lingshu-medical-mllm

/

Lingshu-7B

like 74

Follow

Lingshu: MLLMs for Unified Multimodal Medical Understanding and Reasoning 126

AttributeError: 'Qwen2_5_VLVisionAttention' object has no attribute 'is_causal'

can you share your code that raise this error and the version of your transformers lib?

# Load processor and model
    processor = AutoProcessor.from_pretrained(model_name)
    model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
        model_name,
        torch_dtype=torch.bfloat16,
        attn_implementation="flash_attention_2",
        device_map="auto",
    )
    device = next(model.parameters()).device

    # Load input data
    with open(input_json_path, 'r') as f:
        data = json.load(f)

    results = []
    for entry in data:
        img_path = entry.get("img_path")
        prompt_text = entry.get("prompt")

        # Prepare inference messages
        messages = [
            {
                "role": "user",
                "content": [
                    {"type": "image", "image": img_path},
                    {"type": "text", "text": prompt_text}
                ],
            }
        ]

        # Apply chat template
        text = processor.apply_chat_template(
            messages, tokenize=False, add_generation_prompt=True
        )

        # Process vision inputs
        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(device)

        # Run generation
        generated_ids = model.generate(**inputs, max_new_tokens=128)
        generated_sequences = [out_ids[len(in_ids):] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)]
        output_texts = processor.batch_decode(
            generated_sequences,
            skip_special_tokens=True,
            clean_up_tokenization_spaces=False,
        )

        # Attach response
        response = output_texts[0].strip()
        result_entry = entry.copy()
        result_entry["model_response"] = response
        results.append(result_entry)

python -c "import transformers; print(transformers.version)"

4.53.0

# Load processor and model
    processor = AutoProcessor.from_pretrained(model_name)
    model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
        model_name,
        torch_dtype=torch.bfloat16,
        attn_implementation="flash_attention_2",
        device_map="auto",
    )
    device = next(model.parameters()).device

    # Load input data
    with open(input_json_path, 'r') as f:
        data = json.load(f)

    results = []
    for entry in data:
        img_path = entry.get("img_path")
        prompt_text = entry.get("prompt")

        # Prepare inference messages
        messages = [
            {
                "role": "user",
                "content": [
                    {"type": "image", "image": img_path},
                    {"type": "text", "text": prompt_text}
                ],
            }
        ]

        # Apply chat template
        text = processor.apply_chat_template(
            messages, tokenize=False, add_generation_prompt=True
        )

        # Process vision inputs
        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(device)

        # Run generation
        generated_ids = model.generate(**inputs, max_new_tokens=128)
        generated_sequences = [out_ids[len(in_ids):] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)]
        output_texts = processor.batch_decode(
            generated_sequences,
            skip_special_tokens=True,
            clean_up_tokenization_spaces=False,
        )

        # Attach response
        response = output_texts[0].strip()
        result_entry = entry.copy()
        result_entry["model_response"] = response
        results.append(result_entry)

python -c "import transformers; print(transformers.version)"

4.53.0

We tried this transformers version and reported the same error. So the reason is that the latest transformers is not compatible with our model. You may need to use transformers<=4.52.1
Thanks

# Load processor and model
    processor = AutoProcessor.from_pretrained(model_name)
    model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
        model_name,
        torch_dtype=torch.bfloat16,
        attn_implementation="flash_attention_2",
        device_map="auto",
    )
    device = next(model.parameters()).device

    # Load input data
    with open(input_json_path, 'r') as f:
        data = json.load(f)

    results = []
    for entry in data:
        img_path = entry.get("img_path")
        prompt_text = entry.get("prompt")

        # Prepare inference messages
        messages = [
            {
                "role": "user",
                "content": [
                    {"type": "image", "image": img_path},
                    {"type": "text", "text": prompt_text}
                ],
            }
        ]

        # Apply chat template
        text = processor.apply_chat_template(
            messages, tokenize=False, add_generation_prompt=True
        )

        # Process vision inputs
        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(device)

        # Run generation
        generated_ids = model.generate(**inputs, max_new_tokens=128)
        generated_sequences = [out_ids[len(in_ids):] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)]
        output_texts = processor.batch_decode(
            generated_sequences,
            skip_special_tokens=True,
            clean_up_tokenization_spaces=False,
        )

        # Attach response
        response = output_texts[0].strip()
        result_entry = entry.copy()
        result_entry["model_response"] = response
        results.append(result_entry)

python -c "import transformers; print(transformers.version)"

4.53.0

https://github.com/huggingface/transformers/issues/39095
It should be a transformers bug, which also faced in original qwen2_5_vl models.