biondizzle 152648789d fix: use checkpoint input_scale for activation global scale (not hardcoded 1/2688) ...

The checkpoint stores input_scale per projection — the pre-computed
activation normalization factor. Using 1/2688 was wrong for most layers
(e.g. down_proj input_scale=0.031 vs 1/2688=0.000372 — 83x off).
This caused under-quantized activations and garbage output.

2026-05-16 21:46:00 +00:00

..

patches

fix: use checkpoint input_scale for activation global scale (not hardcoded 1/2688)

2026-05-16 21:46:00 +00:00

nvfp4_cutedsl.py

fix: use checkpoint input_scale for activation global scale (not hardcoded 1/2688)

2026-05-16 21:46:00 +00:00