[Feature][NPU]: Under the same request, the quality of the raw image on the NPU differs from that on the NVIDIA GPU.

[dazedsanjin]

🚀 The feature, motivation and pitch

When deploying and verifying Qwen-Image-2512, I found that the quality of the raw images on the NPU（910B）was lower than that on the NVIDIA（A800）under the same request.

The request information is as follows:

response = requests.post(
    "http://localhost:8080/v1/images/generations",
    json={
        "prompt": "a black and white cat wearing a princess tiara",
        "size": "512x512",
        "num_inference_steps": 50,
        "seed": 42,
        "n": 1
    }
)

The image generated by A800 is as follows:

The image generated by 910B is as follows:

Is there a slight difference in the accuracy of the generated data?

Alternatives

No response

Additional context

No response

Before submitting a new issue...

• Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the documentation page, which can answer lots of frequently asked questions.

[hsliuustc0106]

@gcanlin WDYT

[gcanlin]

Could you try setting generator to CPU both on NPU and GPU? BTW, I think the current difference could be acceptable.

response = requests.post(
    "http://localhost:8080/v1/images/generations",
    json={
        "prompt": "a black and white cat wearing a princess tiara",
        "size": "512x512",
        "num_inference_steps": 50,
        "seed": 42,
        "generator_device": "cpu",
        "n": 1
    }
)

Update: It depends on #1183. You may need cherry-pick it to v0.14.0.

[dazedsanjin]

Replacing the generator_device field with cpu does indeed bring it closer to the NVIDIA GPU.
The image generated by A800 is as follows:

The image generated by 910B is as follows:

Could you please explain why there are differences when using GPUs? Is it mainly due to differences in the underlying operator implementation?

[gcanlin]

Could you please explain why there are differences when using GPUs? Is it mainly due to differences in the underlying operator implementation?

For the first difference, that's caused by the different behavior of torch.Generator between GPUs and NPUs.

For the second one, yes, that's caused by the underlying operator that can't be equivalent totally on different hardware, like conv3d.

[hsliuustc0106]

Could you please explain why there are differences when using GPUs? Is it mainly due to differences in the underlying operator implementation?

For the first difference, that's caused by the different behavior of torch.Generator between GPUs and NPUs.

For the second one, yes, that's caused by the underlying operator that can't be equivalent totally on different hardware, like conv3d.

I think you need add some explainations in #886