多GPU运行Qwen2-VL模型触发CUDA断言错误的解决方法

问题场景 #

拥有4块GPU，运行Qwen2-VL-2B-Instruct模型时触发CUDA设备断言错误，核心报错点为inputs_embeds[image_mask] = image_embeds处索引越界。

使用代码：

model_name="Qwen/Qwen2-VL-2B-Instruct"
model = Qwen2VLForConditionalGeneration.from_pretrained(
          model_name, torch_dtype="auto", device_map="auto"
        )
model = nn.DataParallel(model)
processor = AutoProcessor.from_pretrained(model_name)

messages = [
        {
            "role": "user",
            "content": [
                {
                    "type": "image",
                    "image": file
                },
                {
                    "type": "text",
                    "text": """Describe the image"""
                }
            ]
        }
]
text = processor.apply_chat_template(
            messages, tokenize=False, add_generation_prompt=True)
image_inputs, video_inputs = process_vision_info(messages)
inputs = processor(
            text=[text],
            images=image_inputs,
            videos=video_inputs,
            padding=True,
            return_tensors="pt",
        )
with torch.no_grad():
    generated_ids = model.module.generate(**inputs, max_new_tokens=128)

错误详情 #

../aten/src/ATen/native/cuda/IndexKernel.cu:92: operator(): block: [35,0,0], thread: [31,0,0] Assertion `-sizes[i] <= index && index < sizes[i] && "index out of bounds"` failed.
ERROR:  CUDA error: device-side assert triggered
Compile with `TORCH_USE_CUDA_DSA` to enable device-side assertions.

Traceback (most recent call last):
  File "/home/ubuntu/projects/mistral-qaC/services/VisionService.py", line 104, in ask_vision
    generated_ids = self.model.module.generate(
                    ^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/ubuntu/projects/upper/lib/python3.12/site-packages/torch/utils/_contextlib.py", line 116, in decorate_context
    return func(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^
  File "/home/ubuntu/projects/upper/lib/python3.12/site-packages/transformers/generation/utils.py", line 2015, in generate
    result = self._sample(
             ^^^^^^^^^^^^^
  File "/home/ubuntu/projects/upper/lib/python3.12/site-packages/transformers/generation/utils.py", line 2965, in _sample
    outputs = self(**model_inputs, return_dict=True)
              ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/ubuntu/projects/upper/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1553, in _wrapped_call_impl
    return self._call_impl(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/ubuntu/projects/upper/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1562, in _call_impl
    return forward_call(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/ubuntu/projects/upper/lib/python3.12/site-packages/accelerate/hooks.py", line 169, in new_forward
    output = module._old_forward(*args, **kwargs)
             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/ubuntu/projects/upper/lib/python3.12/site-packages/transformers/models/qwen2_vl/modeling_qwen2_vl.py", line 1598, in forward
    inputs_embeds[image_mask] = image_embeds
    ~~~~~~~~~~~~~^^^^^^^^^^^^
RuntimeError: CUDA error: device-side assert triggered
Compile with `TORCH_USE_CUDA_DSA` to enable device-side assertions.

已尝试方案 #

• 使用CUDA_LAUNCH_BLOCKING=1 python script.py运行脚本，无效
• 确认模型和输入设备均为cuda:0
• 生成前执行torch.cuda.synchronize()和torch.cuda.empty_cache()

输入张量形状 #

• input_ids: torch.Size([1, 759])
• attention_mask: torch.Size([1, 759])
• pixel_values: torch.Size([2940, 1176])
• image_grid_thw: torch.Size([1, 3])

环境版本 #

transformers==4.45.0.dev0
torch==2.4.1+cu124

解决方法 #

1. 移除nn.DataParallel包装 #

transformers库的from_pretrained配合device_map="auto"已经自动实现多GPU模型分配，额外套nn.DataParallel会导致模型张量拆分逻辑冲突，引发索引越界。修改代码如下：

model_name="Qwen/Qwen2-VL-2B-Instruct"
model = Qwen2VLForConditionalGeneration.from_pretrained(
          model_name, torch_dtype="auto", device_map="auto"
        )
# 移除该行：model = nn.DataParallel(model)
processor = AutoProcessor.from_pretrained(model_name)

# ... 其余代码不变 ...

with torch.no_grad():
    # 直接调用model.generate，无需使用model.module.generate
    generated_ids = model.generate(**inputs, max_new_tokens=128)

2. 修正图像输入处理逻辑 #

自定义的process_vision_info函数可能返回格式错误的图像输入，导致pixel_values形状异常（当前缺少batch维度）。改为直接从messages中提取图像：

# 替换原有的image_inputs, video_inputs = process_vision_info(messages)
image_inputs = []
for msg in messages:
    for content_item in msg["content"]:
        if content_item["type"] == "image":
            image_inputs.append(content_item["image"])
video_inputs = []  # 无视频输入时设为空列表

3. 确保输入张量与模型设备对齐 #

将处理后的输入张量移至模型所在设备：

inputs = processor(
            text=[text],
            images=image_inputs,
            videos=video_inputs,
            padding=True,
            return_tensors="pt",
        )
# 添加该行代码
inputs = {k: v.to(model.device) for k, v in inputs.items()}

4. 验证pixel_values形状 #

处理后的pixel_values应符合模型要求的格式（通常为[batch_size, channels, height, width]），若仍异常，需检查图像输入是否为PIL.Image或numpy数组格式，避免传入扁平化的张量。

内容的提问来源于stack exchange，提问作者Cihan Yalçın