biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Misc] Replace os environ to monkeypatch in test suite (#14516)

Browse Source 
Signed-off-by: sibi <85477603+t-sibiraj@users.noreply.github.com>
Signed-off-by: Aaron Pham <contact@aarnphm.xyz>
Co-authored-by: Cyrus Leung <cyrus.tl.leung@gmail.com>
Co-authored-by: Aaron Pham <contact@aarnphm.xyz>
This commit is contained in:
Sibi
2025-03-17 11:35:57 +08:00
committed by GitHub
parent 1e799b7ec1
commit a73e183e36
43 changed files with 1900 additions and 1658 deletions
Download Patch File Download Diff File Expand all files Collapse all files

129
tests/kernels/test_attention_selector.py
View File

@@ -5,13 +5,12 @@ from unittest.mock import Mock, patch
import pytest
import torch
from tests.kernels.utils import override_backend_env_variable
from vllm.attention.selector import _cached_get_attn_backend, get_attn_backend
from vllm.platforms.cpu import CpuPlatform
from vllm.platforms.cuda import CudaPlatform
from vllm.platforms.openvino import OpenVinoPlatform
from vllm.platforms.rocm import RocmPlatform
from vllm.utils import STR_FLASH_ATTN_VAL, STR_INVALID_VAL
from vllm.utils import STR_BACKEND_ENV_VAR, STR_FLASH_ATTN_VAL, STR_INVALID_VAL
@pytest.fixture(autouse=True)
@@ -25,87 +24,111 @@ def clear_cache():
"name", ["TORCH_SDPA", "ROCM_FLASH", "XFORMERS", "FLASHINFER", "OPENVINO"])
@pytest.mark.parametrize("use_v1", [True, False])
@pytest.mark.parametrize("device", ["cpu", "openvino", "hip", "cuda"])
def test_env(name: str, use_v1: bool, device: str, monkeypatch):
def test_env(
name: str,
use_v1: bool,
device: str,
monkeypatch: pytest.MonkeyPatch,
):
"""Test that the attention selector can be set via environment variable.
Note that we do not test FlashAttn because it is the default backend.
"""
monkeypatch.setenv("VLLM_USE_V1", "1" if use_v1 else "0")
override_backend_env_variable(monkeypatch, name)
with monkeypatch.context() as m:
m.setenv("VLLM_USE_V1", "1" if use_v1 else "0")
m.setenv(STR_BACKEND_ENV_VAR, name)
if device == "cpu":
with patch("vllm.attention.selector.current_platform", CpuPlatform()):
backend = get_attn_backend(16, torch.float16, torch.float16, 16,
False)
assert backend.get_name() == "TORCH_SDPA"
elif device == "hip":
with patch("vllm.attention.selector.current_platform", RocmPlatform()):
backend = get_attn_backend(16, torch.float16, torch.float16, 16,
False)
EXPECTED = "ROCM_ATTN_VLLM_V1" if use_v1 else "ROCM_FLASH"
assert backend.get_name() == EXPECTED
elif device == "openvino":
with patch("vllm.attention.selector.current_platform",
OpenVinoPlatform()), patch.dict('sys.modules',
{'openvino': Mock()}):
backend = get_attn_backend(16, torch.float16, torch.float16, 16,
False)
assert backend.get_name() == "OPENVINO"
else:
if name in ["XFORMERS", "FLASHINFER"]:
if device == "cpu":
with patch("vllm.attention.selector.current_platform",
CudaPlatform()):
CpuPlatform()):
backend = get_attn_backend(16, torch.float16, torch.float16,
16, False)
EXPECTED = "FLASH_ATTN_VLLM_V1" if use_v1 else name
assert backend.get_name() == "TORCH_SDPA"
elif device == "hip":
with patch("vllm.attention.selector.current_platform",
RocmPlatform()):
backend = get_attn_backend(16, torch.float16, torch.float16,
16, False)
EXPECTED = "ROCM_ATTN_VLLM_V1" if use_v1 else "ROCM_FLASH"
assert backend.get_name() == EXPECTED
elif device == "openvino":
with patch("vllm.attention.selector.current_platform",
OpenVinoPlatform()), patch.dict('sys.modules',
{'openvino': Mock()}):
backend = get_attn_backend(16, torch.float16, torch.float16,
16, False)
assert backend.get_name() == "OPENVINO"
else:
if name in ["XFORMERS", "FLASHINFER"]:
with patch("vllm.attention.selector.current_platform",
CudaPlatform()):
