[Perf][MoE] Write FlashInfer combine into final output (#47156)

Signed-off-by: snordmann <snordmann@nvidia.com>
Co-authored-by: Codex <codex@openai.com>
This commit is contained in:
Samuel Nordmann
2026-07-16 17:21:03 +03:00
committed by GitHub
co-authored by Codex
parent 3e90d015ba
commit d08eebad16
3 changed files with 61 additions and 2 deletions
@@ -161,8 +161,8 @@ class FlashInferNVLinkOneSidedPrepareAndFinalize(mk.FusedMoEPrepareAndFinalizeMo
ep_size, self.runtime_max_tokens_per_rank, hidden_size
)
combined_output = self.all2all_manager.moe_alltoall.combine( # type: ignore[attr-defined]
self.all2all_manager.combine_into( # type: ignore[attr-defined]
payload=fused_expert_output,
runtime_max_tokens_per_rank=self.runtime_max_tokens_per_rank,
output=output,
)
output.copy_(combined_output)