@fugood/llama.node 0.3.9 → 0.3.11

package/src/llama.cpp/src/llama-model.cpp

@@ -1093,8 +1093,20 @@ void llama_model::load_hparams(llama_model_loader & ml) {
             {
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
                 switch (hparams.n_layer) {
-                    case 28: type = LLM_TYPE_6B; break;
-                    case 40: type = LLM_TYPE_9B; break;
+                    case 28: {
+                        if (hparams.n_head(0) == 16) {
+                            type = LLM_TYPE_1_5B;
+                        } else {
+                            type = LLM_TYPE_6B;
+                        }
+                    } break;
+                    case 40: {
+                        if (hparams.n_head(0) == 24) {
+                            type = LLM_TYPE_4B;
+                        } else {
+                            type = LLM_TYPE_9B;
+                        }
+                    } break;
                     default: type = LLM_TYPE_UNKNOWN;
                 }
             } break;
@@ -1263,6 +1275,8 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
 
     const bool use_mmap_buffer = true;
 
+    LLAMA_LOG_INFO("%s: loading model tensors, this can take a while... (mmap = %s)\n", __func__, use_mmap_buffer ? "true" : "false");
+
     // build a list of buffer types for the CPU and GPU devices
     pimpl->cpu_buft_list = make_cpu_buft_list(devices);
     for (auto * dev : devices) {
@@ -1303,10 +1317,12 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
     const int act_gpu_layers = devices.empty() ? 0 : std::min(n_gpu_layers, (int)n_layer + 1);
     auto get_layer_buft_list = [&](int il) -> llama_model::impl::layer_dev {
         if (il < i_gpu_start || (il - i_gpu_start) >= act_gpu_layers) {
+            LLAMA_LOG_DEBUG("load_tensors: layer %3d assigned to device %s\n", il, ggml_backend_dev_name(cpu_dev));
             return {cpu_dev, &pimpl->cpu_buft_list};
         }
         const int layer_gpu = std::upper_bound(splits.begin(), splits.begin() + n_devices(), float(il - i_gpu_start)/act_gpu_layers) - splits.begin();
         auto * dev = devices.at(layer_gpu);
+        LLAMA_LOG_DEBUG("load_tensors: layer %3d assigned to device %s\n", il, ggml_backend_dev_name(dev));
         return {dev, &pimpl->gpu_buft_list.at(dev)};
     };
 
@@ -2203,6 +2219,50 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         layer.rope_short = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_SHORT, "weight", i), { n_embd_head/2 }, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
                     }
                 } break;
+            case LLM_ARCH_PHIMOE:
+                {
+                    const int64_t n_embd_head = n_embd / n_head;
+
+                    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), { n_embd, n_vocab }, 0);
+
+                    // output
+                    output_norm   = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), { n_embd }, 0);
+                    output_norm_b = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "bias"),   {n_embd}, 0);
+                    output        = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), { n_embd, n_vocab }, 0);
+                    output_b      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "bias"),   { n_vocab }, 0);
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        auto & layer = layers[i];
+
+                        layer.attn_norm   = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), { n_embd }, 0);
+                        layer.attn_norm_b = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "bias",   i), { n_embd }, 0);
+
+                        layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), { n_embd, n_embd + 2 * n_embd_gqa }, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        if (layer.wqkv == nullptr) {
+                            layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd}, 0);
+                            layer.bq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "bias",   i), {n_embd}, 0);
+
+                            layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_gqa}, 0);
+                            layer.bk = create_tensor(tn(LLM_TENSOR_ATTN_K, "bias",   i), {n_embd_gqa}, 0);
+
+                            layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_gqa}, 0);
+                            layer.bv = create_tensor(tn(LLM_TENSOR_ATTN_V, "bias",   i), {n_embd_gqa}, 0);
+                        }
+                        layer.wo   = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), { n_embd, n_embd }, 0);
+                        layer.bo   = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias",   i), { n_embd }, 0);
+
+                        layer.ffn_norm   = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), { n_embd }, 0);
+                        layer.ffn_norm_b = create_tensor(tn(LLM_TENSOR_FFN_NORM, "bias",   i), { n_embd }, 0);
+
+                        layer.ffn_gate_inp  = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,  "weight", i), {n_embd, n_expert},         0);
+                        layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd, n_ff,   n_expert}, 0);
+                        layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff,   n_embd, n_expert}, 0);
+                        layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {n_embd, n_ff,   n_expert}, 0);
+
+                        layer.rope_long  = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_LONG,  "weight", i), { n_embd_head/2 }, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
+                        layer.rope_short = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_SHORT, "weight", i), { n_embd_head/2 }, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
+                     }
+                } break;
             case LLM_ARCH_PLAMO:
                 {
                     tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
@@ -3022,9 +3082,17 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         auto & layer = layers[i];
 
