biondizzle b531a98f8f Fix scale assembly: per-expert 128-row fixed slots, no dynamic sizing ...

- Reverted from full-buffer swizzle to per-expert 128-row slots
- Scatter into e*128 fixed positions (cudagraph-compatible, fixed shape)
- Clamp local_row to 127 for experts with >128 tokens (GEMM uses expert_offsets)
- Buffer sized for num_experts*128 rows (not max_tokens*top_k)
- Add _warmup_done guard to only run warmup once (not 60x)

2026-05-17 11:10:59 +00:00

..

patches

Fix scale assembly: per-expert 128-row fixed slots, no dynamic sizing

2026-05-17 11:10:59 +00:00

nvfp4_cutedsl.py

Fix scale assembly: per-expert 128-row fixed slots, no dynamic sizing

2026-05-17 11:10:59 +00:00