fix bugs (#2954)

docs/source/Instruction/命令行参数.md

@@ -68,6 +68,7 @@
 - 🔥stream: 流式输出，默认为`False`
 - stop_words: 额外的停止词，默认为`[]`
 - logprobs: 是否输出logprobs，默认为False
+- top_logprobs: 默认为None
 
 ### 量化参数
 以下为拉起模型时量化的参数，具体含义可以查看[量化](https://huggingface.co/docs/transformers/main/en/main_classes/quantization)文档。这里不包含`swift export`中涉及的`gptq`、`awq`量化参数

docs/source_en/Instruction/Command-line-parameters.md

@@ -69,6 +69,7 @@ Refer to the [generation_config](https://huggingface.co/docs/transformers/main_c
 - 🔥stream: Stream output, default is `False`.
 - stop_words: Additional stop words, default is `[]`.
 - logprobs: Whether to output logprobs, default is False.
+- top_logprobs: Default is `None`.
 
 ### Quantization Arguments
 

examples/custom/sft.sh

@@ -21,4 +21,5 @@ swift sft \
     --warmup_ratio 0.05 \
     --dataloader_num_workers 4 \
     --max_length 2048 \
-    --output_dir output
+    --output_dir output \
+    --dataset_num_proc 4

examples/train/rlhf/cpo.sh

@@ -24,4 +24,5 @@ swift rlhf \
     --output_dir output \
     --warmup_ratio 0.05 \
     --dataloader_num_workers 4 \
-    --deepspeed zero2
+    --deepspeed zero2 \
+    --dataset_num_proc 4

examples/train/rlhf/dpo.sh

@@ -24,4 +24,5 @@ swift rlhf \
     --output_dir output \
     --warmup_ratio 0.05 \
     --dataloader_num_workers 4 \
-    --deepspeed zero2
+    --deepspeed zero2 \
+    --dataset_num_proc 4

examples/train/rlhf/kto.sh

@@ -23,4 +23,5 @@ swift rlhf \
     --output_dir output \
     --warmup_ratio 0.05 \
     --dataloader_num_workers 4 \
-    --deepspeed zero2
+    --deepspeed zero2 \
+    --dataset_num_proc 4

examples/train/rlhf/orpo.sh

@@ -24,4 +24,5 @@ swift rlhf \
     --output_dir output \
     --warmup_ratio 0.05 \
     --dataloader_num_workers 4 \
-    --deepspeed zero2
+    --deepspeed zero2 \
+    --dataset_num_proc 4

examples/train/rlhf/ppo.sh

@@ -27,4 +27,5 @@ swift rlhf \
     --warmup_ratio 0.05 \
     --dataloader_num_workers 4 \
     --deepspeed zero2 \
-    --response_length 512
+    --response_length 512 \
+    --dataset_num_proc 4

examples/train/rlhf/rm.sh

@@ -24,4 +24,5 @@ swift rlhf \
     --output_dir output \
     --warmup_ratio 0.05 \
     --dataloader_num_workers 4 \
-    --deepspeed zero2
+    --deepspeed zero2 \
+    --dataset_num_proc 4

examples/train/rlhf/simpo.sh

@@ -22,4 +22,5 @@ swift rlhf \
     --output_dir output \
     --warmup_ratio 0.05 \
     --dataloader_num_workers 4 \
-    --deepspeed zero2
+    --deepspeed zero2 \
+    --dataset_num_proc 4

swift/llm/argument/base_args/generation_args.py

@@ -35,6 +35,7 @@ class GenerationArguments:
     stream: bool = False
     stop_words: List[str] = field(default_factory=list)
     logprobs: bool = False
+    top_logprobs: Optional[int] = None
 
     def get_request_config(self):
         if getattr(self, 'task_type') != 'causal_lm':
@@ -50,4 +51,5 @@ def get_request_config(self):
             stop=self.stop_words,
             stream=self.stream,
             repetition_penalty=self.repetition_penalty,
-            logprobs=self.logprobs)
+            logprobs=self.logprobs,
+            top_logprobs=self.top_logprobs)

swift/llm/infer/infer_engine/lmdeploy_engine.py

@@ -181,7 +181,9 @@ def _prepare_generation_config(self, request_config: RequestConfig) -> LmdeployG
             if request_config.top_logprobs is not None:
                 kwargs['logprobs'] = max(1, request_config.top_logprobs)
 
-        return LmdeployGenerationConfig(**kwargs)
+        res = LmdeployGenerationConfig(**kwargs)
+        res.top_logprobs = request_config.top_logprobs
+        return res
 
     async def _infer_stream_async(
             self, template: Template, inputs: Dict[str, Any],
@@ -204,7 +206,8 @@ async def _infer_stream_async(
                 if not delta_text and not is_finished:
                     continue
 
