[1/N] Optimize All Reduce - Benchmark different AR operations (#13797)

Co-authored-by: luoyuan.luo <luoyuan.luo@antgroup.com>
2026-01-26 22:44:13 +08:00
parent 6c0f9b4824
commit 7bb41989fa
2 changed files with 353 additions and 2 deletions
--- a/python/sglang/srt/distributed/device_communicators/custom_all_reduce.py
+++ b/python/sglang/srt/distributed/device_communicators/custom_all_reduce.py
@@ -325,12 +325,12 @@ class CustomAllreduce:
        # little performance improvement over NCCL.
        if not _is_hip:
            if self.world_size == 2 or self.full_nvlink:
-                return inp_size < self.max_size
+                return inp_size <= self.max_size
            return False

        if _is_hip:
            if self.full_nvlink:
-                return inp_size < self.max_size
+                return inp_size <= self.max_size
            return False

        return False