                         layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+                        layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.bqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "bias", i),   {n_embd + 2*n_embd_gqa}, llama_model_loader::TENSOR_NOT_REQUIRED);
 
-                        layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, 0);
-                        layer.bqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "bias", i),   {n_embd + 2*n_embd_gqa}, 0);
+                        if (layer.wqkv == nullptr) {
+                            layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd_head_k * n_head}, 0);
+                            layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_k_gqa}, 0);
+                            layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_v_gqa}, 0);
+                            layer.bq = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "bias", i), {n_embd},     llama_model_loader::TENSOR_NOT_REQUIRED);
+                            layer.bk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "bias", i), {n_embd_gqa}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                            layer.bv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "bias", i), {n_embd_gqa}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        }
 
                         layer.wo   = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
 
@@ -3911,8 +3979,10 @@ uint64_t llama_model_size(const struct llama_model * model) {
     return model->size();
 }
 
-const char * llama_model_chat_template(const struct llama_model * model) {
-    const auto & it = model->gguf_kv.find(LLM_KV(model->arch)(LLM_KV_TOKENIZER_CHAT_TEMPLATE));
+const char * llama_model_chat_template(const struct llama_model * model, const char * name) {
+    const auto key = name ? LLM_KV(model->arch, name)(LLM_KV_TOKENIZER_CHAT_TEMPLATE_N)
+        : LLM_KV(model->arch)(LLM_KV_TOKENIZER_CHAT_TEMPLATE);
+    const auto & it = model->gguf_kv.find(key);
     if (it == model->gguf_kv.end()) {
         return nullptr;
     }

package/src/llama.cpp/src/llama-sampling.cpp

@@ -1433,13 +1433,30 @@ static void llama_sampler_grammar_apply(struct llama_sampler * smpl, llama_token
     }
 }
 
+// Fwd declare to break reset --> init_impl --> llama_sampler_grammar_i --> reset cycle.
+static struct llama_sampler * llama_sampler_init_grammar_impl(
+        const struct llama_vocab * vocab,
+                      const char * grammar_str,
+                      const char * grammar_root,
+                              bool lazy,
+                     const char ** trigger_words,
+                            size_t num_trigger_words,
+               const llama_token * trigger_tokens,
+                            size_t num_trigger_tokens);
+
 static void llama_sampler_grammar_reset(struct llama_sampler * smpl) {
     auto * ctx = (llama_sampler_grammar *) smpl->ctx;
     if (!ctx->grammar) {
         return;
     }
 
-    auto * grammar_new = llama_grammar_init_impl(ctx->grammar->vocab, ctx->grammar_str.c_str(), ctx->grammar_root.c_str());
+    std::vector<const char *>  trigger_words;
+    for (auto & word : ctx->grammar->trigger_words) {
+        trigger_words.push_back(word.c_str());
+    }
+    auto * grammar_new = llama_grammar_init_impl(ctx->grammar->vocab, ctx->grammar_str.c_str(), ctx->grammar_root.c_str(),
+                                                 ctx->grammar->lazy, trigger_words.data(), trigger_words.size(),
+                                                 ctx->grammar->trigger_tokens.data(), ctx->grammar->trigger_tokens.size());
 
     llama_grammar_free_impl(ctx->grammar);
     ctx->grammar = grammar_new;
@@ -1448,7 +1465,7 @@ static void llama_sampler_grammar_reset(struct llama_sampler * smpl) {
 static struct llama_sampler * llama_sampler_grammar_clone(const struct llama_sampler * smpl) {
     const auto * ctx = (const llama_sampler_grammar *) smpl->ctx;
 
-    auto * result = llama_sampler_init_grammar(ctx->vocab, nullptr, nullptr);
+    auto * result = llama_sampler_init_grammar_impl(ctx->vocab, nullptr, nullptr, false, nullptr, 0, nullptr, 0);
 
