llamafactory-cliapi 推理Qwen3-VL-32B-Instruct报错：openai.BadRequestError: Error code: 400 - {'detail': 'Invalid or inaccessible file path.'}

[fuqiang-benz]

Reminder

• I have read the above rules and searched the existing issues.

System Info

环境：使用llamafactory:npu-a2 docker镜像启动
版本：
llamafactory ：0.9.4.dev0
torch : 2.7.1+cpu
torch_npu : 2.7.1
openai : 1.5.0

启动命令：
llamafactory-cli api examples/inference/qwen3_vl.yaml

qwen3_vl.yaml:
model_name_or_path: /home/LLM/llm_models/Qwen3-VL-32B-Instruct
template: qwen3_vl

请求脚本：
def main(): client = OpenAI( api_key="{}".format(os.getenv("API_KEY", "0")), base_url="http://183.242.165.15:{}/v1".format(os.getenv("API_PORT", 11245)), ) messages = [] messages.append( { "role": "user", "content": [ {"type": "text", "text": "Output the color and number of each box."}, { "type": "image_url", "image_url": {"url": "/home/LLM/llm_models/qwen2_5_VL_32B_images/1.png"}, }, ], } ) result = client.chat.completions.create(messages=messages, model="test") messages.append(result.choices[0].message) print("Round 1:", result.choices[0].message.content)

发现奇怪现象：
image_url 传入链接时可以正常识别
image_url 传入本地路径报错：openai.BadRequestError: Error code: 400 - {'detail': 'Invalid or inaccessible file path.'}

Reproduction

Put your message here.

Others

No response