Latest commits

Author: MaggotHATE

Makefile

@@ -479,7 +479,6 @@ OBJS_GGUF_LLAMA = \
     $(TMP)$(PREFIX)_llama-hparams.o \
     $(TMP)$(PREFIX)_llama-impl.o \
     $(TMP)$(PREFIX)_llama-io.o \
-    $(TMP)$(PREFIX)_llama-kv-cache.o \
     $(TMP)$(PREFIX)_llama-kv-cache-unified.o \
     $(TMP)$(PREFIX)_llama-kv-cache-unified-iswa.o \
     $(TMP)$(PREFIX)_llama-kv-cache-recurrent.o \

base_sampling2/master/src/llama-arch.cpp

@@ -200,7 +200,6 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_TOKENIZER_HF_JSON,              "tokenizer.huggingface.json"              },
     { LLM_KV_TOKENIZER_RWKV,                 "tokenizer.rwkv.world"                    },
     { LLM_KV_TOKENIZER_CHAT_TEMPLATE,        "tokenizer.chat_template"                 },
-    { LLM_KV_TOKENIZER_CHAT_TEMPLATE_N,      "tokenizer.chat_template.%s"              },
     { LLM_KV_TOKENIZER_FIM_PRE_ID,           "tokenizer.ggml.fim_pre_token_id"         },
     { LLM_KV_TOKENIZER_FIM_SUF_ID,           "tokenizer.ggml.fim_suf_token_id"         },
     { LLM_KV_TOKENIZER_FIM_MID_ID,           "tokenizer.ggml.fim_mid_token_id"         },
@@ -1707,8 +1706,14 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
 LLM_KV::LLM_KV(llm_arch arch, const char * suffix) : arch(arch), suffix(suffix) {}
 
 std::string LLM_KV::operator()(llm_kv kv) const {
-    return suffix ? ::format(LLM_KV_NAMES.at(kv), LLM_ARCH_NAMES.at(arch), suffix)
-        : ::format(LLM_KV_NAMES.at(kv), LLM_ARCH_NAMES.at(arch));
+    std::string name = ::format(LLM_KV_NAMES.at(kv), LLM_ARCH_NAMES.at(arch));
+
+    if (suffix != nullptr) {
+        name += ".";
+        name += suffix;
+    }
+
+    return name;
 }
 
 std::string LLM_TN_IMPL::str() const {

base_sampling2/master/src/llama-arch.h

@@ -196,7 +196,6 @@ enum llm_kv {
     LLM_KV_TOKENIZER_HF_JSON,
     LLM_KV_TOKENIZER_RWKV,
     LLM_KV_TOKENIZER_CHAT_TEMPLATE,
-    LLM_KV_TOKENIZER_CHAT_TEMPLATE_N,
     LLM_KV_TOKENIZER_FIM_PRE_ID,
     LLM_KV_TOKENIZER_FIM_SUF_ID,
     LLM_KV_TOKENIZER_FIM_MID_ID,

base_sampling2/master/src/llama-graph.cpp

@@ -659,6 +659,28 @@ ggml_tensor * llm_graph_context::build_ffn(
                 cur = ggml_mul(ctx0, x0, x1);
                 cb(cur, "ffn_mul", il);
             } break;
+        case LLM_FFN_GEGLU:
+            {
+                // Split into two equal parts
+                int64_t split_point = cur->ne[0] / 2;
+                ggml_tensor * output_ffn_up = ggml_cont(ctx0, ggml_view_2d(
+                                                ctx0, cur, split_point,
+                                                cur->ne[1], cur->nb[1], 0
+                                            ));
+                ggml_tensor * output_ffn_gate = ggml_cont(ctx0, ggml_view_2d(
+                                                ctx0, cur, split_point,
+                                                cur->ne[1], cur->nb[1],
+                                                split_point * ggml_element_size(cur)
+                                            ));
+
+                // Apply GELU activation function to the first part
+                output_ffn_up = ggml_gelu(ctx0, output_ffn_up);
+                cb(output_ffn_up, "ffn_gelu", il);
+
+                // Element-wise multiplication between the activated part and the gate part
+                cur = ggml_mul(ctx0, output_ffn_up, output_ffn_gate);
+                cb(cur, "ffn_geglu", il);
+            } break;
     }
 
     if (gate && type_gate == LLM_FFN_PAR) {

base_sampling2/master/src/llama-graph.h

@@ -36,6 +36,7 @@ enum llm_ffn_op_type {
     LLM_FFN_RELU,
     LLM_FFN_RELU_SQR,
     LLM_FFN_SWIGLU,
+    LLM_FFN_GEGLU,
 };
 
 enum llm_ffn_gate_type {

base_sampling2/master/src/llama-model.cpp

@@ -13788,7 +13788,7 @@ uint64_t llama_model_size(const llama_model * model) {
 }
 
 const char * llama_model_chat_template(const llama_model * model, const char * name) {
-    const auto key = name ? LLM_KV(model->arch, name)(LLM_KV_TOKENIZER_CHAT_TEMPLATE_N)
+    const auto key = name ? LLM_KV(model->arch, name)(LLM_KV_TOKENIZER_CHAT_TEMPLATE)
         : LLM_KV(model->arch)(LLM_KV_TOKENIZER_CHAT_TEMPLATE);
     const auto & it = model->gguf_kv.find(key);
     if (it == model->gguf_kv.end()) {