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

[Bugfix] VLLM_V1 supports passing other compilation levels (#19340)

Browse Source 
Signed-off-by: Richard Zou <zou3519@gmail.com>
This commit is contained in:
Richard Zou
2025-07-29 09:35:58 -04:00
committed by GitHub
parent ab714131e4
commit 04e38500ee
5 changed files with 88 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files

55
tests/compile/test_config.py
View File

@@ -26,6 +26,8 @@ def test_use_cudagraphs_dynamic(monkeypatch):
assert not vllm_config.compilation_config.use_cudagraph
# forked needed to workaround https://github.com/vllm-project/vllm/issues/21073
@pytest.mark.forked
# NB: We don't test VLLM_DISABLE_COMPILE_CACHE=0 because that depends
# on the state of the cache directory on the current machine, which
# may be influenced by other tests.
@@ -33,8 +35,8 @@ def test_use_cudagraphs_dynamic(monkeypatch):
def test_VLLM_DISABLE_COMPILE_CACHE(vllm_runner, monkeypatch, val):
assert vllm.envs.VLLM_USE_V1
# spawn means that the counters are in the same process.
monkeypatch.setenv('VLLM_WORKER_MULTIPROC_METHOD', "spawn")
# 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)
compilation_config = {
@@ -50,6 +52,8 @@ def test_VLLM_DISABLE_COMPILE_CACHE(vllm_runner, monkeypatch, val):
pass
# forked needed to workaround https://github.com/vllm-project/vllm/issues/21073
@pytest.mark.forked
@pytest.mark.parametrize("enabled", [True, False])
def test_use_cudagraphs(vllm_runner, monkeypatch, enabled):
assert vllm.envs.VLLM_USE_V1
@@ -72,3 +76,50 @@ def test_use_cudagraphs(vllm_runner, monkeypatch, enabled):
compilation_config=compilation_config,
gpu_memory_utilization=0.4) as _):
pass
# forked needed to workaround https://github.com/vllm-project/vllm/issues/21073
@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')
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 _):
pass
# forked needed to workaround https://github.com/vllm-project/vllm/issues/21073
@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')
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 _):
pass
# forked needed to workaround https://github.com/vllm-project/vllm/issues/21073
@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')
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 _):
pass