biondizzle acc20dffd7 CRITICAL FIX: don't fold input_scale into NVFP4 weight dequant ...

input_scale is the activation quantization scale (for FP8 inputs).
Since we use BF16 activations, the weight dequant is simply:
  lut[weight] * weight_scale * weight_scale_2

Folding input_scale in produced weights ~4000x too small,
causing all attention and FFN outputs to be effectively zero.

2026-05-31 22:03:55 +00:00

37 KiB

Raw Blame History

View Raw