fix: 量化配置未传递给 LLaMA-Factory 导致退化为全精度加载

- 新增 QuantizationArgs 配置类，与 LLaMA-Factory QuantizationArguments 对齐
- 移除 load_in_4bit/load_in_8bit（由 quantization_bit 内部处理）
- 新增 quantization_method 支持多种量化后端（bnb, gptq, awq 等）
- 修复 dict_to_argv 跳过 None 值，避免 HfArgumentParser 报错
- 使用 exclude_none=True 防止空字段传递给 run_exp()
- 训练和推理时自动展平嵌套的量化配置

weclone/train/train_sft.py

@@ -40,7 +40,11 @@ def main():
     formatted_config = json.dumps(train_config.model_dump(mode="json"), indent=4, ensure_ascii=False)
     logger.info(f"Fine-tuning configuration:\n{formatted_config}")
 
-    run_exp(train_config.model_dump(mode="json"))
+    # Build config dict and remove nested 'quantization' key (its fields are already flattened at top level)
+    config_dict = train_config.model_dump(mode="json", exclude_none=True)
+    config_dict.pop("quantization", None)
+
+    run_exp(config_dict)
 
 
 if __name__ == "__main__":

weclone/utils/config.py

@@ -47,6 +47,17 @@ def load_base_config() -> WcConfig:
     return wc_config
 
 
+def _flatten_quantization_args(train_sft_args) -> dict:
+    """Extract quantization sub-config and return as a flat dict with non-None values.
+
+    LLaMA-Factory expects quantization params at the top level of the argument
+    namespace (e.g. ``--quantization_bit 4``), not nested under a ``quantization``
+    prefix. This helper flattens the nested ``QuantizationArgs`` model so it can
+    be merged directly into the config dict passed to HfArgumentParser.
+    """
+    return train_sft_args.quantization.get_non_none_dict()
+
+
 def create_config_by_arg_type(arg_type: str, wc_config: WcConfig) -> BaseModel:
     """Create corresponding configuration object based on argument type, merge common_config"""
     if arg_type == "cli_args":
@@ -55,7 +66,13 @@ def create_config_by_arg_type(arg_type: str, wc_config: WcConfig) -> BaseModel:
     common_config = wc_config.common_args.model_dump()
 
     if arg_type == "web_demo" or arg_type == "api_service":
-        config_dict = {**common_config, **wc_config.infer_args.model_dump()}
+        # Inherit quantization settings from train_sft_args for inference
+        quant_dict = _flatten_quantization_args(wc_config.train_sft_args)
+        config_dict = {
+            **common_config,
+            **wc_config.infer_args.model_dump(),
+            **quant_dict,
+        }
         return WCInferConfig(**config_dict)
 
     elif arg_type == "vllm":
@@ -66,12 +83,18 @@ def create_config_by_arg_type(arg_type: str, wc_config: WcConfig) -> BaseModel:
 
     elif arg_type == "train_sft":
         common_config["include_type"] = wc_config.make_dataset_args.include_type
-        config_dict = {**common_config, **wc_config.train_sft_args.model_dump()}
+
+        # Merge training params; flatten the nested quantization sub-config
+        train_dict = wc_config.train_sft_args.model_dump()
+        # Remove the nested "quantization" dict — its fields are flattened below
+        train_dict.pop("quantization", None)
+        train_dict.update(_flatten_quantization_args(wc_config.train_sft_args))
+
+        config_dict = {**common_config, **train_dict}
         return WCTrainSftConfig(**config_dict)
 
     elif arg_type == "make_dataset":
         make_dataset_config = wc_config.make_dataset_args.model_dump()
-        # TODO: Should the following three parameters be moved to common?
         train_sft_args = wc_config.train_sft_args
         extra_values = {
             "dataset": train_sft_args.dataset,

weclone/utils/config_models.py

@@ -1,5 +1,5 @@
 from enum import Enum
-from typing import TYPE_CHECKING, List, Optional
+from typing import TYPE_CHECKING, List, Literal, Optional
 
