@fugood/llama.node 1.2.5 → 1.3.0-rc.0

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

@@ -478,7 +478,8 @@ void llama_model::load_hparams(llama_model_loader & ml) {
     ml.get_key(LLM_KV_GENERAL_NAME, name, false);
 
     // everything past this point is not vocab-related
-    if (hparams.vocab_only) {
+    // for CLIP models, we only need to load tensors, no hparams
+    if (hparams.vocab_only || ml.get_arch() == LLM_ARCH_CLIP) {
         return;
     }
 
@@ -11358,8 +11359,8 @@ struct llm_build_gemma3n_iswa : public llm_graph_context {
     }
 };
 
-struct llm_build_gemma_embedding_iswa : public llm_graph_context {
-    llm_build_gemma_embedding_iswa(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+struct llm_build_gemma_embedding : public llm_graph_context {
+    llm_build_gemma_embedding(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
         const int64_t n_embd_head = hparams.n_embd_head_k;
 
         ggml_tensor * cur;
@@ -11376,8 +11377,7 @@ struct llm_build_gemma_embedding_iswa : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
 
-        // TODO: support cacheless iSWA embeddings [TAG_NO_CACHE_ISWA]
-        auto * inp_attn = build_attn_inp_kv_iswa();
+        auto * inp_attn = build_attn_inp_no_cache();
 
         ggml_tensor * inp_out_ids = build_inp_out_ids();
 
@@ -19378,7 +19378,7 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
         case LLM_ARCH_NOMIC_BERT_MOE:
         case LLM_ARCH_NEO_BERT:
         case LLM_ARCH_WAVTOKENIZER_DEC:
-        //case LLM_ARCH_GEMMA_EMBEDDING: // TODO: disabled until the cacheless SWA logic is fixed [TAG_NO_CACHE_ISWA]
+        case LLM_ARCH_GEMMA_EMBEDDING:
         case LLM_ARCH_DREAM:
         case LLM_ARCH_LLADA:
         case LLM_ARCH_LLADA_MOE:
@@ -19671,7 +19671,7 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
             } break;
         case LLM_ARCH_GEMMA_EMBEDDING:
             {
-                llm = std::make_unique<llm_build_gemma_embedding_iswa>(*this, params);
+                llm = std::make_unique<llm_build_gemma_embedding>(*this, params);
             } break;
         case LLM_ARCH_STARCODER2:
             {
@@ -20014,6 +20014,7 @@ int32_t llama_n_head(const llama_model * model) {
 llama_rope_type llama_model_rope_type(const llama_model * model) {
     switch (model->arch) {
         // these models do not use RoPE
+        case LLM_ARCH_CLIP:
         case LLM_ARCH_GPT2:
         case LLM_ARCH_GPTJ:
         case LLM_ARCH_MPT:

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

@@ -701,6 +701,7 @@ 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;
 
@@ -714,12 +715,14 @@ 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)
+    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
@@ -881,6 +884,9 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
         // do not quantize relative position bias (T5)
         quantize &= name.find("attn_rel_b.weight") == std::string::npos;
 
+        // do not quantize specific multimodal tensors
+        quantize &= name.find(".position_embd.") == std::string::npos;
+
         ggml_type new_type;
         void * new_data;
         size_t new_size;

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

@@ -124,6 +124,9 @@ static int llama_model_load(const std::string & fname, std::vector<std::string>
         } catch(const std::exception & e) {
             throw std::runtime_error("error loading model hyperparameters: " + std::string(e.what()));
         }
+        if (model.arch == LLM_ARCH_CLIP) {
+            throw std::runtime_error("CLIP cannot be used as main model, use it with --mmproj instead");
+        }
         try {
             model.load_vocab(ml);
         } catch(const std::exception & e) {
@@ -312,6 +315,7 @@ struct llama_model * llama_model_load_from_splits(
         LLAMA_LOG_ERROR("%s: list of splits is empty\n", __func__);
         return nullptr;
     }
+    splits.reserve(n_paths);
     for (size_t i = 0; i < n_paths; ++i) {
         splits.push_back(paths[i]);
     }