@fugood/llama.node 1.4.2 → 1.4.4

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

@@ -423,8 +423,8 @@ static buft_list_t make_gpu_buft_list(ggml_backend_dev_t dev, llama_split_mode s
 }
 
 struct llama_model::impl {
-    impl() {}
-    ~impl() {}
+    impl() = default;
+    ~impl() = default;
 
     uint64_t n_elements = 0;
 
@@ -461,7 +461,7 @@ llama_model::llama_model(const llama_model_params & params) : params(params), pi
     pimpl->has_tensor_overrides = params.tensor_buft_overrides && params.tensor_buft_overrides[0].pattern;
 }
 
-llama_model::~llama_model() {}
+llama_model::~llama_model() = default;
 
 void llama_model::load_stats(llama_model_loader & ml) {
     pimpl->n_elements = ml.n_elements;
@@ -663,8 +663,10 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                     hparams.swa_type             = LLAMA_SWA_TYPE_NONE;
                     hparams.n_no_rope_layer_step = hparams.n_layer; // always use rope
                 } else {
-                    hparams.swa_type      = LLAMA_SWA_TYPE_CHUNKED;
-                    hparams.n_swa         = 8192;
+                    hparams.swa_type                = LLAMA_SWA_TYPE_CHUNKED;
+                    hparams.n_swa                   = 8192;
+                    hparams.n_attn_temp_floor_scale = 8192;
+                    hparams.f_attn_temp_scale       = 0.1f;
                     hparams.set_swa_pattern(4);   // pattern: 3 chunked - 1 full
                 }
 
@@ -1626,6 +1628,10 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                 }
                 ml.get_key(LLM_KV_ROPE_SCALING_YARN_LOG_MUL, hparams.rope_yarn_log_mul, false);
 
+                // (optional) temperature tuning - used by mistral-large
+                ml.get_key(LLM_KV_ATTENTION_TEMPERATURE_SCALE,  hparams.f_attn_temp_scale,       false);
+                ml.get_key(LLM_KV_ATTENTION_TEMPERATURE_LENGTH, hparams.n_attn_temp_floor_scale, false);
+
                 switch (hparams.n_layer) {
                     case 27: type = LLM_TYPE_16B; break;
                     case 60: type = LLM_TYPE_236B; break;
@@ -2247,6 +2253,42 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                     default: type = LLM_TYPE_UNKNOWN;
                 }
             } break;
+        case LLM_ARCH_MISTRAL3:
+            {
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+                ml.get_key(LLM_KV_ATTENTION_TEMPERATURE_SCALE, hparams.f_attn_temp_scale, false);
+
+                ml.get_key(LLM_KV_ROPE_SCALING_YARN_BETA_FAST,   hparams.yarn_beta_fast, false);
+                ml.get_key(LLM_KV_ROPE_SCALING_YARN_BETA_SLOW,   hparams.yarn_beta_slow, false);
+                ml.get_key(LLM_KV_ROPE_SCALING_YARN_LOG_MUL,     hparams.rope_yarn_log_mul, false);
+
+                // TODO: maybe add n_attn_temp_floor_scale as a separate KV?
+                if (hparams.f_attn_temp_scale != 0.0f) {
+                    hparams.n_attn_temp_floor_scale = hparams.n_ctx_orig_yarn;
+                    if (hparams.n_attn_temp_floor_scale == 0) {
+                        throw std::runtime_error("invalid n_ctx_orig_yarn for attention temperature scaling");
+                    }
+                }
+
+                // TODO: this seems to be correct with the case of mscale == mscale_all_dims == 1.0f
+                //       but may need further verification with other values
+                if (hparams.rope_yarn_log_mul != 0.0f) {
+                    float factor = 1.0f / hparams.rope_freq_scale_train;
+                    float mscale = 1.0f;
+                    float mscale_all_dims = hparams.rope_yarn_log_mul;
+                    static auto get_mscale = [](float scale, float mscale) {
+                        return scale <= 1.0f ? 1.0f : (0.1f * mscale * logf(scale) + 1.0f);
+                    };
+                    hparams.yarn_attn_factor = get_mscale(factor, mscale) / get_mscale(factor, mscale_all_dims);
+                }
+
+                switch (hparams.n_layer) {
+                    case 26: type = LLM_TYPE_3B; break;
+                    case 34: type = LLM_TYPE_8B; break;
+                    case 40: type = LLM_TYPE_14B; break;
+                    default: type = LLM_TYPE_UNKNOWN;
+                }
+            } break;
         default: throw std::runtime_error("unsupported model architecture");
     }
 