 from loguru import logger
 from pydantic import BaseModel, Field, model_validator
@@ -135,7 +135,7 @@ class MakeDatasetArgs(BaseConfigModel):
     max_image_num: int = Field(2, description="Maximum number of images per single data entry")
     blocked_words: List[str] = Field([], description="List of blocked words")
     add_time: bool = Field(False, description="Whether to add time to the dataset")
-    add_relation: bool = Field(False, description="Whether to add chat member relationship to the dataset")
+    add_relation: bool = Field(False, description="Whether to add chat relation to the dataset")
     single_combine_strategy: CombineStrategy = Field(
         CombineStrategy.TIME_WINDOW,
         description="Strategy for combining single person's messages into a single sentence",
@@ -163,6 +163,40 @@ class MakeDatasetArgs(BaseConfigModel):
     )
     clean_batch_size: int = Field(10, description="Batch size for data cleaning")
     vision_api: VisionApiConfig = Field(VisionApiConfig())
+    pure_text: bool = Field(False, description="Whether to use pure text mode (disable vision encoder)")
+
+
+class QuantizationArgs(BaseConfigModel):
+    """Quantization arguments aligned with LLaMA-Factory QuantizationArguments.
+
+    These parameters are passed directly to LLaMA-Factory's HfArgumentParser
+    for both training and inference. LLaMA-Factory internally maps
+    ``quantization_bit`` to ``BitsAndBytesConfig(load_in_4bit/load_in_8bit)``,
+    so there is no need to expose ``load_in_4bit`` / ``load_in_8bit`` directly.
+
+    Reference: LLaMA-Factory src/llamafactory/hparams/model_args.py QuantizationArguments
+    """
+
+    quantization_method: Optional[str] = Field(
+        None,
+        description="Quantization method: bnb, gptq, awq, aqlm, quanto, eetq, hqq, mxfp4, fp8",
+    )
+    quantization_bit: Optional[int] = Field(
+        None,
+        description="Number of bits for on-the-fly quantization (e.g. 4 or 8)",
+    )
+    quantization_type: Optional[Literal["nf4", "fp4"]] = Field(
+        None,
+        description="Quantization data type for bitsandbytes int4 training: nf4 or fp4",
+    )
+    double_quantization: Optional[bool] = Field(
+        None,
+        description="Whether to use double quantization in bitsandbytes int4 training",
+    )
+
+    def get_non_none_dict(self) -> dict:
+        """Return only the non-None fields as a dict, for merging into other configs."""
+        return {k: v for k, v in self.model_dump().items() if v is not None}
 
 
 class TrainSftArgs(BaseConfigModel):
@@ -190,6 +224,10 @@ class TrainSftArgs(BaseConfigModel):
     plot_loss: bool = Field(True, description="Whether to plot loss curve")
     fp16: bool = Field(True, description="Whether to use fp16")
     flash_attn: str = Field("fa2", description="Flash Attention type")
+    quantization: QuantizationArgs = Field(
+        default_factory=QuantizationArgs,
+        description="Quantization settings for on-the-fly quantization (QLoRA, etc.)",
+    )
     preprocessing_num_workers: int = Field(16, description="Number of preprocessing worker processes")
     dataloader_num_workers: int = Field(4, description="Number of dataloader worker processes")
     deepspeed: Optional[str] = Field(
@@ -207,6 +245,12 @@ class InferArgs(BaseConfigModel):
 
 class VllmArgs(BaseConfigModel):
     gpu_memory_utilization: float = Field(default=0.9, description="vllm GPU memory utilization")
+    quantization: Optional[str] = Field(
+        default=None, description="Quantization method for vLLM, e.g. 'awq', 'gptq'"
+    )
+    load_format: Optional[str] = Field(
+        default=None, description="Format for loading weights, e.g. 'awq', 'gptq'"
+    )
 
 
 class TestModelArgs(BaseConfigModel):
@@ -235,7 +279,7 @@ class WcConfig(BaseModel):
 
 
 class WCInferConfig(CommonArgs, InferArgs):
-    """Final configuration model for Web Demo"""
+    """Final configuration model for Web Demo / API Service (based on LLaMA-Factory ChatModel)"""
 
     pass
 

weclone/utils/tools.py

@@ -1,7 +1,8 @@
 def dict_to_argv(d):
     argv = []
     for k, v in d.items():
+        if v is None:
+            continue
         argv.append("--" + k)
-        if v is not None:
         argv.append(str(v))
     return argv