xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2026-01-23 18:34:28 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Bugfix] Move current_platform import to avoid python import cache. (#16601)

Browse Source 
Signed-off-by: iwzbi <wzbi@zju.edu.cn>
This commit is contained in:
Wenzheng Bi 2025-10-09 18:46:19 +08:00 committed by GitHub
parent 0426e3c5e1
commit ec10fd0abc
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
2 changed files with 8 additions and 7 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

12
tests/kernels/attention/test_attention_selector.py
View File

@ -84,12 +84,12 @@ def test_env(
m.setenv("VLLM_MLA_DISABLE", "1" if use_mla else "0")
if device == "cpu":
with patch("vllm.attention.selector.current_platform", CpuPlatform()):
with patch("vllm.platforms.current_platform", CpuPlatform()):
backend = get_attn_backend(16, torch.float16, None, block_size)
assert backend.get_name() == "TORCH_SDPA"
elif device == "hip":
with patch("vllm.attention.selector.current_platform", RocmPlatform()):
with patch("vllm.platforms.current_platform", RocmPlatform()):
if use_mla:
# ROCm MLA backend logic:
# - TRITON_MLA: supported when block_size != 1
@ -126,7 +126,7 @@ def test_env(
assert backend.get_name() == expected
elif device == "cuda":
with patch("vllm.attention.selector.current_platform", CudaPlatform()):
with patch("vllm.platforms.current_platform", CudaPlatform()):
if use_mla:
# CUDA MLA backend logic:
# - CUTLASS_MLA: only supported with block_size == 128
@ -214,12 +214,12 @@ def test_env(
def test_fp32_fallback(device: str):
"""Test attention backend selection with fp32."""
if device == "cpu":
with patch("vllm.attention.selector.current_platform", CpuPlatform()):
with patch("vllm.platforms.current_platform", CpuPlatform()):
backend = get_attn_backend(16, torch.float32, None, 16)
assert backend.get_name() == "TORCH_SDPA"
elif device == "cuda":
with patch("vllm.attention.selector.current_platform", CudaPlatform()):
with patch("vllm.platforms.current_platform", CudaPlatform()):
backend = get_attn_backend(16, torch.float32, None, 16)
assert backend.get_name() == "FLEX_ATTENTION"
@ -277,7 +277,7 @@ def test_invalid_env(monkeypatch: pytest.MonkeyPatch):
"""Test that invalid attention backend names raise ValueError."""
with (
monkeypatch.context() as m,
patch("vllm.attention.selector.current_platform", CudaPlatform()),
patch("vllm.platforms.current_platform", CudaPlatform()),
):
m.setenv(STR_BACKEND_ENV_VAR, STR_INVALID_VAL)

3
vllm/attention/selector.py
View File

@ -14,7 +14,6 @@ import vllm.envs as envs
from vllm.attention.backends.abstract import AttentionBackend
from vllm.attention.backends.registry import _Backend, backend_name_to_enum
from vllm.logger import init_logger
from vllm.platforms import current_platform
from vllm.utils import STR_BACKEND_ENV_VAR, resolve_obj_by_qualname
logger = init_logger(__name__)
@ -192,6 +191,8 @@ def _cached_get_attn_backend(
)
# get device-specific attn_backend
from vllm.platforms import current_platform
attention_cls = current_platform.get_attn_backend_cls(
selected_backend,
head_size,