.front()) ? "" : ""; // skip BOS for first message @@ -21896,7 +21980,7 @@ static int32_t llama_chat_apply_template_internal( if (add_ass) { ss << "assistant\n"; } - } else if (tmpl == "gemma" || tmpl == "gemma2" || tmpl_contains("")) { + } else if (tmpl == LLM_CHAT_TEMPLATE_GEMMA) { // google/gemma-7b-it std::string system_prompt = ""; for (auto message : chat) { @@ -21918,7 +22002,7 @@ static int32_t llama_chat_apply_template_internal( if (add_ass) { ss << "model\n"; } - } else if (tmpl == "orion" || tmpl_contains("'\\n\\nAssistant: ' + eos_token")) { + } else if (tmpl == LLM_CHAT_TEMPLATE_ORION) { // OrionStarAI/Orion-14B-Chat std::string system_prompt = ""; for (auto message : chat) { @@ -21938,7 +22022,7 @@ static int32_t llama_chat_apply_template_internal( ss << message->content << ""; } } - } else if (tmpl == "openchat" || tmpl_contains("GPT4 Correct ")) { + } else if (tmpl == LLM_CHAT_TEMPLATE_OPENCHAT) { // openchat/openchat-3.5-0106, for (auto message : chat) { std::string role(message->role); @@ -21952,13 +22036,13 @@ static int32_t llama_chat_apply_template_internal( if (add_ass) { ss << "GPT4 Correct Assistant:"; } - } else if (tmpl == "vicuna" || tmpl == "vicuna-orca" || (tmpl_contains("USER: ") && tmpl_contains("ASSISTANT: "))) { + } else if (tmpl == LLM_CHAT_TEMPLATE_VICUNA || tmpl == LLM_CHAT_TEMPLATE_VICUNA_ORCA) { // eachadea/vicuna-13b-1.1 (and Orca variant) for (auto message : chat) { std::string role(message->role); if (role == "system") { // Orca-Vicuna variant uses a system prefix - if (tmpl == "vicuna-orca" || tmpl_contains("SYSTEM: ")) { + if (tmpl == LLM_CHAT_TEMPLATE_VICUNA_ORCA) { ss << "SYSTEM: " << message->content << "\n"; } else { ss << message->content << "\n\n"; @@ -21972,7 +22056,7 @@ static int32_t llama_chat_apply_template_internal( if (add_ass) { ss << "ASSISTANT:"; } - } else if (tmpl == "deepseek" || (tmpl_contains("### Instruction:") && tmpl_contains("<|EOT|>"))) { + } else if (tmpl == LLM_CHAT_TEMPLATE_DEEPSEEK) { // deepseek-ai/deepseek-coder-33b-instruct for (auto message : chat) { std::string role(message->role); @@ -21987,7 +22071,7 @@ static int32_t llama_chat_apply_template_internal( if (add_ass) { ss << "### Response:\n"; } - } else if (tmpl == "command-r" || (tmpl_contains("<|START_OF_TURN_TOKEN|>") && tmpl_contains("<|USER_TOKEN|>"))) { + } else if (tmpl == LLM_CHAT_TEMPLATE_COMMAND_R) { // CohereForAI/c4ai-command-r-plus for (auto message : chat) { std::string role(message->role); @@ -22002,7 +22086,7 @@ static int32_t llama_chat_apply_template_internal( if (add_ass) { ss << "<|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>"; } - } else if (tmpl == "llama3" || (tmpl_contains("<|start_header_id|>") && tmpl_contains("<|end_header_id|>"))) { + } else if (tmpl == LLM_CHAT_TEMPLATE_LLAMA_3) { // Llama 3 for (auto message : chat) { std::string role(message->role); @@ -22011,7 +22095,7 @@ static int32_t llama_chat_apply_template_internal( if (add_ass) { ss << "<|start_header_id|>assistant<|end_header_id|>\n\n"; } - } else if (tmpl == "chatglm3" || tmpl_contains("[gMASK]sop")) { + } else if (tmpl == LLM_CHAT_TEMPLATE_CHATGML_3) { // chatglm3-6b ss << "[gMASK]" << "sop"; for (auto message : chat) { @@ -22021,7 +22105,7 @@ static int32_t llama_chat_apply_template_internal( if (add_ass) { ss << "<|assistant|>"; } - } else if (tmpl == "chatglm4" || tmpl_contains("[gMASK]")) { + } else