[XPU] Enable XPU eplb (#44987)

Signed-off-by: mayuyuace <qiming1.zhang@intel.com>
Signed-off-by: Qiming Zhang <qiming1.zhang@intel.com>
This commit is contained in:
Qiming Zhang
2026-09-17 13:19:13 +08:00
committed by GitHub
parent 2743dc4f92
commit e6b1a5e5e3
16 changed files with 267 additions and 109 deletions
+6 -2
View File
@@ -13,6 +13,7 @@ from vllm.config import VllmConfig, set_current_vllm_config
from vllm.distributed.parallel_state import (
init_distributed_environment,
)
from vllm.platforms import current_platform
from vllm.utils.system_utils import update_environment_variables
mp.set_start_method("spawn", force=True)
@@ -73,13 +74,16 @@ def distributed_run(fn, world_size, *args):
def set_env_vars_and_device(env: dict[str, str]) -> None:
update_environment_variables(env)
local_rank = os.environ["LOCAL_RANK"]
device = torch.device(f"cuda:{local_rank}")
accelerator_type = torch.accelerator.current_accelerator().type
device = torch.device(f"{accelerator_type}:{local_rank}")
torch.accelerator.set_device_index(device)
backend = "xccl" if current_platform.is_xpu() else "nccl"
# Create a minimal vllm config for init_distributed_environment
vllm_config = VllmConfig()
with set_current_vllm_config(vllm_config):
init_distributed_environment()
init_distributed_environment(backend=backend)
atexit.register(_destroy_process_group_if_initialized)
# Ensure each worker process has the same random seed
random.seed(42)