support llama3.

Author: wangzhaode

include/llm.hpp

@@ -313,6 +313,19 @@ class Yi_6b : public Llama2_7b {
     virtual std::vector<int> tokenizer(const std::string& query) override;
     virtual bool is_stop(int token_id) override;
 };
+
+class Llama3_8b : public Llama2_7b {
+public:
+    Llama3_8b() {
+        model_name_ = "Llama3_8b";
+        layer_nums_ = 32;
+        key_value_shape_ = {2, 1, 8, 0, 128};
+        hidden_size_ = 4096;
+    }
+private:
+    virtual std::vector<int> tokenizer(const std::string& query) override;
+    virtual bool is_stop(int token_id) override;
+};
 // Llm end
 
 // Embedding start

src/llm.cpp

@@ -86,6 +86,9 @@ Llm* Llm::createLLM(const std::string& path, std::string model_type) {
     } else if (model_type.find("yi") != std::string::npos) {
         llm = new Yi_6b;
         llm->model_name_ = "Yi_6b";
+    } else if (model_type.find("llama3") != std::string::npos) {
+        llm = new Llama3_8b;
+        llm->model_name_ = "Llama3_8b";
     }
     if (!llm) {
         std::cerr << "model type can't judge!" << std::endl;
@@ -796,6 +799,18 @@ std::vector<int> Yi_6b::tokenizer(const std::string& query) {
 bool Yi_6b::is_stop(int token_id) {
     return token_id == 7 || token_id == 64001;
 }
+
+std::vector<int> Llama3_8b::tokenizer(const std::string& query) {
+    // <|begin_of_text|><|start_header_id|>user<|end_header_id|>\n\n+query+<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n
+    auto ids = tokenizer_encode(query);
+    ids.insert(ids.begin(), {128000, 128006, 882, 128007, 271});
+    ids.insert(ids.end(), {128009, 128006, 78191, 128007, 271});
+    return ids;
+}
+
+bool Llama3_8b::is_stop(int token_id) {
+    return token_id == 128001 || token_id == 128009;
+}
 // Llm end
 
 // Embedding start