From c36434a921abf41e98adbd3f913ff8daa07ae148 Mon Sep 17 00:00:00 2001
From: zhangkaihuo <zhangkaihuo@gmail.com>
Date: Thu, 19 Sep 2024 15:36:16 +0800
Subject: [PATCH] only add skip layers

---
 examples/llava/minicpmv-cli.cpp |  1 +
 llama.cpp                       | 17 ++++++++++-------
 llama.h                         |  1 +
 3 files changed, 12 insertions(+), 7 deletions(-)

diff --git a/examples/llava/minicpmv-cli.cpp b/examples/llava/minicpmv-cli.cpp
index f8aeba563f50c..e58b264324215 100644
--- a/examples/llava/minicpmv-cli.cpp
+++ b/examples/llava/minicpmv-cli.cpp
@@ -81,6 +81,7 @@ static struct llava_context * llava_init_context(gpt_params * params) {
         //load last model
         llama_model_params model_params = llama_model_params_from_gpt_params(*params);
         model_params.init_time = false;
+        model_params.has_vocab = false;
         //llama_model * model2 = llama_load_model_from_file(params->model.c_str(), model_params);
         //llama_model * model2 = llama_load_model_from_file("/Users/zkh/Downloads/last_16/ggml-model-Q4_0.gguf", model_params);
         model2 = llama_load_model_from_file(params->skip_model.c_str(), model_params);
diff --git a/llama.cpp b/llama.cpp
index 62d9efd6a13cd..ec36da68e12c0 100644
--- a/llama.cpp
+++ b/llama.cpp
@@ -4123,7 +4123,8 @@ static llama_token llama_byte_to_token(const llama_vocab & vocab, uint8_t ch);
 
 static void llm_load_vocab(
         llama_model_loader & ml,
-        llama_model & model) {
+        llama_model & model,
+        bool has_vocab) {
     auto & vocab = model.vocab;
 
     struct gguf_context * ctx = ml.meta;
@@ -4296,10 +4297,11 @@ static void llm_load_vocab(
     } else if (vocab.type == LLAMA_VOCAB_TYPE_WPM) {
         vocab.linefeed_id = vocab.special_pad_id;
     } else {
-        printf("====unknow vocab type\n");
-        // const std::vector<int> ids = llama_tokenize_internal(vocab, "\xC4\x8A", false); // U+010A
-        // GGML_ASSERT(!ids.empty() && "model vocab missing newline token");
-        // vocab.linefeed_id = ids[0];
+        if(has_vocab){
+            const std::vector<int> ids = llama_tokenize_internal(vocab, "\xC4\x8A", false); // U+010A
+            GGML_ASSERT(!ids.empty() && "model vocab missing newline token");
+            vocab.linefeed_id = ids[0];
+        }
     }
 
     // special tokens
@@ -5930,12 +5932,12 @@ static int llama_model_load(const std::string & fname, llama_model & model, llam
             throw std::runtime_error("error loading model hyperparameters: " + std::string(e.what()));
         }
         try {
-            llm_load_vocab(ml, model);
+            llm_load_vocab(ml, model, params.has_vocab);
         } catch(const std::exception & e) {
             throw std::runtime_error("error loading model vocabulary: " + std::string(e.what()));
         }
 
-        if (model.skip_layers == 0){
+        if (params.has_vocab){
             llm_load_print_meta(ml, model);
         }
 
@@ -14861,6 +14863,7 @@ struct llama_model_params llama_model_default_params() {
         /*.use_mmap                    =*/ true,
         /*.use_mlock                   =*/ false,
         /*.init_time                   =*/ true,
+        /*.has_vocab                   =*/ true,
     };
 
 #ifdef GGML_USE_METAL
diff --git a/llama.h b/llama.h
index 545397b167acf..c21a00564c217 100644
--- a/llama.h
+++ b/llama.h
@@ -236,6 +236,7 @@ extern "C" {
         bool use_mmap;   // use mmap if possible
         bool use_mlock;  // force system to keep model in RAM
         bool init_time;
+        bool has_vocab;
     };
 
     struct llama_context_params {