[Bugfix][MiniMax-M3] Fence decode indexer shared-memory stage reuse (#60194)

Signed-off-by: Vladimir Anisimov <vanisimov@nvidia.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: OpenAI Codex <noreply@openai.com>
Co-authored-by: Thien Tran <gau.nernst@yahoo.com.sg>
This commit is contained in:
vovani
2026-10-07 01:35:12 +00:00
committed by GitHub
co-authored by Claude Opus 5.5 OpenAI Codex Thien Tran
parent 4b4eb6277a
commit fe51194e2d
2 changed files with 35 additions and 8 deletions
@@ -287,6 +287,8 @@ class IndexDecodeScoreKernel:
cute.arch.barrier(barrier_id=self.BAR_MMA, number_of_threads=128)
for q in cutlass.range_constexpr(Q_TILES):
cute.copy(ldsm_atom, sQ_ldsm[None, (q, None)], rQ[None, None, q])
# Order ldmatrix reads before the producer refills this stage with TMA.
cute.arch.fence_proxy("async.shared", space="cta")
cute.arch.mbarrier_arrive(tma_empty_mbar)
tma_stage = 1 % self.num_stages
@@ -342,6 +344,7 @@ class IndexDecodeScoreKernel:
rC[None, m, n],
)
cute.arch.fence_proxy("async.shared", space="cta")
cute.arch.mbarrier_arrive(tma_empty_mbar + tma_stage)
k_start = block_id * BLOCK_K + warp_id * 32