@@ -2560,6 +2602,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
             case LLM_ARCH_MINICPM:
             case LLM_ARCH_GRANITE:
             case LLM_ARCH_GRANITE_MOE:
+            case LLM_ARCH_MISTRAL3:
                 {
                     tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
 
@@ -6487,7 +6530,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                             layer.ssm_in         = create_tensor(tn(LLM_TENSOR_SSM_IN,         "weight", i), { n_embd, qkvz_dim }, 0);
                             layer.ssm_conv1d     = create_tensor(tn(LLM_TENSOR_SSM_CONV1D,     "weight", i), { hparams.ssm_d_conv, conv_dim }, 0);
                             layer.ssm_dt         = create_tensor(tn(LLM_TENSOR_SSM_DT,         "bias",   i), { hparams.ssm_dt_rank }, 0);
-                            layer.ssm_a          = create_tensor(tn(LLM_TENSOR_SSM_A,                    i), { hparams.ssm_dt_rank }, 0);
+                            layer.ssm_a          = create_tensor(tn(LLM_TENSOR_SSM_A_NOSCAN,             i), { hparams.ssm_dt_rank }, 0);
                             layer.ssm_beta_alpha = create_tensor(tn(LLM_TENSOR_SSM_BETA_ALPHA, "weight", i), { n_embd, ba_dim }, 0);
                             layer.ssm_norm       = create_tensor(tn(LLM_TENSOR_SSM_NORM,       "weight", i), { head_v_dim }, 0);
                             layer.ssm_out        = create_tensor(tn(LLM_TENSOR_SSM_OUT,        "weight", i), { value_dim, n_embd }, 0);
@@ -7522,6 +7565,10 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
             {
                 llm = std::make_unique<llm_build_qwen3next>(*this, params);
             } break;
+        case LLM_ARCH_MISTRAL3:
+            {
+                llm = std::make_unique<llm_build_mistral3>(*this, params);
+            } break;
         default:
             GGML_ABORT("fatal error");
     }
@@ -7690,6 +7737,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
         case LLM_ARCH_ARCEE:
         case LLM_ARCH_ERNIE4_5:
         case LLM_ARCH_ERNIE4_5_MOE:
+        case LLM_ARCH_MISTRAL3:
             return LLAMA_ROPE_TYPE_NORM;
 
         // the pairs of head values are offset by n_rot/2

package/src/llama.cpp/src/llama-quant.cpp

@@ -666,7 +666,6 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
 
     std::map<int, std::string> mapped;
     int blk_id = 0;
-    int pruned_attention_w = 0;
 
     // make a list of weights
     std::vector<const llama_model_loader::llama_tensor_weight *> tensors;
@@ -674,11 +673,6 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
     for (const auto & it : ml.weights_map) {
         const std::string remapped_name(remap_layer(it.first, prune_list, mapped, blk_id));
         if (remapped_name.empty()) {
-            if (it.first.find("attn_v.weight") != std::string::npos ||
-                it.first.find("attn_qkv.weight") != std::string::npos ||
-                it.first.find("attn_kv_b.weight") != std::string::npos) {
-                    pruned_attention_w++;
-            }
             LLAMA_LOG_DEBUG("%s: pruning tensor %s\n", __func__, it.first.c_str());
             continue;
         }
@@ -703,7 +697,6 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
         });
     }
 
-    bool is_clip_model = false;
     for (const auto * it : tensors) {
         const struct ggml_tensor * tensor = it->tensor;
 
