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

Convert formatting to use ruff instead of yapf + isort (#26247)

Browse Source 
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
This commit is contained in:
Harry Mellor
2025-10-05 15:06:22 +01:00
committed by GitHub
parent 17edd8a807
commit d6953beb91
1508 changed files with 115244 additions and 94146 deletions
Download Patch File Download Diff File Expand all files Collapse all files

172
tests/compile/test_config.py
View File

@@ -9,11 +9,11 @@ from vllm.utils import _is_torch_equal_or_newer
def test_version():
assert _is_torch_equal_or_newer('2.8.0.dev20250624+cu128', '2.8.0.dev')
assert _is_torch_equal_or_newer('2.8.0a0+gitc82a174', '2.8.0.dev')
assert _is_torch_equal_or_newer('2.8.0', '2.8.0.dev')
assert _is_torch_equal_or_newer('2.8.1', '2.8.0.dev')
assert not _is_torch_equal_or_newer('2.7.1', '2.8.0.dev')
assert _is_torch_equal_or_newer("2.8.0.dev20250624+cu128", "2.8.0.dev")
assert _is_torch_equal_or_newer("2.8.0a0+gitc82a174", "2.8.0.dev")
assert _is_torch_equal_or_newer("2.8.0", "2.8.0.dev")
assert _is_torch_equal_or_newer("2.8.1", "2.8.0.dev")
assert not _is_torch_equal_or_newer("2.7.1", "2.8.0.dev")
def test_use_cudagraphs_dynamic(monkeypatch):
@@ -21,7 +21,7 @@ def test_use_cudagraphs_dynamic(monkeypatch):
vllm_config = VllmConfig()
assert vllm_config.compilation_config.use_cudagraph
monkeypatch.setenv('VLLM_USE_V1', '0')
monkeypatch.setenv("VLLM_USE_V1", "0")
vllm_config = VllmConfig()
assert not vllm_config.compilation_config.use_cudagraph
@@ -44,19 +44,23 @@ def test_VLLM_DISABLE_COMPILE_CACHE(vllm_runner, monkeypatch, val):
assert vllm.envs.VLLM_USE_V1
# Disable multiprocessing so that the counter is in the same process
monkeypatch.setenv('VLLM_ENABLE_V1_MULTIPROCESSING', '0')
monkeypatch.setenv('VLLM_DISABLE_COMPILE_CACHE', val)
monkeypatch.setenv("VLLM_ENABLE_V1_MULTIPROCESSING", "0")
monkeypatch.setenv("VLLM_DISABLE_COMPILE_CACHE", val)
compilation_config = {
"use_cudagraph": False, # speed things up a bit
}
with (
compilation_counter.expect(num_cache_entries_updated=0,
num_compiled_artifacts_saved=0),
# loading the model causes compilation (if enabled) to happen
vllm_runner('facebook/opt-125m',
compilation_config=compilation_config,
gpu_memory_utilization=0.4) as _):
compilation_counter.expect(
num_cache_entries_updated=0, num_compiled_artifacts_saved=0
),
# loading the model causes compilation (if enabled) to happen
vllm_runner(
"facebook/opt-125m",
compilation_config=compilation_config,
gpu_memory_utilization=0.4,
) as _,
):
pass
@@ -67,22 +71,25 @@ def test_use_cudagraphs(vllm_runner, monkeypatch, enabled):
assert vllm.envs.VLLM_USE_V1
# Disable multiprocessing so that the counter is in the same process
monkeypatch.setenv('VLLM_ENABLE_V1_MULTIPROCESSING', '0')
monkeypatch.setenv("VLLM_ENABLE_V1_MULTIPROCESSING", "0")
compilation_config = {
"cudagraph_capture_sizes": [100],
"use_cudagraph": enabled,
}
with (
compilation_counter.expect(
num_graphs_seen=1,
num_gpu_runner_capture_triggers=1 if enabled else 0,
num_cudagraph_captured=13 if enabled else 0,
),
# loading the model causes compilation (if enabled) to happen
vllm_runner('facebook/opt-125m',
compilation_config=compilation_config,
gpu_memory_utilization=0.4) as _):
compilation_counter.expect(
num_graphs_seen=1,
num_gpu_runner_capture_triggers=1 if enabled else 0,
num_cudagraph_captured=13 if enabled else 0,
),
# loading the model causes compilation (if enabled) to happen
vllm_runner(
"facebook/opt-125m",
compilation_config=compilation_config,
gpu_memory_utilization=0.4,
) as _,
):
pass
@@ -90,14 +97,17 @@ def test_use_cudagraphs(vllm_runner, monkeypatch, enabled):
@pytest.mark.forked
def test_dynamo_as_is(vllm_runner, monkeypatch):
# Disable multiprocessing so that the counter is in the same process
monkeypatch.setenv('VLLM_ENABLE_V1_MULTIPROCESSING', '0')
monkeypatch.setenv("VLLM_ENABLE_V1_MULTIPROCESSING", "0")
with (
compilation_counter.expect(dynamo_as_is_count=1),
# loading the model causes compilation (if enabled) to happen
vllm_runner('facebook/opt-125m',
compilation_config={"level": 1},
gpu_memory_utilization=0.4) as _):
compilation_counter.expect(dynamo_as_is_count=1),
# loading the model causes compilation (if enabled) to happen
vllm_runner(
"facebook/opt-125m",
compilation_config={"level": 1},
gpu_memory_utilization=0.4,
) as _,
):
pass
@@ -105,14 +115,16 @@ def test_dynamo_as_is(vllm_runner, monkeypatch):
@pytest.mark.forked
def test_no_compilation(vllm_runner, monkeypatch):