backend = get_attn_backend(16, torch.float16,
torch.float16, 16, False)
EXPECTED = "FLASH_ATTN_VLLM_V1" if use_v1 else name
assert backend.get_name() == EXPECTED
def test_flash_attn(monkeypatch):
def test_flash_attn(monkeypatch: pytest.MonkeyPatch):
"""Test FlashAttn validation."""
# TODO: When testing for v1, pipe in `use_v1` as an argument to
# get_attn_backend
override_backend_env_variable(monkeypatch, STR_FLASH_ATTN_VAL)
with monkeypatch.context() as m:
m.setenv(STR_BACKEND_ENV_VAR, STR_FLASH_ATTN_VAL)
# Unsupported CUDA arch
with patch("torch.cuda.get_device_capability", return_value=(7, 5)):
# Unsupported CUDA arch
monkeypatch.setattr(torch.cuda, "get_device_capability", lambda:
(7, 5))
backend = get_attn_backend(16, torch.float16, None, 16, False)
assert backend.get_name() != STR_FLASH_ATTN_VAL
# Unsupported data type
backend = get_attn_backend(16, torch.float8_e4m3fn, None, 16, False)
assert backend.get_name() != STR_FLASH_ATTN_VAL
# Reset the monkeypatch for subsequent tests
monkeypatch.undo()
# Unsupported kv cache data type
backend = get_attn_backend(16, torch.float16, "fp8", 16, False)
assert backend.get_name() != STR_FLASH_ATTN_VAL
# Unsupported data type
backend = get_attn_backend(16, torch.float8_e4m3fn, None, 16, False)
assert backend.get_name() != STR_FLASH_ATTN_VAL
# Unsupported block size
backend = get_attn_backend(16, torch.float16, None, 8, False)
assert backend.get_name() != STR_FLASH_ATTN_VAL
# Unsupported kv cache data type
backend = get_attn_backend(16, torch.float16, "fp8", 16, False)
assert backend.get_name() != STR_FLASH_ATTN_VAL
# flash-attn is not installed
with patch.dict('sys.modules', {'vllm_flash_attn': None}):
# Unsupported block size
backend = get_attn_backend(16, torch.float16, None, 8, False)
assert backend.get_name() != STR_FLASH_ATTN_VAL
# flash-attn is not installed
import sys
original_module = sys.modules.get('vllm_flash_attn')
monkeypatch.setitem(sys.modules, 'vllm_flash_attn', None)
backend = get_attn_backend(16, torch.float16, None, 16, False)
assert backend.get_name() != STR_FLASH_ATTN_VAL
# Unsupported head size
backend = get_attn_backend(17, torch.float16, None, 16, False)
assert backend.get_name() != STR_FLASH_ATTN_VAL
# Restore the original module if it existed
if original_module is not None:
monkeypatch.setitem(sys.modules, 'vllm_flash_attn',
original_module)
else:
monkeypatch.delitem(sys.modules, 'vllm_flash_attn', raising=False)
# Attention-free models should bypass env and use PlaceholderAttention
backend = get_attn_backend(16, torch.float16, torch.float16, 16, True)
assert backend.get_name() != STR_FLASH_ATTN_VAL
# Unsupported head size
backend = get_attn_backend(17, torch.float16, None, 16, False)
assert backend.get_name() != STR_FLASH_ATTN_VAL
# Attention-free models should bypass env and use PlaceholderAttention
backend = get_attn_backend(16, torch.float16, torch.float16, 16, True)
assert backend.get_name() != STR_FLASH_ATTN_VAL
@pytest.mark.parametrize("use_v1", [True, False])
def test_invalid_env(use_v1: bool, monkeypatch):
"""Ignore the invalid env variable if it is set."""
monkeypatch.setenv("VLLM_USE_V1", "1" if use_v1 else "0")
override_backend_env_variable(monkeypatch, STR_INVALID_VAL)
def test_invalid_env(use_v1: bool, monkeypatch: pytest.MonkeyPatch):
with patch("vllm.attention.selector.current_platform", CudaPlatform()):
with monkeypatch.context() as m, patch(
"vllm.attention.selector.current_platform", CudaPlatform()):
m.setenv("VLLM_USE_V1", "1" if use_v1 else "0")
m.setenv(STR_BACKEND_ENV_VAR, STR_INVALID_VAL)
# Test with head size 32
backend = get_attn_backend(32, torch.float16, None, 16, False)
EXPECTED = "FLASH_ATTN_VLLM_V1" if use_v1 else "FLASH_ATTN"
assert backend.get_name() == EXPECTED