From 400c1dfe6eb42b2b3ff79c8a1f773227c84094ec Mon Sep 17 00:00:00 2001 From: Mick Date: Sat, 17 Oct 2026 18:44:31 +0800 Subject: [PATCH] [Perf] distributed: skip the DLPack re-wrap of custom all-reduce results (#44251) Co-authored-by: Mick Qian Co-authored-by: Claude Opus 5.5 --- python/sglang/kernels/ops/communication/all_reduce.py | 4 ++++ 1 file changed, 4 insertions(+) diff --git a/python/sglang/kernels/ops/communication/all_reduce.py b/python/sglang/kernels/ops/communication/all_reduce.py index ed6e036bd8..ae6d3cf266 100644 --- a/python/sglang/kernels/ops/communication/all_reduce.py +++ b/python/sglang/kernels/ops/communication/all_reduce.py @@ -228,6 +228,10 @@ def custom_all_reduce( ) -> torch.Tensor: module = get_all_reduce_module(input.dtype, comm.world_size) result = module.all_reduce(comm, input, algo.algo_name, graph_params, use_multicast) + # Already a torch.Tensor once tvm-ffi loads torch's DLPack exchange API; + # re-wrapping it is a pure round trip (~15 us per call on GB300). + if isinstance(result, torch.Tensor): + return result return torch.from_dlpack(result)