if (tmpl == LLM_CHAT_TEMPLATE_CHATGML_4) { ss << "[gMASK]" << ""; for (auto message : chat) { std::string role(message->role); @@ -22030,7 +22114,7 @@ static int32_t llama_chat_apply_template_internal( if (add_ass) { ss << "<|assistant|>"; } - } else if (tmpl == "minicpm" || tmpl_contains(LU8("<用户>"))) { + } else if (tmpl == LLM_CHAT_TEMPLATE_MINICPM) { // MiniCPM-3B-OpenHermes-2.5-v2-GGUF for (auto message : chat) { std::string role(message->role); @@ -22042,7 +22126,7 @@ static int32_t llama_chat_apply_template_internal( ss << trim(message->content); } } - } else if (tmpl == "deepseek2" || tmpl_contains("'Assistant: ' + message['content'] + eos_token")) { + } else if (tmpl == LLM_CHAT_TEMPLATE_DEEPSEEK_2) { // DeepSeek-V2 for (auto message : chat) { std::string role(message->role); @@ -22057,7 +22141,7 @@ static int32_t llama_chat_apply_template_internal( if (add_ass) { ss << "Assistant:"; } - } else if (tmpl == "exaone3" || (tmpl_contains("[|system|]") && tmpl_contains("[|assistant|]") && tmpl_contains("[|endofturn|]"))) { + } else if (tmpl == LLM_CHAT_TEMPLATE_EXAONE_3) { // ref: https://huggingface.co/LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct/discussions/8#66bae61b1893d14ee8ed85bb // EXAONE-3.0-7.8B-Instruct for (auto message : chat) { @@ -22073,7 +22157,7 @@ static int32_t llama_chat_apply_template_internal( if (add_ass) { ss << "[|assistant|]"; } - } else if (tmpl == "rwkv-world" || tmpl_contains("rwkv-world")) { + } else if (tmpl == LLM_CHAT_TEMPLATE_RWKV_WORLD) { // this template requires the model to have "\n\n" as EOT token for (auto message : chat) { std::string role(message->role); @@ -22083,7 +22167,7 @@ static int32_t llama_chat_apply_template_internal( ss << message->content << "\n\n"; } } - } else if (tmpl == "granite" || tmpl_contains("<|start_of_role|>")) { + } else if (tmpl == LLM_CHAT_TEMPLATE_GRANITE) { // IBM Granite template for (const auto & message : chat) { std::string role(message->role); @@ -22135,7 +22219,11 @@ int32_t llama_chat_apply_template( } std::string formatted_chat; - int32_t res = llama_chat_apply_template_internal(curr_tmpl, chat_vec, formatted_chat, add_ass); + llm_chat_template detected_tmpl = llama_chat_detect_template(curr_tmpl); + if (detected_tmpl == LLM_CHAT_TEMPLATE_UNKNOWN) { + return -1; + } + int32_t res = llama_chat_apply_template_internal(detected_tmpl, chat_vec, formatted_chat, add_ass); if (res < 0) { return res; } @@ -22145,6 +22233,15 @@ int32_t llama_chat_apply_template( return res; } +int32_t llama_chat_builtin_templates(const char ** output, size_t len) { + auto it = LLM_CHAT_TEMPLATES.begin(); + for (size_t i = 0; i < std::min(len, LLM_CHAT_TEMPLATES.size()); i++) { + output[i] = it->first.c_str(); + std::advance(it, 1); + } + return (int32_t) LLM_CHAT_TEMPLATES.size(); +} + // // sampling // @@ -22191,32 +22288,23 @@ int llama_split_prefix(char * dest, size_t maxlen, const char * split_path, int } const char * llama_print_system_info(void) { - ggml_cpu_init(); // some ARM features are detected at runtime - static std::string s; - s = ""; - s += "AVX = " + std::to_string(ggml_cpu_has_avx()) + " | "; - s += "AVX_VNNI = " + std::to_string(ggml_cpu_has_avx_vnni()) + " | "; - s += "AVX2 = " + std::to_string(ggml_cpu_has_avx2()) + " | "; - s += "AVX512 = " + std::to_string(ggml_cpu_has_avx512()) + " | "; - s += "AVX512_VBMI = " + std::to_string(ggml_cpu_has_avx512_vbmi()) + " | "; - s += "AVX512_VNNI = " + std::to_string(ggml_cpu_has_avx512_vnni()) + " | "; - s += "AVX512_BF16 = " + std::to_string(ggml_cpu_has_avx512_bf16()) + " | "; - s += "AMX_INT8 = " + std::to_string(ggml_cpu_has_amx_int8()) + " | "; - s += "FMA = " + std::to_string(ggml_cpu_has_fma()) + " | "; - s += "NEON = " + std::to_string(ggml_cpu_has_neon()) + " | "; - s += "SVE = " + std::to_string(ggml_cpu_has_sve()) + " | "; - s += "ARM_FMA = " + std::to_string(ggml_cpu_has_arm_fma()) + " | "; - s += "F16C = " + std::to_string(ggml_cpu_has_f16c()) + " | "; - s += "FP16_VA = " + std::to_string(ggml_cpu_has_fp16_va()) + " | "; - s += "RISCV_VECT = " + std::to_string(ggml_cpu_has_riscv_v()) + " | "; - s += "WASM_SIMD = " + std::to_string(ggml_cpu_has_wasm_simd()) + " | "; - s += "SSE3 = " + std::to_string(ggml_cpu_has_sse3()) + " | "; - s += "SSSE3 = " + std::to_string(ggml_cpu_has_ssse3()) + " | "; - s += "VSX = " + std::to_string(ggml_cpu_has_vsx()) + " | "; - s += "MATMUL_INT8 = " + std::to_string(ggml_cpu_has_matmul_int8()) + " | "; - s += "LLAMAFILE = " + std::to_string(ggml_cpu_has_llamafile()) + " | "; + for (size_t i = 0; i < ggml_backend_reg_count(); i++) { + auto * reg = ggml_backend_reg_get(i); + auto * get_features_fn = (ggml_backend_get_features_t) ggml_backend_reg_get_proc_address(reg, "ggml_backend_get_features"); + if (get_features_fn) { + ggml_backend_feature * features = get_features_fn(reg); + s += ggml_backend_reg_name(reg); + s += " : "; + for (; features->name; features++) { + s += features->name; + s += " = "; + s += features->value; + s += " | "; + } + } + } return s.c_str(); } diff --git a/examples/talk-llama/llama.h b/examples/talk-llama/llama.h index 90791d5f..168c3fa1 100644 --- a/examples/talk-llama/llama.h +++ b/examples/talk-llama/llama.h @@ -185,7 +185,8 @@ extern "C" { LLAMA_ROPE_SCALING_TYPE_NONE = 0, LLAMA_ROPE_SCALING_TYPE_LINEAR = 1, LLAMA_ROPE_SCALING_TYPE_YARN = 2, - LLAMA_ROPE_SCALING_TYPE_MAX_VALUE = LLAMA_ROPE_SCALING_TYPE_YARN, + LLAMA_ROPE_SCALING_TYPE_LONGROPE = 3, + LLAMA_ROPE_SCALING_TYPE_MAX_VALUE = LLAMA_ROPE_SCALING_TYPE_LONGROPE, }; enum llama_pooling_type { @@ -272,6 +273,9 @@ extern "C" { }; struct llama_model_params { + // NULL-terminated list of devices to use for offloading (if NULL, all available devices are used) + ggml_backend_dev_t * devices; + int32_t n_gpu_layers; // number of layers to store in VRAM enum llama_split_mode split_mode; // how to split the model across multiple GPUs @@ -987,6 +991,9 @@ extern "C" { char * buf, int32_t length); + // Get list of built-in chat templates + LLAMA_API int32_t llama_chat_builtin_templates(const char ** output, size_t len); + // // Sampling API // diff --git a/examples/talk-llama/unicode.cpp b/examples/talk-llama/unicode.cpp index 50b35bbb..3d459263 100644 --- a/examples/talk-llama/unicode.cpp +++ b/examples/talk-llama/unicode.cpp @@ -201,7 +201,18 @@ static std::unordered_map unicode_utf8_to_byte_map() { } static inline std::wstring unicode_wstring_from_utf8(const std::string & s) { +#if defined(__clang__) + // disable C++17 deprecation warning for std::codecvt_utf8 +# pragma clang diagnostic push +# pragma clang diagnostic ignored "-Wdeprecated-declarations" +#endif + std::wstring_convert> conv; + +#if defined(__clang__) +# pragma clang diagnostic pop +#endif + return conv.from_bytes(s); }