Merge branch 'hiyouga:main' into main

requirements.txt

@@ -1,8 +1,8 @@
 torch>=1.13.1
 transformers>=4.29.1
 datasets>=2.12.0
-accelerate>=0.19.0
-peft>=0.3.0
+accelerate>=0.21.0
+peft>=0.4.0
 trl>=0.4.7
 sentencepiece
 jieba

src/llmtuner/__init__.py

@@ -1,4 +1,4 @@
 from llmtuner.chat import ChatModel
 
 
-__version__ = "0.1.2"
+__version__ = "0.1.3"

src/llmtuner/extras/save_and_load.py

@@ -1,6 +1,6 @@
 import os
 import torch
-from typing import Dict
+from typing import Dict, Optional
 
 from transformers.trainer import WEIGHTS_NAME, WEIGHTS_INDEX_NAME
 from transformers.modeling_utils import load_sharded_checkpoint
@@ -12,12 +12,12 @@ from llmtuner.extras.logging import get_logger
 logger = get_logger(__name__)
 
 
-def get_state_dict(model: torch.nn.Module) -> Dict[str, torch.Tensor]: # get state dict containing trainable parameters
+def get_state_dict(model: torch.nn.Module, trainable_only: Optional[bool] = True) -> Dict[str, torch.Tensor]:
     state_dict = model.state_dict()
     filtered_state_dict = {}
 
     for k, v in model.named_parameters():
-        if v.requires_grad:
+        if (not trainable_only) or v.requires_grad:
             filtered_state_dict[k] = state_dict[k].cpu().clone().detach()
 
     return filtered_state_dict

src/llmtuner/tuner/core/loader.py

@@ -27,8 +27,8 @@ logger = get_logger(__name__)
 
 check_min_version("4.29.1")
 require_version("datasets>=2.12.0", "To fix: pip install datasets>=2.12.0")
-require_version("accelerate>=0.19.0", "To fix: pip install accelerate>=0.19.0")
-require_version("peft>=0.3.0", "To fix: pip install peft>=0.3.0")
+require_version("accelerate>=0.21.0", "To fix: pip install accelerate>=0.21.0")
+require_version("peft>=0.4.0", "To fix: pip install peft>=0.4.0")
 require_version("trl>=0.4.7", "To fix: pip install trl>=0.4.7")
 
 
@@ -81,9 +81,6 @@ def load_model_and_tokenizer(
 
         elif model_args.quantization_bit == 4:
             require_version("bitsandbytes>=0.39.0", "To fix: pip install bitsandbytes>=0.39.0")
-            require_version("transformers>=4.30.1", "To fix: pip install transformers>=4.30.1")
-            require_version("accelerate>=0.20.3", "To fix: pip install accelerate>=0.20.3")
-            require_version("peft>=0.4.0.dev0", "To fix: pip install git+https://github.com/huggingface/peft.git")
             config_kwargs["load_in_4bit"] = True
             config_kwargs["quantization_config"] = BitsAndBytesConfig(
                 load_in_4bit=True,

src/llmtuner/tuner/core/trainer.py

@@ -56,7 +56,7 @@ class PeftTrainer(Seq2SeqTrainer):
             backbone_model.config.use_cache = True
             backbone_model.save_pretrained(
                 output_dir,
-                state_dict=get_state_dict(backbone_model),
+                state_dict=get_state_dict(backbone_model, trainable_only=(self.finetuning_args.finetuning_type != "full")),
                 safe_serialization=self.args.save_safetensors
             )
             backbone_model.config.use_cache = False

src/llmtuner/webui/chat.py

@@ -84,6 +84,12 @@ class WebChatModel(ChatModel):
             query, history, prefix, max_new_tokens=max_new_tokens, top_p=top_p, temperature=temperature
         ):
             response += new_text
+            response = self.postprocess(response)
             new_history = history + [(query, response)]
             chatbot[-1] = [query, response]
             yield chatbot, new_history
+
+    def postprocess(self, response: str) -> str:
+        response = response.replace("<", "&lt;")
+        response = response.replace(">", "&gt;")
+        return response