xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2026-05-27 07:34:25 +08:00
Code Issues Packages Projects Releases Wiki Activity

[XPU] Use spawn with XPU multiprocessing (#20649)

Browse Source 
Signed-off-by: Dmitry Rogozhkin <dmitry.v.rogozhkin@intel.com>
This commit is contained in:
Dmitry Rogozhkin 2025-07-09 00:34:28 -07:00 committed by GitHub
parent 6bbf1795b7
commit e760fcef22
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
3 changed files with 15 additions and 5 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

7
tests/utils.py
View File

@ -818,14 +818,15 @@ def create_new_process_for_each_test(
Args: Args:
method: The process creation method. Can be either "spawn" or "fork". method: The process creation method. Can be either "spawn" or "fork".
If not specified, If not specified, it defaults to "spawn" on ROCm and XPU
it defaults to "spawn" on ROCm platforms and "fork" otherwise. platforms and "fork" otherwise.
Returns: Returns:
A decorator to run test functions in separate processes. A decorator to run test functions in separate processes.
""" """
if method is None: if method is None:
method = "spawn" if current_platform.is_rocm() else "fork" use_spawn = current_platform.is_rocm() or current_platform.is_xpu()
method = "spawn" if use_spawn else "fork"
assert method in ["spawn", assert method in ["spawn",
"fork"], "Method must be either 'spawn' or 'fork'" "fork"], "Method must be either 'spawn' or 'fork'"

4
tests/v1/e2e/test_cascade_attention.py
View File

@ -5,10 +5,10 @@ import pytest
from vllm import LLM, SamplingParams from vllm import LLM, SamplingParams
from ...utils import fork_new_process_for_each_test from ...utils import create_new_process_for_each_test
@fork_new_process_for_each_test @create_new_process_for_each_test()
@pytest.mark.parametrize("attn_backend", @pytest.mark.parametrize("attn_backend",
["FLASH_ATTN_VLLM_V1", "FLASHINFER_VLLM_V1"]) ["FLASH_ATTN_VLLM_V1", "FLASHINFER_VLLM_V1"])
def test_cascade_attention(example_system_message, monkeypatch, attn_backend): def test_cascade_attention(example_system_message, monkeypatch, attn_backend):

9
vllm/utils/__init__.py
View File

@ -1535,6 +1535,13 @@ def cuda_is_initialized() -> bool:
return torch.cuda.is_initialized() return torch.cuda.is_initialized()
def xpu_is_initialized() -> bool:
"""Check if XPU is initialized."""
if not torch.xpu._is_compiled():
return False
return torch.xpu.is_initialized()
def cuda_get_device_properties(device, def cuda_get_device_properties(device,
names: Sequence[str], names: Sequence[str],
init_cuda=False) -> tuple[Any, ...]: init_cuda=False) -> tuple[Any, ...]:
@ -2848,6 +2855,8 @@ def _maybe_force_spawn():
reason = None reason = None
if cuda_is_initialized(): if cuda_is_initialized():
reason = "CUDA is initialized" reason = "CUDA is initialized"
elif xpu_is_initialized():
reason = "XPU is initialized"
elif is_in_ray_actor(): elif is_in_ray_actor():
# even if we choose to spawn, we need to pass the ray address # even if we choose to spawn, we need to pass the ray address
# to the subprocess so that it knows how to connect to the ray cluster. # to the subprocess so that it knows how to connect to the ray cluster.