Try to allow NCCL cumem for multi node nvlink case (#11987)

This commit is contained in:
fzyzcjy
2025-11-01 03:48:25 +08:00
committed by GitHub
parent a80bcb5a68
commit 30ad107028

View File

@@ -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"