@@ -717,32 +710,10 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
         } else if (name == LLM_TN(model.arch)(LLM_TENSOR_OUTPUT, "weight")) {
             qs.has_output = true;
         }
-
-        is_clip_model |= name.rfind("mm.", 0) == 0; // check the "mm." prefix
     }
 
     qs.n_ffn_down = qs.n_ffn_gate = qs.n_ffn_up = (int)model.hparams.n_layer;
 
-    // sanity checks for models that have attention layers
-    if (qs.n_attention_wv != 0 && !is_clip_model)
-    {
-        const auto & n_head_kv_iter = model.hparams.n_head_kv_arr.begin();
-        // attention layers have a non-zero number of kv heads
-        int32_t n_layer_attn = model.hparams.n_layer - std::count(n_head_kv_iter, n_head_kv_iter + model.hparams.n_layer, 0);
-        if (llama_model_has_encoder(&model)) {
-            // now n_layer_attn is the number of attention layers in the encoder
-            // for each decoder block, there are 2 attention layers
-            n_layer_attn += 2 * model.hparams.dec_n_layer;
-        }
-
-        // note: for linear-attention models (such as Qwen3 Next) this is the number of linear layers
-        const int32_t n_layer_recr = std::count(model.hparams.recurrent_layer_arr.begin(), model.hparams.recurrent_layer_arr.end(), true);
-
-        LLAMA_LOG_INFO("%s: n_layer_attn = %d, n_layer_recr = %d, pruned_attention_w = %d\n", __func__, n_layer_attn, n_layer_recr, pruned_attention_w);
-
-        GGML_ASSERT((qs.n_attention_wv == n_layer_attn - pruned_attention_w - n_layer_recr) && "n_attention_wv is unexpected");
-    }
-
     size_t total_size_org = 0;
     size_t total_size_new = 0;
 

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

@@ -3253,8 +3253,7 @@ void llama_vocab::impl::print_info() const {
 llama_vocab::llama_vocab() : pimpl(new impl(*this)) {
 }
 
-llama_vocab::~llama_vocab() {
-}
+llama_vocab::~llama_vocab() = default;
 
 void llama_vocab::load(llama_model_loader & ml, const LLM_KV & kv) {
     pimpl->load(ml, kv);

package/src/llama.cpp/src/models/deepseek2.cpp

@@ -30,6 +30,12 @@ llm_build_deepseek2::llm_build_deepseek2(const llama_model & model, const llm_gr
     // {n_embd, n_tokens}
     inpL = build_inp_embd(model.tok_embd);
 
+    // (optional) temperature tuning - used by mistral-large
+    ggml_tensor * inp_attn_scale = nullptr;
+    if (hparams.f_attn_temp_scale != 0.0f) {
+        inp_attn_scale = build_inp_attn_scale();
+    }
+
     // inp_pos - contains the positions
     ggml_tensor * inp_pos = build_inp_pos();
 
@@ -128,6 +134,12 @@ llm_build_deepseek2::llm_build_deepseek2(const llama_model & model, const llm_gr
                 ggml_tensor * Vcur = kv_cmpr;
                 cb(Vcur, "Vcur", il);
 
+                if (inp_attn_scale) {
+                    // apply llama 4 temperature scaling
+                    Qcur = ggml_mul(ctx0, Qcur, inp_attn_scale);
+                    cb(Qcur, "Qcur_attn_temp_scaled", il);
+                }
+
                 // note: MLA with the absorption optimzation converts into MQA (ie: GQA with 1 group)
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, NULL,
@@ -160,6 +172,12 @@ llm_build_deepseek2::llm_build_deepseek2(const llama_model & model, const llm_gr
                 ggml_tensor * Kcur = ggml_concat(ctx0, ggml_repeat(ctx0, k_pe, q_pe), k_nope, 0);
                 cb(Kcur, "Kcur", il);
 
