[TRTLLM-7318][feat] MnnvlThroughput AlltoAll implementation.#7499

Merged

bobboli merged 62 commits intoNVIDIA:mainfrom

bobboli:alltoall

Oct 27, 2025

Commits on Sep 22, 2025

Commits on Sep 24, 2025

Use smem for topk_target_ranks and topk_send_indices.
bobboli
committed

Commits on Oct 7, 2025

DISABLE_SYNC_FOR_PROFILING includes st of flags.
bobboli
committed

Commits on Oct 8, 2025

Commits on Oct 9, 2025

Dispatch and combine should use different completion flags, otherwise there is racing.
bobboli
committed
Support the case when combine payload is in workspace.
bobboli
committed
If prepareCombineKernel does not do copy, only one block is enough. Otherwise too many blocks.
bobboli
committed

Commits on Oct 13, 2025

Commits on Oct 19, 2025

Commits on Oct 20, 2025

Fix merge error.
bobboli
committed

Commits on Oct 21, 2025

Commits on Oct 22, 2025

Commits on Oct 23, 2025

Commits on Oct 24, 2025

Commits on Oct 25, 2025

Fix: Requires 8 GPUs to run the a2a unittest.
bobboli
committed

Commits on Oct 27, 2025