[Test] Fix one-sided MNNVL alltoall test workspace under-reservation (#44946)

Signed-off-by: Yongye Zhu <zyy1102000@gmail.com>

Yongye Zhu committed Jun 10, 2026 at 12:43 UTC 2ba68d9bf704434c8eb9364fefc4459819347077

1 file changed +5

tests/distributed/test_mnnvl_alltoall.py

@@ -747,6 +747,11 @@ def _one_sided_data_worker(rank, world_size):

747 top_k=experts_per_token,

748 num_experts=num_experts,

749 hidden_size=hidden_size,

750 + # Account for the fp8 block-scale payload (a1q_scale: hidden//16 bytes

751 + # per token) that is dispatched alongside the nvfp4 hidden states.

752 + # Without this the dispatch region is under-reserved and the combine

753 + # payload overflows the per-rank workspace.

754 + dispatch_scale_bytes_per_token=hidden_size // 16,

755 )

756 assert manager.initialized

757 assert manager.moe_alltoall is not None