[XPU][UT]Using the device_control_env_var to restrict visible device on different platform (#57581)
Signed-off-by: yisheng <yi.sheng@intel.com> Co-authored-by: Kunshang Ji <kunshang.ji@intel.com>
This commit is contained in:
@@ -15,6 +15,7 @@ import pytest
|
|||||||
from tests.utils import multi_gpu_test
|
from tests.utils import multi_gpu_test
|
||||||
from vllm.config import VllmConfig
|
from vllm.config import VllmConfig
|
||||||
from vllm.engine.arg_utils import EngineArgs
|
from vllm.engine.arg_utils import EngineArgs
|
||||||
|
from vllm.platforms import current_platform
|
||||||
from vllm.v1.core.sched.output import SchedulerOutput
|
from vllm.v1.core.sched.output import SchedulerOutput
|
||||||
from vllm.v1.executor import multiproc_executor
|
from vllm.v1.executor import multiproc_executor
|
||||||
from vllm.v1.executor.multiproc_executor import MultiprocExecutor
|
from vllm.v1.executor.multiproc_executor import MultiprocExecutor
|
||||||
@@ -378,8 +379,8 @@ def test_multiproc_executor_properties():
|
|||||||
def test_multiproc_executor_multi_node():
|
def test_multiproc_executor_multi_node():
|
||||||
"""Test MultiprocExecutor with multi-node configuration.
|
"""Test MultiprocExecutor with multi-node configuration.
|
||||||
This simulates 2 nodes with TP=4:
|
This simulates 2 nodes with TP=4:
|
||||||
- Node 0 (rank 0): Uses GPUs 0,1 (CUDA_VISIBLE_DEVICES=0,1) with TP=2
|
- Node 0 (rank 0): Uses GPUs 0,1 (device_control_env_var=0,1) with TP=2
|
||||||
- Node 1 (rank 1): Uses GPUs 2,3 (CUDA_VISIBLE_DEVICES=2,3) with TP=2
|
- Node 1 (rank 1): Uses GPUs 2,3 (device_control_env_var=2,3) with TP=2
|
||||||
Total world_size = 4, nnodes = 2
|
Total world_size = 4, nnodes = 2
|
||||||
"""
|
"""
|
||||||
# Python 3.14+ changed default multiprocessing start method to 'forkserver'
|
# Python 3.14+ changed default multiprocessing start method to 'forkserver'
|
||||||
@@ -392,15 +393,20 @@ def test_multiproc_executor_multi_node():
|
|||||||
# symm_mem does not work for simulating multi instance in single node
|
# symm_mem does not work for simulating multi instance in single node
|
||||||
os.environ["VLLM_ALLREDUCE_USE_SYMM_MEM"] = "0"
|
os.environ["VLLM_ALLREDUCE_USE_SYMM_MEM"] = "0"
|
||||||
|
|
||||||
|
# Use the platform-specific device control env var (e.g.
|
||||||
|
# CUDA_VISIBLE_DEVICES on CUDA, ZE_AFFINITY_MASK on XPU) so that each
|
||||||
|
# node only sees its own devices on different platform.
|
||||||
|
device_control_env_var = current_platform.device_control_env_var
|
||||||
|
|
||||||
def run_node(node_rank: int, result_queue: multiprocessing.Queue, port: int):
|
def run_node(node_rank: int, result_queue: multiprocessing.Queue, port: int):
|
||||||
"""Run a single node's executor."""
|
"""Run a single node's executor."""
|
||||||
executor = None
|
executor = None
|
||||||
try:
|
try:
|
||||||
# Set CUDA_VISIBLE_DEVICES for this node
|
# Restrict visible devices for this node
|
||||||
if node_rank == 0:
|
if node_rank == 0:
|
||||||
os.environ["CUDA_VISIBLE_DEVICES"] = "0,1"
|
os.environ[device_control_env_var] = "0,1"
|
||||||
else:
|
else:
|
||||||
os.environ["CUDA_VISIBLE_DEVICES"] = "2,3"
|
os.environ[device_control_env_var] = "2,3"
|
||||||
|
|
||||||
# Create config for this node
|
# Create config for this node
|
||||||
vllm_config = create_vllm_config(
|
vllm_config = create_vllm_config(
|
||||||
|
|||||||
Reference in New Issue
Block a user