fix: apply fp8 weight scale factors when loading scaled safetensors

[fszontagh]

Summary

fp8 safetensors are dequantised to F16, but the companion per-tensor scale was never applied, so every fp8 weight loaded off by its scale factor. The scales are ~1e-3, so weights end up orders of magnitude too large, activations saturate and generation returns a pure white image. Nothing errors: the run reports 1/1 images saved.

Two spellings ship in the wild and both are now handled, for F8 weights only, so int8_tensorwise checkpoints keep their existing .weight_scale path:

• <module>.scale_weight (ComfyUI)
• <module>.weight_scale (diffusers)

.scale_input is an activation hint with no meaning once the weight is materialised, so it is dropped alongside the weight scale. scripts/convert_fp8_scale_to_bf16.py only understood .weight_scale and had the same blind spot, so it is updated to match.

Related Issue / Discussion

None.

Additional Information

Affected checkpoints found locally, both spellings:

checkpoint	scale key	paired
qwen_2.5_vl_7b_fp8_scaled.safetensors	.scale_weight	358
umt5_xxl_fp8_e4m3fn_scaled.safetensors (WAN encoder)	.scale_weight	168
qwen3vl_8b_fp8_scaled.safetensors	.weight_scale	252

Before/after on two pipelines, same command and seed, measured on the decoded output:

run	mean	std	colours
Qwen-Image-Edit, before	255.0	0.00	1
Qwen-Image-Edit, after	86.4	70.08	177315
WAN 2.1 VACE video, before	255.0	0.00	1
WAN 2.1 VACE video, after	117.4	90.16	11982

For Qwen-Image-Edit the fixed output differs from the same prompt run through the GGUF build of the same encoder by mean 1.43, i.e. the scaling is numerically right rather than merely non-blank.

SD 1.5 was re-run as a regression check and is unchanged. The change is inert on checkpoints that carry no scale tensors.

Checklist

• I have read and confirmed this PR follows the contribution guidelines.