fix bugs (#2207)

Author: Jintao-Huang

swift/llm/deploy.py

@@ -275,7 +275,7 @@ async def inference_vllm_async(request: Union[ChatCompletionRequest, CompletionR
     request_id = request_info['request_id']
 
     kwargs = {'max_tokens': request.max_tokens}
-    for key in ['n', 'best_of', 'frequency_penalty', 'length_penalty', 'presence_penalty', 'num_beams']:
+    for key in ['n', 'best_of', 'frequency_penalty', 'length_penalty', 'presence_penalty']:
         kwargs[key] = getattr(request, key)
     for key in ['temperature', 'top_k', 'top_p', 'repetition_penalty']:
         new_value = getattr(request, key)
@@ -292,9 +292,6 @@ async def inference_vllm_async(request: Union[ChatCompletionRequest, CompletionR
             kwargs['logprobs'] = max(1, request.top_logprobs)
 
     generation_config = VllmGenerationConfig(**kwargs)
-    if generation_config.use_beam_search and request.stream:
-        error_msg = 'Streaming generation does not support beam search.'
-        raise ValueError(error_msg)
     tokenizer = template.tokenizer
     if tokenizer.eos_token is not None and tokenizer.eos_token not in generation_config.stop:
         generation_config.stop.append(tokenizer.eos_token)

swift/llm/export.py

@@ -255,18 +255,18 @@ def llm_export(args: ExportArguments) -> None:
         if args.quant_method == 'awq':
             from awq import AutoAWQForCausalLM
             model, template = prepare_model_template(
-                args, device_map=args.quant_device_map, verbose=False, automodel_class=AutoAWQForCausalLM)
+                args, device_map=args.quant_device_map, task='export', automodel_class=AutoAWQForCausalLM)
             awq_model_quantize(model, template.tokenizer, args.quant_batch_size)
             model.save_quantized(args.quant_output_dir)
         elif args.quant_method == 'gptq':
-            model, template = prepare_model_template(args, device_map=args.quant_device_map, verbose=False)
+            model, template = prepare_model_template(args, device_map=args.quant_device_map, task='export')
             gptq_quantizer = gptq_model_quantize(model, template.tokenizer, args.quant_batch_size)
             model.config.quantization_config.pop('dataset', None)
             gptq_quantizer.save(model, args.quant_output_dir)
         elif args.quant_method == 'bnb':
             args.quantization_bit = args.quant_bits
             args.bnb_4bit_compute_dtype, args.load_in_4bit, args.load_in_8bit = args.select_bnb()
-            model, template = prepare_model_template(args, device_map=args.quant_device_map, verbose=False)
+            model, template = prepare_model_template(args, device_map=args.quant_device_map, task='export')
             model.save_pretrained(args.quant_output_dir)
         else:
             raise ValueError(f'args.quant_method: {args.quant_method}')

swift/llm/infer.py

@@ -109,7 +109,7 @@ def merge_lora(args: InferArguments,
         if device_map is None:
             device_map = args.merge_device_map
         logger.info(f'merge_device_map: {device_map}')
-        model, template = prepare_model_template(args, device_map=device_map, verbose=False)
+        model, template = prepare_model_template(args, device_map=device_map, task='export')
         logger.info('Merge LoRA...')
         Swift.merge_and_unload(model)
         model = model.model
@@ -133,7 +133,7 @@ def merge_lora(args: InferArguments,
 def prepare_model_template(args: InferArguments,
                            *,
                            device_map: Optional[str] = None,
-                           verbose: bool = True,
+                           task: Literal['infer', 'export'] = 'infer',
                            automodel_class=None) -> Tuple[PreTrainedModel, Template]:
     from .sft import get_default_device_map
     if is_torch_npu_available():
@@ -188,25 +188,7 @@ def prepare_model_template(args: InferArguments,
         revision=args.model_revision,
         quant_method=args.quant_method,
         **kwargs)
-    if verbose:
-        logger.info(f'model_config: {model.config}')
-
-    generation_config = GenerationConfig(
-        max_new_tokens=args.max_new_tokens,
-        temperature=args.temperature,
-        top_k=args.top_k,
-        top_p=args.top_p,
-        do_sample=args.do_sample,
-        repetition_penalty=args.repetition_penalty,
-        num_beams=args.num_beams,
-        pad_token_id=tokenizer.pad_token_id,
-        eos_token_id=tokenizer.eos_token_id)
-    set_generation_config(model, generation_config)
-    logger.info(f'model.generation_config: {model.generation_config}')
 
