@fugood/llama.node 1.1.5 → 1.1.6

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

@@ -211,7 +211,10 @@ static ggml_type llama_tensor_get_type(quantize_state_impl & qs, ggml_type new_t
             const int64_t nx = tensor->ne[0];
             const int64_t qk_k = ggml_blck_size(new_type);
 
-            if (arch == LLM_ARCH_FALCON || nx % qk_k != 0) {
+            if (ftype == LLAMA_FTYPE_MOSTLY_MXFP4_MOE) {
+                new_type = GGML_TYPE_Q8_0;
+            }
+            else if (arch == LLM_ARCH_FALCON || nx % qk_k != 0) {
                 new_type = GGML_TYPE_Q8_0;
             }
             else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS ||
@@ -223,6 +226,14 @@ static ggml_type llama_tensor_get_type(quantize_state_impl & qs, ggml_type new_t
                 new_type = GGML_TYPE_Q6_K;
             }
         }
+    } else if (ftype == LLAMA_FTYPE_MOSTLY_MXFP4_MOE) {
+        // MoE   tensors -> MXFP4
+        // other tensors -> Q8_0
+        if (tensor->ne[2] > 1) {
+            new_type = GGML_TYPE_MXFP4;
+        } else {
+            new_type = GGML_TYPE_Q8_0;
+        }
     } else if (name == "token_embd.weight" || name == "per_layer_token_embd.weight") {
         if (qs.params->token_embedding_type < GGML_TYPE_COUNT) {
             new_type = qs.params->token_embedding_type;
@@ -533,6 +544,8 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
         case LLAMA_FTYPE_MOSTLY_BF16: default_type = GGML_TYPE_BF16; break;
         case LLAMA_FTYPE_ALL_F32:     default_type = GGML_TYPE_F32;  break;
 
+        case LLAMA_FTYPE_MOSTLY_MXFP4_MOE: default_type = GGML_TYPE_MXFP4; break;
+
         // K-quants
         case LLAMA_FTYPE_MOSTLY_Q2_K_S:
         case LLAMA_FTYPE_MOSTLY_Q2_K:    default_type = GGML_TYPE_Q2_K;    break;
@@ -984,6 +997,29 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
                 const float * imatrix_03 = imatrix ? imatrix + i03 * n_per_row : nullptr;
 
                 new_size += llama_tensor_quantize_impl(new_type, f32_data_03, new_data_03, chunk_size, nrows, n_per_row, imatrix_03, workers, nthread_use);
+
+                // TODO: temporary sanity check that the F16 -> MXFP4 is lossless
+#if 1
+                if (new_type == GGML_TYPE_MXFP4) {
+                    auto * x = f32_data_03;
+
+                    //LLAMA_LOG_INFO("nrows = %d, n_per_row = %d\n", nrows, n_per_row);
+                    std::vector<float> deq(nrows*n_per_row);
+                    const ggml_type_traits * qtype = ggml_get_type_traits(new_type);
+                    qtype->to_float(new_data_03, deq.data(), deq.size());
+
+                    double err = 0.0f;
+                    for (int i = 0; i < (int) deq.size(); ++i) {
+                        err += fabsf(deq[i] - x[i]);
+                        //if (fabsf(deq[i] - x[i]) > 0.00001 && i < 256) {
+                        if (deq[i] != x[i]) {
+                            LLAMA_LOG_INFO("deq[%d] = %f, x[%d] = %f\n", i, deq[i], i, x[i]);
+                        }
+                    }
+                    //LLAMA_LOG_INFO("err = %f\n", err);
+                    GGML_ASSERT(err == 0.00000);
+                }
+#endif
             }
             LLAMA_LOG_INFO("size = %8.2f MiB -> %8.2f MiB\n", ggml_nbytes(tensor)/1024.0/1024.0, new_size/1024.0/1024.0);
         }

