nvfp4-megamoe-kernel

Files

biondizzle 84655d066a CUDA graph: Fix MoE bincount and per-call allocations (Hazard #4 )

1. Replace torch.bincount with scatter_add_ into pre-allocated buffer
   - bincount produces data-dependent shapes → breaks graph capture
   - scatter_add_ with pre-allocated _tokens_per_expert_buf (fixed shape)
   - Pre-allocated _ones_buf to avoid per-call torch.ones()

2. Replace torch.full for l1_gsa with pre-allocated buffer + fill_
   - torch.full allocates every call → breaks graph capture
   - Use self._l1_gsa_buf.fill_(l1_gs) instead

2026-06-03 17:37:03 +00:00

_archive

Cleanup Step 2: Archive Lineage P code, fix broken imports

2026-06-02 19:27:07 +00:00

cache

Cleanup Step 2: Archive Lineage P code, fix broken imports

2026-06-02 19:27:07 +00:00

decode

CUDA graph: Fix remaining sync violations from B200 detector run 2

2026-06-03 17:20:34 +00:00

kernels

Fix compressor: do not add positional bias to KV content