[Bugfix] Deterministic MoE combine (reduce_scatterv) under VLLM_BATCH_INVARIANT (#45683)

Signed-off-by: shijuzhao <shijuzhao@tencent.com>
Co-authored-by: shijuzhao <shijuzhao@tencent.com>
This commit is contained in:
shijuzhao
2026-08-21 21:41:31 +08:00
committed by GitHub
co-authored by shijuzhao
parent c8438a3d40
commit cda3868f5d
2 changed files with 68 additions and 1 deletions
@@ -451,7 +451,14 @@ class CudaCommunicator(DeviceCommunicatorBase):
output = torch.empty(
output_shape, dtype=input_tensor.dtype, device=input_tensor.device
)
if sizes is not None and sizes.count(sizes[0]) != len(sizes):
use_deterministic_rs = envs.VLLM_BATCH_INVARIANT and world_size > 2
if use_deterministic_rs:
# Reduce to a fixed root (0) for determinism
reduced = torch.empty_like(input_tensor)
sizes = sizes if sizes else [chunk_size] * world_size
pynccl_comm.reduce(reduced, input_tensor, root=0)
pynccl_comm.scatter(output, reduced, sizes, root=0)
elif sizes is not None and sizes.count(sizes[0]) != len(sizes):
pynccl_comm.reduce_scatterv(output, input_tensor, sizes=sizes)
else:
pynccl_comm.reduce_scatter(output, input_tensor)