Merge pull request #1 from SCIR-HI/main

同步master

README.md

@@ -227,7 +227,9 @@ https://wandb.ai/thinksoso/llama_med/runs/a5wgcnzt/overview?workspace=user-think
 4. Q: 模型运行的结果不同、效果有限
    
    A: 由于生成模型生成多样性的考量，多次运行的结果可能会有差异。当前开源的模型由于LLaMA及Alpaca中文语料有限，且知识结合的方式较为粗糙，目前我们在进行相关改进研究，完成后欢迎大家的关注。
+5. Q: 模型无法运行/推理内容完全无法接受
    
+   A: 请确定已安装requirements中的依赖、配置好cuda环境并添加环境变量、正确输入下载好的模型以及lora的存储位置；推理内容如存在重复生成或部分错误内容属于llama-based模型的偶发现象，与llama模型的中文能力、训练数据规模以及超参设置均有一定的关系，未来我们会不断迭代缓解此问题。如存在严重问题，请将运行的文件名、模型名、lora等配置信息详细描述在issue中，谢谢大家。
 
 
 

finetune.py

@@ -25,6 +25,9 @@ from transformers import LlamaForCausalLM, LlamaTokenizer
 
 from utils.prompter import Prompter
 
+from transformers import Seq2SeqTrainer, TrainerCallback, TrainingArguments, TrainerState, TrainerControl
+from transformers.trainer_utils import PREFIX_CHECKPOINT_DIR
+
 
 def train(
     # model/data params
@@ -198,7 +201,7 @@ def train(
         if os.path.exists(checkpoint_name):
             print(f"Restarting from {checkpoint_name}")
             adapters_weights = torch.load(checkpoint_name)
-            model = set_peft_model_state_dict(model, adapters_weights)
+            set_peft_model_state_dict(model, adapters_weights)
         else:
             print(f"Checkpoint {checkpoint_name} not found")
 
@@ -223,6 +226,21 @@ def train(
         model.is_parallelizable = True
         model.model_parallel = True
 
+    class SavePeftModelCallback(TrainerCallback):
+        def on_save(
+            self,
+            args: TrainingArguments,
+            state: TrainerState,
+            control: TrainerControl,
+            **kwargs,
+        ):
+            checkpoint_folder = os.path.join(args.output_dir, f"{PREFIX_CHECKPOINT_DIR}-{state.global_step}")
+            kwargs["model"].save_pretrained(checkpoint_folder)
+            pytorch_model_path = os.path.join(checkpoint_folder, "pytorch_model.bin")
+            if os.path.exists(pytorch_model_path):
+                os.remove(pytorch_model_path)
+            return control
+
     trainer = transformers.Trainer(
         model=model,
         train_dataset=train_data,
@@ -251,16 +269,10 @@ def train(
         data_collator=transformers.DataCollatorForSeq2Seq(
             tokenizer, pad_to_multiple_of=8, return_tensors="pt", padding=True
         ),
+        callbacks=[SavePeftModelCallback],
     )
     model.config.use_cache = False
 
-    old_state_dict = model.state_dict
-    model.state_dict = (
-        lambda self, *_, **__: get_peft_model_state_dict(
-            self, old_state_dict()
-        )
-    ).__get__(model, type(model))
-
     if torch.__version__ >= "2" and sys.platform != "win32":
         model = torch.compile(model)
 

requirements.txt

@@ -1,12 +1,13 @@
-accelerate
-appdirs
-bitsandbytes
-black
-black[jupyter]
-datasets
-fire
-git+https://github.com/huggingface/peft.git
-git+https://github.com/huggingface/transformers.git
-gradio
-sentencepiece
+accelerate==0.20.1
+appdirs==1.4.4
+bitsandbytes==0.37.2
+black==23.3.0
+black[jupyter]==23.3.0
+datasets==2.11.0
+fire==0.5.0
+peft==0.3.0
+transformers==4.30.1
+gradio==3.33.1
+sentencepiece==0.1.97
+scipy==1.10.1
 wandb