-                logprobs = self._get_logprobs(output.logprobs, output.token_ids[token_idx:], generation_config.logprobs)
+                logprobs = self._get_logprobs(output.logprobs, output.token_ids[token_idx:],
+                                              generation_config.top_logprobs)
                 token_idx = len(output.token_ids)
 
                 usage_info = self._get_usage_info(len(inputs['input_ids']), output.num_token)
@@ -233,7 +236,7 @@ async def _infer_full_async(self, template: Template, inputs: Dict[str, Any],
                 pass
 
         response = template.decode(output.token_ids)
-        logprobs = self._get_logprobs(output.logprobs, output.token_ids, generation_config.logprobs)
+        logprobs = self._get_logprobs(output.logprobs, output.token_ids, generation_config.top_logprobs)
 
         usage_info = self._get_usage_info(len(inputs['input_ids']), output.num_token)
         toolcall = self._get_toolcall(response, template.tools_prompt)

swift/llm/infer/infer_engine/vllm_engine.py

@@ -219,8 +219,10 @@ def _get_logprobs(self,
                       logprobs_list: Optional[List[Dict[int, float]]],
                       token_ids: List[int],
                       top_logprobs: Optional[int] = None) -> Optional[Dict[str, Any]]:
-        if logprobs_list is None:
+        if logprobs_list is None or len(token_ids) == 0:
             return None
+        if len(token_ids) > 0:
+            logprobs_list = logprobs_list[-len(token_ids):]
         for logprobs in logprobs_list:
             for token_id, logprob in logprobs.items():
                 logprobs[token_id] = logprob.logprob
@@ -244,7 +246,9 @@ def _prepare_generation_config(self, request_config: RequestConfig) -> SamplingP
         for key in ['n', 'best_of', 'frequency_penalty', 'presence_penalty', 'seed']:
             kwargs[key] = getattr(request_config, key)
 
-        return SamplingParams(**kwargs)
+        res = SamplingParams(**kwargs)
+        res.top_logprobs = request_config.top_logprobs
+        return res
 
     async def _infer_stream_async(self, template: Template, inputs: Dict[str, Any], generation_config: SamplingParams,
                                   **kwargs) -> AsyncIterator[ChatCompletionStreamResponse]:
@@ -271,7 +275,7 @@ async def _infer_stream_async(self, template: Template, inputs: Dict[str, Any],
             choices = []
             for output in result.outputs:
                 logprobs = self._get_logprobs(output.logprobs, output.token_ids[token_idxs[output.index]:],
-                                              generation_config.logprobs)
+                                              generation_config.top_logprobs)
                 token_idxs[output.index] = len(output.token_ids)
                 toolcall = None
                 if output.is_finished:
@@ -301,7 +305,7 @@ async def _infer_full_async(self,
         for output in result.outputs:
             output.token_ids = list(output.token_ids)
             response = template.decode(output.token_ids)
-            logprobs = self._get_logprobs(output.logprobs, output.token_ids, generation_config.logprobs)
+            logprobs = self._get_logprobs(output.logprobs, output.token_ids, generation_config.top_logprobs)
             toolcall = self._get_toolcall(response, template.tools_prompt)
             choice = ChatCompletionResponseChoice(
                 index=output.index,

swift/llm/model/patcher.py

@@ -210,3 +210,20 @@ def __new_init__(self, *args, **kwargs):
         yield
     finally:
         PreTrainedModel.from_pretrained = from_pretrained
+
+
+@contextmanager
+def patch_automodel_for_awq():
+    from_pretrained = PreTrainedModel.from_pretrained
+
+    @classmethod
+    def _new_from_pretrained(cls, *args, **kwargs):
+        kwargs.pop('use_cache', None)
+        return from_pretrained.__func__(cls, *args, **kwargs)
+
+    PreTrainedModel.from_pretrained = _new_from_pretrained
+
+    try:
+        yield
+    finally:
+        PreTrainedModel.from_pretrained = from_pretrained

swift/llm/model/register.py

@@ -19,7 +19,7 @@
 
 from swift.utils import get_dist_setting, get_logger, is_mp, is_unsloth_available, patch_getattr, use_torchacc
 from .constant import ModelType
-from .patcher import patch_automodel_for_sequence_classification
+from .patcher import patch_automodel_for_awq, patch_automodel_for_sequence_classification
 from .utils import AttnImpl, HfConfigFactory, ModelInfo, safe_snapshot_download
 
 GetModelTokenizerFunction = Callable[..., Tuple[Optional[PreTrainedModel], PreTrainedTokenizerBase]]
@@ -200,13 +200,15 @@ def get_model_tokenizer_from_local(model_dir: str,
             except ValueError:
                 model = None
 
+        automodel_class = automodel_class or AutoModelForCausalLM
         if model is None:
             if model_info.task_type == 'seq_cls':
                 context = partial(patch_automodel_for_sequence_classification, model_meta=kwargs['model_meta'])
+            elif 'AutoAWQFor' in automodel_class.__name__:
+                context = patch_automodel_for_awq
             else:
                 context = nullcontext
             with context():
-                automodel_class = automodel_class or AutoModelForCausalLM
                 model = automodel_class.from_pretrained(
                     model_dir, config=model_config, torch_dtype=torch_dtype, trust_remote_code=True, **model_kwargs)