swift微调Qwen2.5-VL，合并模型后，分别使用swift和transform推理结果不一致 #4009
Description
使用swift微调多模态大模型，合并模型后，分别使用swift和transform推理结果不一致，想问下原因是什么
min和max pixels、超参数设置都一样；swift prompt包含了，transform prompt未包含
- swift代码
model = 'checkpoint-540-merged'
model, tokenizer = get_model_tokenizer(model, torch.bfloat16, attn_impl='flash_attn')
template_type = None
template_type = template_type or model.model_meta.template
template = get_template(template_type, tokenizer, default_system=system_prompt)
engine = PtEngine.from_model_template(model, template, max_batch_size=1)
request_config = RequestConfig(max_tokens=1024, temperature=0.1, top_p=0.001, top_k=1,repetition_penalty=1.05)
infer_requests = [
InferRequest(messages=[{'role': 'user', 'content': prompt_use}],
images=data["images"]),
]
resp_list = engine.infer(infer_requests, request_config)
- transformer代码
model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
self.qwen_model_path,
torch_dtype=torch.bfloat16,
attn_implementation="flash_attention_2",
device_map=self.DEVICE
)
min_pixels = 256 * 28 * 28
max_pixels = 1280 * 28 * 28
processor = AutoProcessor.from_pretrained(model_path, min_pixels=min_pixels, max_pixels=max_pixels)
messages = [
{"role": "system", "content": system_prompt},
{
"role": "user",
"content": content,
}
]
# Preparation for inference
text = processor.apply_chat_template(
messages, tokenize=False, add_generation_prompt=True, add_vision_id=True
)
image_inputs, video_inputs = process_vision_info(messages)
inputs = processor(
text=[text],
images=image_inputs,
videos=video_inputs,
padding=True,
return_tensors="pt",
)
inputs = inputs.to(model.device)
generated_ids = model.generate(**inputs, max_new_tokens=2048, temperature=0.1, top_p=0.001, top_k=1, repetition_penalty=1.05)
generated_ids_trimmed = [
out_ids[len(in_ids) :] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
]
output_text = processor.batch_decode(
generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=True
)