初步支持QWEN1.5模型

src/model.cpp

@@ -103,6 +103,9 @@ namespace fastllm {
             model->model_type = "internlm";
         } else if (modelType == "llama") {
             model = (basellm*)(new LlamaModel());
+        } else if (modelType == "qwen2") {
+            model = new LlamaModel();
+            model->model_type = "qwen";
         } else if (modelType == "qwen") {
             model = (basellm *) (new QWenModel());
             model->weight.tokenizer.type = Tokenizer::TokenizerType::QWEN;

tools/fastllm_pytools/hf_model.py

@@ -58,6 +58,8 @@ def create(model,
         if modelInfo["chat_format"] == "chatml":
             modelInfo["im_end_id"] = tokenizer.im_end_id
             modelInfo["im_start_id"] = tokenizer.im_start_id
+    elif (modelInfo["model_type"] == "qwen2"):
+        modelInfo["eos_token_id"] = "151645"
     if (modelInfo["model_type"] == "chatglm" and hasattr(tokenizer, "build_chat_input")):
         # chatglm3
         modelInfo["pre_prompt"] = "";

tools/fastllm_pytools/torch2flm.py

@@ -115,6 +115,8 @@ def tofile(exportPath,
         if modelInfo["chat_format"] == "chatml":
             modelInfo["im_end_id"] = tokenizer.im_end_id
             modelInfo["im_start_id"] = tokenizer.im_start_id
+    elif (modelInfo["model_type"] == "qwen2"):
+        modelInfo["eos_token_id"] = "151645"
     if (modelInfo["model_type"] == "chatglm" and hasattr(tokenizer, "build_chat_input")):
         # chatglm3
         modelInfo["pre_prompt"] = "";

tools/scripts/qwen2_2flm.py

@@ -0,0 +1,14 @@
+import sys
+import torch
+from transformers import AutoTokenizer, AutoModelForCausalLM
+from fastllm_pytools import torch2flm
+
+if __name__ == "__main__":
+    model_name = sys.argv[3] if len(sys.argv) >= 4 else 'qwen/Qwen1.5-7B-Chat'
+    tokenizer = AutoTokenizer.from_pretrained(model_name)
+    model = AutoModelForCausalLM.from_pretrained(model_name, device_map="cpu", torch_dtype=torch.float16)
+    dtype = sys.argv[2] if len(sys.argv) >= 3 else "float16"
+    exportPath = sys.argv[1] if len(sys.argv) >= 2 else "qwen1.5-7b-' + dtype + '.flm"
+    # add custom code here
+    torch2flm.tofile(exportPath, model, tokenizer, pre_prompt="<|im_start|>system\nYou are a helpful assistant.<|im_end|>", user_role="<|im_start|>user\n",
+                     bot_role="<|im_end|><|im_start|>assistant\n", history_sep="<|im_end|>\n", dtype = dtype)