From 30ad107028dba2f7417391d30a4c769cdb662186 Mon Sep 17 00:00:00 2001 From: fzyzcjy <5236035+fzyzcjy@users.noreply.github.com> Date: Sat, 1 Nov 2025 03:48:25 +0800 Subject: [PATCH] Try to allow NCCL cumem for multi node nvlink case (#11987) --- python/sglang/srt/entrypoints/engine.py | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/python/sglang/srt/entrypoints/engine.py b/python/sglang/srt/entrypoints/engine.py index d2c4bf890..34b071486 100644 --- a/python/sglang/srt/entrypoints/engine.py +++ b/python/sglang/srt/entrypoints/engine.py @@ -677,7 +677,8 @@ class Engine(EngineBase): def _set_envs_and_config(server_args: ServerArgs): # Set global environments os.environ["TF_CPP_MIN_LOG_LEVEL"] = "3" - os.environ["NCCL_CUMEM_ENABLE"] = str(int(server_args.enable_symm_mem)) + if "NCCL_CUMEM_ENABLE" not in os.environ: + os.environ["NCCL_CUMEM_ENABLE"] = str(int(server_args.enable_symm_mem)) if not server_args.enable_symm_mem: os.environ["NCCL_NVLS_ENABLE"] = str(int(server_args.enable_nccl_nvls)) os.environ["CUDA_DEVICE_MAX_CONNECTIONS"] = "8"