biondizzle 3cd910193c Rewrite scale assembly: no .item() calls, no Python loops, fully GPU ...

Apply to_blocked swizzle on entire padded buffer at once instead of
per-expert loops. No .item()/.cpu() calls. Fully cudagraph-safe.

2026-05-17 09:59:12 +00:00

..

patches

Fix OOB in scale assembly: size padded_x_sf for max tokens, fix top_k/max_num_tokens passing, support variable-size expert blocks

2026-05-17 09:56:28 +00:00

nvfp4_cutedsl.py

Rewrite scale assembly: no .item() calls, no Python loops, fully GPU

2026-05-17 09:59:12 +00:00