# Disable multiprocessing so that the counter is in the same process
monkeypatch.setenv('VLLM_ENABLE_V1_MULTIPROCESSING', '0')
monkeypatch.setenv("VLLM_ENABLE_V1_MULTIPROCESSING", "0")
with (
compilation_counter.expect(num_graphs_seen=0,
dynamo_as_is_count=0),
# loading the model causes compilation (if enabled) to happen
vllm_runner('facebook/opt-125m',
compilation_config={"level": 0},
gpu_memory_utilization=0.4) as _):
compilation_counter.expect(num_graphs_seen=0, dynamo_as_is_count=0),
# loading the model causes compilation (if enabled) to happen
vllm_runner(
"facebook/opt-125m",
compilation_config={"level": 0},
gpu_memory_utilization=0.4,
) as _,
):
pass
@@ -120,77 +132,73 @@ def test_no_compilation(vllm_runner, monkeypatch):
@pytest.mark.forked
def test_enforce_eager(vllm_runner, monkeypatch):
# Disable multiprocessing so that the counter is in the same process
monkeypatch.setenv('VLLM_ENABLE_V1_MULTIPROCESSING', '0')
monkeypatch.setenv("VLLM_ENABLE_V1_MULTIPROCESSING", "0")
with (
compilation_counter.expect(num_graphs_seen=0,
dynamo_as_is_count=0),
# loading the model causes compilation (if enabled) to happen
vllm_runner('facebook/opt-125m',
enforce_eager=True,
gpu_memory_utilization=0.4) as _):
compilation_counter.expect(num_graphs_seen=0, dynamo_as_is_count=0),
# loading the model causes compilation (if enabled) to happen
vllm_runner(
"facebook/opt-125m", enforce_eager=True, gpu_memory_utilization=0.4
) as _,
):
pass
def test_splitting_ops_dynamic():
# Default config
config = VllmConfig()
assert config.compilation_config.cudagraph_mode == \
CUDAGraphMode.FULL_AND_PIECEWISE
assert config.compilation_config.cudagraph_mode == CUDAGraphMode.FULL_AND_PIECEWISE
assert config.compilation_config.splitting_ops_contain_attention()
# When use_inductor_graph_partition=True
if _is_torch_equal_or_newer('2.9.0.dev'):
if _is_torch_equal_or_newer("2.9.0.dev"):
# inductor graph partition is only available in PyTorch 2.9+.
# this is a fast config check so we are not using pytest.skip.
config = VllmConfig(compilation_config=CompilationConfig(
use_inductor_graph_partition=True,
splitting_ops=["silly_attention"]))
config = VllmConfig(
compilation_config=CompilationConfig(
use_inductor_graph_partition=True, splitting_ops=["silly_attention"]
)
)
# should ignore splitting_ops
assert config.compilation_config.splitting_ops == []
# When attn_fusion pass enabled.
config = VllmConfig(compilation_config=CompilationConfig(
pass_config={
"enable_attn_fusion": True,
"enable_noop": True
},
custom_ops=["+quant_fp8"],
cudagraph_mode=CUDAGraphMode.PIECEWISE,
))
config = VllmConfig(
compilation_config=CompilationConfig(
pass_config={"enable_attn_fusion": True, "enable_noop": True},
custom_ops=["+quant_fp8"],
cudagraph_mode=CUDAGraphMode.PIECEWISE,
)
)
assert config.compilation_config.splitting_ops == []
# cudagraph mode also fall back to FULL
assert config.compilation_config.cudagraph_mode == \
CUDAGraphMode.FULL
assert config.compilation_config.cudagraph_mode == CUDAGraphMode.FULL
# splitting_ops can not contain attention ops when attn_fusion
# pass enabled.
with pytest.raises(AssertionError):
config = VllmConfig(compilation_config=CompilationConfig(
pass_config={
"enable_attn_fusion": True,
"enable_noop": True
},
custom_ops=["+quant_fp8"],
cudagraph_mode=CUDAGraphMode.PIECEWISE,
# work around for accessing all attntion ops
splitting_ops=CompilationConfig()._attention_ops,
))
config = VllmConfig(
compilation_config=CompilationConfig(
pass_config={"enable_attn_fusion": True, "enable_noop": True},
custom_ops=["+quant_fp8"],
cudagraph_mode=CUDAGraphMode.PIECEWISE,
# work around for accessing all attntion ops
splitting_ops=CompilationConfig()._attention_ops,
)
)
# When both use_inductor_graph_partition and attn_fusion pass enabled.
if _is_torch_equal_or_newer('2.9.0.dev'):
config = VllmConfig(compilation_config=CompilationConfig(
use_inductor_graph_partition=True,
pass_config={
"enable_attn_fusion": True,
"enable_noop": True
},
custom_ops=["+quant_fp8"],
cudagraph_mode=CUDAGraphMode.PIECEWISE,
))
if _is_torch_equal_or_newer("2.9.0.dev"):
config = VllmConfig(
compilation_config=CompilationConfig(
use_inductor_graph_partition=True,
pass_config={"enable_attn_fusion": True, "enable_noop": True},
custom_ops=["+quant_fp8"],
cudagraph_mode=CUDAGraphMode.PIECEWISE,
)
)
assert config.compilation_config.splitting_ops == []
# enable_attn_fusion is directly support under
# use_inductor_graph_partition=True, and cudagraph_mode
# is unchanged.
assert config.compilation_config.cudagraph_mode == \
CUDAGraphMode.PIECEWISE
assert config.compilation_config.cudagraph_mode == CUDAGraphMode.PIECEWISE