load qwen_image_edit_2509 error

[h123fire]

Checklist

• 1. I have searched for related issues and FAQs (https://nunchaku.tech/docs/nunchaku/faq/faq.html) but was unable to find a solution.
• 2. The issue persists in the latest version.
• 3. Please note that without environment information and a minimal reproducible example, it will be difficult for us to reproduce and address the issue, which may delay our response.
• 4. If your report is a question rather than a bug, please submit it as a discussion at https://github.com/mit-han-lab/nunchaku/discussions/new/choose. Otherwise, this issue will be closed.
• 5. If this is related to ComfyUI, please report it at https://github.com/mit-han-lab/ComfyUI-nunchaku/issues.
• 6. I will do my best to describe the issue in English.

Describe the Bug

[Bug] transformer = NunchakuQwenImageTransformer2DModel.from_pretrained( f"nunchaku-tech/nunchaku-qwen-image-edit-2509/svdq-{get_precision()}_r{rank}-qwen-image-edit-2509.safetensors" )
assert state_dict[k].dtype == model_state_dict[k].dtype
AssertionError

Environment

pytorch 2.5

Reproduction Steps

no

[lmxyy]

what GPU are you using?

[kli017]

same error, any solution?

[stime111]

what GPU are you using?

Same qusetion on 3090 windows platform after i update cuda version from 12.1 to 12.8 , I find it cause by the mismatch of .wscales layers(which is torch.bfloat16 in transformer and torch.float8_e4m3fn in safetensor).
Also i find the shape mismatch between .wscales in transformer and safetensor(torch.Size([48, 9216]) in transformer and torch.Size([192, 9216]) in r32_4step_v2 safetensor, layer transformer_blocks.0.attn.to_qkv.wscales).
Is there any problem during model loading?

[eriksonferreira]

Same problem here, did you guys found any solution? I'm using a A6000 48gb

[fengwk]

I seem to have a similar problem and expect it to be solved

[01-111162-173]

seem question

[vnull777-ops]

same problem with 5060ti 16G

[github-actions]

This issue has been automatically closed due to 30-day inactivity. Please feel free to reopen it with /reopen if needed.