[Refactor] Remove dead or duplicate tests (#58446)
Signed-off-by: yewentao256 <zhyanwentao@126.com>
This commit is contained in:
@@ -341,61 +341,6 @@ def test_fp32_fallback(device: str):
|
||||
get_attn_backend(16, torch.float32, None)
|
||||
|
||||
|
||||
def test_flash_attn(monkeypatch: pytest.MonkeyPatch):
|
||||
"""Test FlashAttn validation."""
|
||||
pytest.skip(
|
||||
"Skipping as current backend selector does not "
|
||||
"handle fallbacks when a backend is explicitly set."
|
||||
)
|
||||
|
||||
attention_config = AttentionConfig(backend=AttentionBackendEnum.FLASH_ATTN)
|
||||
cache_config = CacheConfig(block_size=16)
|
||||
vllm_config = VllmConfig(
|
||||
attention_config=attention_config, cache_config=cache_config
|
||||
)
|
||||
|
||||
with set_current_vllm_config(vllm_config):
|
||||
# Unsupported CUDA arch
|
||||
monkeypatch.setattr(torch.cuda, "get_device_capability", lambda _=None: (7, 5))
|
||||
backend = get_attn_backend(16, torch.float16, None)
|
||||
assert backend.get_name() != "FLASH_ATTN"
|
||||
|
||||
# Reset the monkeypatch for subsequent tests
|
||||
monkeypatch.undo()
|
||||
|
||||
# Unsupported data type
|
||||
backend = get_attn_backend(16, torch.float8_e4m3fn, None)
|
||||
assert backend.get_name() != "FLASH_ATTN"
|
||||
|
||||
# Unsupported kv cache data type
|
||||
backend = get_attn_backend(16, torch.float16, "fp8")
|
||||
assert backend.get_name() != "FLASH_ATTN"
|
||||
|
||||
# Unsupported block size
|
||||
vllm_config.cache_config.block_size = 8
|
||||
backend = get_attn_backend(16, torch.float16, None)
|
||||
assert backend.get_name() != "FLASH_ATTN"
|
||||
|
||||
# flash-attn is not installed
|
||||
import sys
|
||||
|
||||
vllm_config.cache_config.block_size = 16
|
||||
original_module = sys.modules.get("vllm_flash_attn")
|
||||
monkeypatch.setitem(sys.modules, "vllm_flash_attn", None)
|
||||
backend = get_attn_backend(16, torch.float16, None)
|
||||
assert backend.get_name() != "FLASH_ATTN"
|
||||
|
||||
# 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)
|
||||
|
||||
# Unsupported head size
|
||||
backend = get_attn_backend(17, torch.float16, None)
|
||||
assert backend.get_name() != "FLASH_ATTN"
|
||||
|
||||
|
||||
def test_invalid_backend():
|
||||
"""Test that invalid attention backend names raise ValueError."""
|
||||
with (
|
||||
|
||||
@@ -1,73 +0,0 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
|
||||
import pytest
|
||||
import torch
|
||||
|
||||
from vllm.config import AttentionConfig, VllmConfig, set_current_vllm_config
|
||||
from vllm.platforms.rocm import RocmPlatform
|
||||
from vllm.v1.attention.backends.registry import AttentionBackendEnum
|
||||
from vllm.v1.attention.selector import _cached_get_attn_backend, get_attn_backend
|
||||
|
||||
|
||||
@pytest.fixture(autouse=True)
|
||||
def clear_cache():
|
||||
"""Clear lru cache to ensure each test case runs without caching."""
|
||||
_cached_get_attn_backend.cache_clear()
|
||||
|
||||
|
||||
@pytest.mark.skip(reason="Skipped for now. Should be revisited.")
|
||||
def test_selector(monkeypatch: pytest.MonkeyPatch):
|
||||
# Set the current platform to ROCm using monkeypatch
|
||||
monkeypatch.setattr("vllm.v1.attention.selector.current_platform", RocmPlatform())
|
||||
|
||||
# Test standard ROCm attention
|
||||
attention_config = AttentionConfig(backend=AttentionBackendEnum.ROCM_ATTN)
|
||||
vllm_config = VllmConfig(attention_config=attention_config)
|
||||
|
||||
with set_current_vllm_config(vllm_config):
|
||||
backend = get_attn_backend(16, torch.float16, torch.float16, 16, False)
|
||||
assert backend.get_name() == "ROCM_FLASH" or backend.get_name() == "TRITON_ATTN"
|
||||
|
||||
# MLA test for deepseek related
|
||||
# Change the attention backend to triton MLA
|
||||
attention_config = AttentionConfig(backend=AttentionBackendEnum.TRITON_MLA)
|
||||
vllm_config = VllmConfig(attention_config=attention_config)
|
||||
|
||||
with set_current_vllm_config(vllm_config):
|
||||
backend = get_attn_backend(576, torch.bfloat16, "auto", 16, False, use_mla=True)
|
||||
assert backend.get_name() == "TRITON_MLA"
|
||||
|
||||
# If attention backend is None
|
||||
# If use_mla is true
|
||||
# The selected backend is triton MLA
|
||||
attention_config = AttentionConfig(backend=None)
|
||||
vllm_config = VllmConfig(attention_config=attention_config)
|
||||
|
||||
with set_current_vllm_config(vllm_config):
|
||||
backend = get_attn_backend(576, torch.bfloat16, "auto", 16, False, use_mla=True)
|
||||
assert backend.get_name() == "TRITON_MLA"
|
||||
|
||||
# Change the attention backend to AITER MLA
|
||||
attention_config = AttentionConfig(backend=AttentionBackendEnum.ROCM_AITER_MLA)
|
||||
vllm_config = VllmConfig(attention_config=attention_config)
|
||||
|
||||
with set_current_vllm_config(vllm_config):
|
||||
backend = get_attn_backend(576, torch.bfloat16, "auto", 1, False, use_mla=True)
|
||||
assert backend.get_name() == "ROCM_AITER_MLA"
|
||||
|
||||
# If attention backend is None
|
||||
# If use_mla is true
|
||||
# If VLLM_ROCM_USE_AITER is enabled
|
||||
# The selected backend is ROCM_AITER_MLA
|
||||
with monkeypatch.context() as m:
|
||||
m.setenv("VLLM_ROCM_USE_AITER", "1")
|
||||
|
||||
attention_config = AttentionConfig(backend=None)
|
||||
vllm_config = VllmConfig(attention_config=attention_config)
|
||||
|
||||
with set_current_vllm_config(vllm_config):
|
||||
backend = get_attn_backend(
|
||||
576, torch.bfloat16, "auto", 1, False, use_mla=True
|
||||
)
|
||||
assert backend.get_name() == "ROCM_AITER_MLA"
|
||||
Reference in New Issue
Block a user