-    if model.generation_config.num_beams != 1:
-        args.stream = False
-        logger.info('Setting args.stream: False')
     if model.max_model_len is None:
         model.max_model_len = args.max_model_len
     elif args.max_model_len is not None:
@@ -215,6 +197,26 @@ def prepare_model_template(args: InferArguments,
         else:
             raise ValueError('args.max_model_len exceeds the maximum max_model_len supported by the model.'
                              f'args.max_model_len: {args.max_model_len}, model.max_model_len: {model.max_model_len}')
+    if task == 'infer':
+        logger.info(f'model_config: {model.config}')
+        generation_config = GenerationConfig(
+            max_new_tokens=args.max_new_tokens,
+            temperature=args.temperature,
+            top_k=args.top_k,
+            top_p=args.top_p,
+            do_sample=args.do_sample,
+            repetition_penalty=args.repetition_penalty,
+            num_beams=args.num_beams,
+            pad_token_id=tokenizer.pad_token_id,
+            eos_token_id=tokenizer.eos_token_id)
+        model._generation_config_origin = model.generation_config
+        set_generation_config(model, generation_config)
+        logger.info(f'model.generation_config: {model.generation_config}')
+
+        if model.generation_config.num_beams != 1:
+            args.stream = False
+            logger.info('Setting args.stream: False')
+
     # Preparing LoRA
     if is_adapter(args.sft_type) and args.ckpt_dir is not None:
         if isinstance(args, DeployArguments) and args.lora_request_list is not None:
@@ -227,7 +229,7 @@ def prepare_model_template(args: InferArguments,
         model = model.to(model.dtype)
     model.requires_grad_(False)
 
-    if verbose:
+    if task == 'infer':
         show_layers(model)
         logger.info(model)
     logger.info(get_model_info(model))

swift/llm/utils/template.py

@@ -2028,6 +2028,10 @@ def _post_encode(self, model, data: Any) -> Dict[str, Any]:
             res['labels'] = labels[0]
         return res
 
+    @staticmethod
+    def _get_generate_ids(generate_ids: List[int], input_token_len: int) -> List[int]:
+        return generate_ids
+
 
 register_template(TemplateType.llama3_1_omni, Llama3_1OmniTemplate(), lazy_tokenize=True)
 
@@ -2642,7 +2646,7 @@ def _encode(self, example: Dict[str, Any]) -> Tuple[Dict[str, Any], Dict[str, An
         videos_path = example.get('videos') or []
         if len(videos_path) > 0:
             video_processor = self.tokenizer.processor.video_processor
-            video_inputs = video_processor(videos, return_tensors='pt').to(self.model.dtype)
+            video_inputs = video_processor(videos_path, return_tensors='pt').to(self.model.dtype)
             inputs['pixel_values_videos'] = video_inputs['pixel_values_videos']
         if len(images) > 0:
             image_processor = self.tokenizer.processor.image_processor

swift/llm/utils/vllm_utils.py

@@ -204,7 +204,6 @@ def __init__(
             top_k: int = 50,  # -1: all
             top_p: float = 1.,
             repetition_penalty: float = 1.,
-            num_beams: int = 1,
             *,
             n: int = 1,
             logprobs: Optional[int] = None,
@@ -218,12 +217,6 @@ def __init__(
             max_new_tokens = kwargs.pop('max_new_tokens', None)
             if max_new_tokens is not None:
                 max_tokens = max_new_tokens
-            if num_beams > 1:
-                top_k = -1
-                top_p = 1
-                temperature = 0
-                logger.warning('The output of num_beams in vllm may not be consistent with '
-                               'the output of num_beams in transformers.')
             if top_k == 0:
                 top_k = -1
             if stop is None:
@@ -233,11 +226,6 @@ def __init__(
             kwargs['top_k'] = top_k
             kwargs['top_p'] = top_p
             kwargs['repetition_penalty'] = repetition_penalty
-            if num_beams > 1:
-                best_of = kwargs.get('best_of')
-                assert 'use_beam_search' not in kwargs and best_of is None
-                kwargs['use_beam_search'] = True
-                kwargs['best_of'] = num_beams
             kwargs['n'] = n
             kwargs['logprobs'] = logprobs
             kwargs['seed'] = seed
@@ -260,7 +248,6 @@ class VllmGenerationConfig(_VllmGenerationConfigMixin, SamplingParams):
         top_k: int = 50  # -1: all
         top_p: float = 1.
         repetition_penalty: float = 1.
-        num_beams: int = 1
         n: int = 1
         logprobs: Optional[int] = None
         seed: Optional[int] = None
@@ -269,15 +256,6 @@ class VllmGenerationConfig(_VllmGenerationConfigMixin, SamplingParams):
         skip_special_tokens: bool = False
 
         def __post_init__(self):
-            if self.num_beams > 1:
-                self.top_k = -1
-                self.top_p = 1
-                self.temperature = 0
-                logger.warning('The output of num_beams in vllm may not be consistent with '
-                               'the output of num_beams in transformers.')
-                assert self.best_of is None
-                self.use_beam_search = True
-                self.best_of = self.num_beams
             if self.top_k == 0:
                 self.top_k = -1
             if self.stop is None:
@@ -435,10 +413,6 @@ def inference_stream_vllm(
         use_tqdm=use_tqdm,
         **kwargs)
 
-    if generation_config.use_beam_search:
-        error_msg = 'Streaming generation does not support beam search.'
-        raise ValueError(error_msg)
-
     n_finished = 0
     n_steps = 0
     if flush_steps is None: