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

[V0 Deprecation] Remove VLLM_USE_V1 from tests (#26341)

Browse Source 
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
This commit is contained in:
Cyrus Leung
2025-10-07 23:42:31 +08:00
committed by GitHub
parent c0a7b89d8e
commit 1e4ecca1d0
51 changed files with 817 additions and 1275 deletions
Download Patch File Download Diff File Expand all files Collapse all files

29
tests/kernels/attention/test_attention_selector.py
View File

@@ -80,7 +80,6 @@ def test_env(
):
"""Test attention backend selection with valid device-backend pairs."""
with monkeypatch.context() as m:
m.setenv("VLLM_USE_V1", "1")
m.setenv(STR_BACKEND_ENV_VAR, name)
m.setenv("VLLM_MLA_DISABLE", "1" if use_mla else "0")
@@ -212,30 +211,21 @@ def test_env(
@pytest.mark.parametrize("device", ["cpu", "cuda"])
def test_fp32_fallback(
device: str,
monkeypatch: pytest.MonkeyPatch,
):
def test_fp32_fallback(device: str):
"""Test attention backend selection with fp32."""
with monkeypatch.context() as m:
m.setenv("VLLM_USE_V1", "1")
if device == "cpu":
with patch("vllm.attention.selector.current_platform", CpuPlatform()):
backend = get_attn_backend(16, torch.float32, None, 16)
assert backend.get_name() == "TORCH_SDPA"
if device == "cpu":
with patch("vllm.attention.selector.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()):
backend = get_attn_backend(16, torch.float32, None, 16)
assert backend.get_name() == "FLEX_ATTENTION"
elif device == "cuda":
with patch("vllm.attention.selector.current_platform", CudaPlatform()):
backend = get_attn_backend(16, torch.float32, None, 16)
assert backend.get_name() == "FLEX_ATTENTION"
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
pytest.skip(
"Skipping as current backend selector does not "
"handle fallbacks when a backend is set via env var."
@@ -289,7 +279,6 @@ def test_invalid_env(monkeypatch: pytest.MonkeyPatch):
monkeypatch.context() as m,
patch("vllm.attention.selector.current_platform", CudaPlatform()),
):
m.setenv("VLLM_USE_V1", "1")
m.setenv(STR_BACKEND_ENV_VAR, STR_INVALID_VAL)
# Should raise ValueError for invalid backend

14
tests/kernels/test_flex_attention.py
View File

@@ -55,7 +55,6 @@ def test_flex_attention_vs_default_backend(vllm_runner, monkeypatch):
# Run with flex attention
with monkeypatch.context() as m:
m.setenv("VLLM_USE_V1", "1")
m.setenv("VLLM_ATTENTION_BACKEND", "FLEX_ATTENTION")
set_seed(seed)
@@ -72,7 +71,6 @@ def test_flex_attention_vs_default_backend(vllm_runner, monkeypatch):
# Run with default backend
with monkeypatch.context() as m:
m.setenv("VLLM_USE_V1", "1")
set_seed(seed)
with vllm_runner(
model_name,
@@ -113,7 +111,6 @@ def test_encoder_flex_attention_vs_default_backend(vllm_runner, monkeypatch):
# Run with flex attention
with monkeypatch.context() as m:
m.setenv("VLLM_USE_V1", "1")
m.setenv("VLLM_ATTENTION_BACKEND", "FLEX_ATTENTION")
with vllm_runner(
model_name,
@@ -126,17 +123,18 @@ def test_encoder_flex_attention_vs_default_backend(vllm_runner, monkeypatch):
flex_outputs = llm_flex.embed(prompts)
# Run with default backend
with monkeypatch.context() as m:
m.setenv("VLLM_USE_V1", "1")
with vllm_runner(
with (
monkeypatch.context() as m,
vllm_runner(
model_name,
runner="pooling",
dtype=torch.bfloat16,
tensor_parallel_size=1,
max_model_len=100,
enforce_eager=True,
) as llm_default:
default_outputs = llm_default.embed(prompts)
) as llm_default,
):
default_outputs = llm_default.embed(prompts)
check_embeddings_close(
embeddings_0_lst=flex_outputs,