+                if (inp_attn_scale) {
+                    // apply llama 4 temperature scaling
+                    Qcur = ggml_mul(ctx0, Qcur, inp_attn_scale);
+                    cb(Qcur, "Qcur_attn_temp_scaled", il);
+                }
+
                 // note: MLA without the absorption optimization converts into MHA (ie: GQA with full n_head groups)
                 cur = build_attn(inp_attn,
                             model.layers[il].wo, NULL,

package/src/llama.cpp/src/models/mistral3.cpp

@@ -0,0 +1,160 @@
+#include "models.h"
+
+llm_build_mistral3::llm_build_mistral3(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    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);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    // (optional) temperature tuning
+    ggml_tensor * inp_attn_scale = nullptr;
+    if (hparams.f_attn_temp_scale != 0.0f) {
+        inp_attn_scale = build_inp_attn_scale();
+    }
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            // rope freq factors for llama3; may return nullptr for llama2 and other models
+            ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
+
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(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);
+            }
+            ggml_tensor * Kcur = build_lora_mm(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);
+            }
+            ggml_tensor * Vcur = build_lora_mm(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_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, rope_factors,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, rope_factors,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            if (inp_attn_scale) {
+                // apply llama 4 temperature scaling
+                Qcur = ggml_mul(ctx0, Qcur, inp_attn_scale);
+                cb(Qcur, "Qcur_attn_temp_scaled", il);
+            }
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
+            cb(cur, "attn_out", il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network (non-MoE)
+        if (model.layers[il].ffn_gate_inp == nullptr) {
+
+            cur = build_norm(ffn_inp,
+                    model.layers[il].ffn_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "ffn_norm", il);
+
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
+                    model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
+                    model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+                    NULL,
+                    LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        } else {
+            // MoE branch
+            cur = build_norm(ffn_inp,
+                    model.layers[il].ffn_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "ffn_norm", il);
+
+            cur = build_moe_ffn(cur,
+                    model.layers[il].ffn_gate_inp,
+                    model.layers[il].ffn_up_exps,
+                    model.layers[il].ffn_gate_exps,
+                    model.layers[il].ffn_down_exps,
+                    nullptr,
+                    n_expert, n_expert_used,
+                    LLM_FFN_SILU, true,
+                    false, 0.0,
+                    LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
+                    il);
+            cb(cur, "ffn_moe_out", il);
+        }
+        cur = ggml_add(ctx0, cur, ffn_inp);
+        cb(cur, "ffn_out", il);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm, NULL,
+            LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

package/src/llama.cpp/src/models/models.h

@@ -322,6 +322,10 @@ struct llm_build_minimax_m2 : public llm_graph_context {
     llm_build_minimax_m2(const llama_model & model, const llm_graph_params & params);
 };
 
+struct llm_build_mistral3 : public llm_graph_context {
+    llm_build_mistral3(const llama_model & model, const llm_graph_params & params);
+};
+
 struct llm_build_mpt : public llm_graph_context {
     llm_build_mpt(const llama_model & model, const llm_graph_params & params);
 };

package/src/llama.cpp/src/unicode.cpp

@@ -499,7 +499,7 @@ static std::vector<size_t> unicode_regex_split_custom_llama3(const std::string &
 
 // use std::wregex to split the text
 static std::vector<size_t> unicode_regex_split_stl(const std::wstring & wtext, const std::wstring & regex_expr, const std::vector<size_t> & offsets) {
-    std::wregex expr(regex_expr);
+    std::wregex expr(regex_expr, std::regex_constants::optimize | std::regex_constants::nosubs);
     std::vector<size_t> bpe_offsets; // store the offset of each word
     bpe_offsets.reserve(offsets.size()); // Reserve memory for the approximate size
     size_t start = 0;
@@ -529,7 +529,7 @@ static std::vector<size_t> unicode_regex_split_stl(const std::wstring & wtext, c
 
 // use std::regex to split the text
 static std::vector<size_t> unicode_regex_split_stl(const std::string & text, const std::string & regex_expr, const std::vector<size_t> & offsets) {
-    std::regex expr(regex_expr);
+    std::regex expr(regex_expr, std::regex_constants::optimize | std::regex_constants::nosubs);
     std::vector<size_t> bpe_offsets; // store the offset of each word
     bpe_offsets.reserve(offsets.size()); // Reserve memory for the approximate size
     size_t start = 0;