Fix float8_e4m3fnuz input scaling range

[maskyuanzh]

Summary

In diffsynth/core/vram/layers.py, AutoWrappedLinear.fp8_linear() currently initializes the FP8 input scaling threshold to 448 and divides it by 2 for torch.float8_e4m3fnuz:

fp8_max = 448.0
if self.computation_dtype == torch.float8_e4m3fnuz:
    fp8_max = fp8_max / 2.0

This produces a threshold of 224, while the maximum finite value defined for torch.float8_e4m3fnuz is 240.

This PR replaces the manually derived threshold with the maximum finite value defined by the selected dtype:

fp8_max = torch.finfo(self.computation_dtype).max

As a result:

• torch.float8_e4m3fn continues to use 448.
• torch.float8_e4m3fnuz now uses 240 instead of 224.

Scaling behavior

Let (M) be the maximum absolute value of an input row. For torch.float8_e4m3fnuz, the old and new scales are:

$$ s_{\mathrm{old}}(M) = \max\left(\frac{M}{224}, 1\right) $$

$$ s_{\mathrm{new}}(M) = \max\left(\frac{M}{240}, 1\right) $$

Therefore:

• For $M \leq 224$, the behavior is unchanged.
• For $224 < M \leq 240$, the new implementation avoids unnecessary scaling.
• For $M > 240$, the row is still scaled, but the full finite range of torch.float8_e4m3fnuz is used.

[gemini-code-assist]

Code Review

This pull request updates the FP8 linear layer implementation in diffsynth/core/vram/layers.py to dynamically determine the maximum representable value for the selected FP8 data type using torch.finfo(self.computation_dtype).max, replacing the previous hardcoded values and conditional logic. No review comments were provided, and there is no additional feedback.

Important

The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.