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

@@ -2191,6 +2191,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                         || t.first == "<｜fim▁begin｜>" // DeepSeek
                         || t.first == "<PRE>"
                         || t.first == "▁<PRE>"          // CodeLlama
+                        || t.first == "<|code_prefix|>" // GLM-4.5
                         ) {
                     special_fim_pre_id = t.second;
                     if ((id_to_token[t.second].attr & LLAMA_TOKEN_ATTR_CONTROL) == 0) {
@@ -2210,6 +2211,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                         || t.first == "<｜fim▁hole｜>" // DeepSeek
                         || t.first == "<SUF>"
                         || t.first == "▁<SUF>"         // CodeLlama
+                        || t.first == "<|code_suffix|>" // GLM-4.5
                         ) {
                     special_fim_suf_id = t.second;
                     if ((id_to_token[t.second].attr & LLAMA_TOKEN_ATTR_CONTROL) == 0) {
@@ -2229,6 +2231,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                         || t.first == "<｜fim▁end｜>"  // DeepSeek
                         || t.first == "<MID>"
                         || t.first == "▁<MID>"         // CodeLlama
+                        || t.first == "<|code_middle|>" // GLM-4.5
                         ) {
                     special_fim_mid_id = t.second;
                     if ((id_to_token[t.second].attr & LLAMA_TOKEN_ATTR_CONTROL) == 0) {
@@ -2311,6 +2314,8 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                     || t.first == "<|eot_id|>"
                     || t.first == "<|im_end|>"
                     || t.first == "<|end|>"
+                    || t.first == "<|return|>" // o200k_harmony
+                    || t.first == "<|call|>"   // o200k_harmony
                     || t.first == "<end_of_turn>"
                     || t.first == "<|endoftext|>"
                     || t.first == "<|eom_id|>"
@@ -2334,6 +2339,13 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
             }
         }
 
+        // @ngxson : quick hack for gpt-oss, always render these tokens
+        for (const auto & t : token_to_id) {
+            if (t.first == "<|channel|>" || t.first == "<|message|>" || t.first == "<|start|>") {
+                id_to_token[t.second].attr = LLAMA_TOKEN_ATTR_USER_DEFINED;
+            }
+        }
+
         // sanity checks
         if (special_eos_id != LLAMA_TOKEN_NULL && special_eog_ids.count(special_eos_id) == 0) {
             special_eog_ids.insert(special_eos_id);
@@ -2349,6 +2361,36 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
             special_eog_ids.insert(special_eom_id);
             LLAMA_LOG_WARN("%s: special_eom_id is not in special_eog_ids - the tokenizer config may be incorrect\n", __func__);
         }
+
+        // TODO: workaround for o200k_harmony tokenizer: the "<|end|>" token should not be EOG
+        //       we don't have a good way to detect this, so for now, if we have "<|return|>" and "<|call|>" tokens,
+        //       we remove the "<|end|>" token from the EOG list
+        {
+            bool has_return = false;
+            bool has_call   = false;
+            bool has_end    = false;
+
+            llama_token end_id = LLAMA_TOKEN_NULL;
+
+            LLAMA_LOG_INFO("%s: printing all EOG tokens:\n", __func__);
+            for (auto tid : special_eog_ids) {
+                LLAMA_LOG_INFO("%s:   - %d ('%s')\n", __func__, tid, id_to_token[tid].text.c_str());
+
+                if (id_to_token[tid].text == "<|return|>") {
+                    has_return = true;
+                } else if (id_to_token[tid].text == "<|call|>") {
+                    has_call = true;
+                } else if (id_to_token[tid].text == "<|end|>") {
+                    has_end = true;
+                    end_id = tid;
+                }
+            }
+
+            if (has_return && has_call && has_end) {
+                special_eog_ids.erase(end_id);
+                LLAMA_LOG_WARN("%s: special_eog_ids contains both '<|return|>' and '<|call|>' tokens, removing '<|end|>' token from EOG list\n", __func__);
+            }
+        }
     }
 
     // build special tokens cache