biondizzle 8c31e78359 Fix cudagraph: fully fixed-layout per-expert sections, no GPU scalars in Python control flow ...

- Each expert gets max_chunks*128 rows at fixed offsets (e*max_chunks*128)
- Phase 1 scatters into fixed offsets with clamped local_row
- Phase 2 reads from fixed offsets (pure Python arithmetic, no GPU sync)
- padded_x_sf_buf sized for num_experts * max_chunks * 128
- padded_expert_offsets pre-computed in _allocate_buffers

2026-05-17 13:58:58 +00:00

..

patches

Fix warmup: pass local expert IDs (not global), remove incorrect _warmup_done guard

2026-05-17 11:38:19 +00:00

nvfp4_cutedsl.py

Fix cudagraph: fully fixed-layout per-expert sections, no GPU scalars in Python control flow

2026-05-17 13:58:58 +00:00