Qwen Img fp16 NaN issue

[realisticdreamer114514]

Custom Node Testing

• I have tried disabling custom nodes and the issue persists (see how to disable custom nodes if you need help)

Your question

Continuing #10668 but a new issue because I'll point out an obvious likely fix right now, need guidance from technically proficient people here

tldr; Qwen img models will produce NaN in latent, which is the cause of black preview/output at unet infer/vae decode stage across models, when running/cast to fp16 on some hardware. In qwen img models it is likely a unet issue. torch.clamp(65536) when --fp16-unet would be a start to fix this but at which point exactly?

Again, are there other possible fixes that doesn't involve editing the code base e.g. venv dependency/flags/node usage fixes?

Logs

Other

No response

[realisticdreamer114514]

Find a workaround (not a fix because it involves clamping & so precision loss):
in \comfy\ldm\qwen_image\model.py, class QwenImageTransformerBlock(nn.Module):, def forward(
add

        # Clip to prevent overflow for fp16
        if encoder_hidden_states.dtype == torch.float16:
            encoder_hidden_states = encoder_hidden_states.clip(-65504, 65504)
        if hidden_states.dtype == torch.float16:
            hidden_states = hidden_states.clip(-65504, 65504)

before return encoder_hidden_states, hidden_states to clamp values within fp16 range

Update: Workaround doesn't work even if it should (black imgs should be a NaN in latent problem)

[Laidawang]

i think its a problem caused by --force-fp16，My black problem fixed after I removed it.

[github-actions]

This issue is being marked stale because it has not had any activity for 30 days. Reply below within 7 days if your issue still isn't solved, and it will be left open. Otherwise, the issue will be closed automatically.