     // copy the state
     {
@@ -1484,7 +1501,15 @@ static struct llama_sampler_i llama_sampler_grammar_i = {
     /* .free   = */ llama_sampler_grammar_free,
 };
 
-struct llama_sampler * llama_sampler_init_grammar(const struct llama_vocab * vocab, const char * grammar_str, const char * grammar_root) {
+static struct llama_sampler * llama_sampler_init_grammar_impl(
+        const struct llama_vocab * vocab,
+                      const char * grammar_str,
+                      const char * grammar_root,
+                              bool lazy,
+                     const char ** trigger_words,
+                            size_t num_trigger_words,
+               const llama_token * trigger_tokens,
+                            size_t num_trigger_tokens) {
     auto * ctx = new llama_sampler_grammar;
 
     if (grammar_str != nullptr && grammar_str[0] != '\0') {
@@ -1492,7 +1517,7 @@ struct llama_sampler * llama_sampler_init_grammar(const struct llama_vocab * voc
             /* .vocab        = */ vocab,
             /* .grammar_str  = */ grammar_str,
             /* .grammar_root = */ grammar_root,
-            /* .grammar      = */ llama_grammar_init_impl(vocab, grammar_str, grammar_root),
+            /* .grammar      = */ llama_grammar_init_impl(vocab, grammar_str, grammar_root, lazy, trigger_words, num_trigger_words, trigger_tokens, num_trigger_tokens),
         };
     } else {
         *ctx = {
@@ -1509,6 +1534,24 @@ struct llama_sampler * llama_sampler_init_grammar(const struct llama_vocab * voc
     };
 }
 
+struct llama_sampler * llama_sampler_init_grammar(
+        const struct llama_vocab * vocab,
+                      const char * grammar_str,
+                      const char * grammar_root) {
+    return llama_sampler_init_grammar_impl(vocab, grammar_str, grammar_root, /* lazy= */ false, nullptr, 0, nullptr, 0);
+}
+
+struct llama_sampler * llama_sampler_init_grammar_lazy(
+        const struct llama_vocab * vocab,
+                      const char * grammar_str,
+                      const char * grammar_root,
+                     const char ** trigger_words,
+                            size_t num_trigger_words,
+               const llama_token * trigger_tokens,
+                            size_t num_trigger_tokens) {
+    return llama_sampler_init_grammar_impl(vocab, grammar_str, grammar_root, /* lazy= */ true, trigger_words, num_trigger_words, trigger_tokens, num_trigger_tokens);
+}
+
 // penalties
 
 struct llama_sampler_penalties {

package/src/llama.cpp/src/llama-vocab.cpp

@@ -1245,8 +1245,13 @@ struct llama_vocab::impl {
 
     std::vector<llama_token> cache_special_tokens;
     std::vector<std::string> cache_token_to_piece; // llama_token_to_piece(special = true);
-
-    std::map<std::pair<std::string, std::string>, int> bpe_ranks;
+    struct pair_hash {
+        size_t operator()(const std::pair<std::string, std::string> & p) const {
+            return std::hash<std::string>{}(p.first) ^  //create some hash for pair
+                   (std::hash<std::string>{}(p.second) << 1);
+        }
+    };
+    std::unordered_map<std::pair<std::string, std::string>, int, pair_hash> bpe_ranks;
 
     // set of all tokens that cause "end of generation"
     std::set<llama_token> special_eog_ids;
@@ -1523,7 +1528,8 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                 pre_type = LLAMA_VOCAB_PRE_TYPE_COMMAND_R;
                 clean_spaces = false;
             } else if (
-                tokenizer_pre == "qwen2") {
+                    tokenizer_pre == "qwen2" ||
+                    tokenizer_pre == "deepseek-r1-qwen") {
                 pre_type = LLAMA_VOCAB_PRE_TYPE_QWEN2;
                 clean_spaces = false;
             } else if (
@@ -1686,7 +1692,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
         GGML_ASSERT(!ids.empty() && "model vocab missing newline token");
         linefeed_id = ids[0];
     } else {
-        const std::vector<int> ids = tokenize("\xC4\x8A", false); // U+010A
+        const std::vector<int> ids = tokenize("\n", false);
 
         //GGML_ASSERT(!ids.empty() && "model vocab missing newline token");
         if (ids.empty()) {