vllm

[Perf][MoE] Batch per-expert token count GPU sync into single transfer

#61132

Open