Fix deterministic inference all-reduce for tp>1 (#34159)

python/sglang/srt/environ.py

@@ -906,6 +906,10 @@ class Envs:
     # Set to 1: force enable (even without --enable-deterministic-inference)
     # Set to 0: force disable (use default Aiter AR even with --enable-deterministic-inference)
     SGLANG_USE_1STAGE_ALLREDUCE = EnvBool(False)
+    # NCCL channel count pinned on CUDA so the all-reduce reduces a token the
+    # same way whatever else shares its batch. Raise it to buy back bandwidth
+    # on links that can drive more channels.
+    SGLANG_DETERMINISTIC_NCCL_NCHANNELS = EnvInt(8)
     SGLANG_OPT_USE_CUSTOM_ALL_REDUCE_V2 = EnvBool(True)
     # MiniMax-M3 on ROCm force-disables custom all-reduce in its model override
     # (arg_groups/overrides.py) when aiter all-reduce fusion is off. Set this to

python/sglang/srt/server_args.py

@@ -7972,8 +7972,14 @@ class ServerArgs:
                     # symmetric-memory path only below a byte threshold, so
                     # which reduce runs would follow the token count.
                     self.enable_torch_symm_mem = False
+                    # Each channel carries a differently shaped tree and the
+                    # channel count is picked from the message size, so a
+                    # token's reduction order would follow the token count.
+                    nchannels = str(envs.SGLANG_DETERMINISTIC_NCCL_NCHANNELS.get())
+                    os.environ["NCCL_MIN_NCHANNELS"] = nchannels
+                    os.environ["NCCL_MAX_NCHANNELS"] = nchannels
                     logger.warning(
-                        "NCCL_ALGO is set to 'allreduce:tree', and custom and symmetric-memory all reduce are disabled for deterministic inference when TP size > 1."
+                        "NCCL_ALGO is set to 'allreduce:tree', the NCCL channel count is pinned, and custom and symmetric-memory all reduce are disabled for deterministic inference when TP size > 1."
                     )
 
     def _handle_unified_memory_pool(self):