[V0 deprecation] Remove VLLM_USE_V1 usage in most modules (#27955)
Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
This commit is contained in:
@@ -154,26 +154,6 @@ AUDIO_ASSETS = AudioTestAssets()
|
||||
"""Singleton instance of {class}`AudioTestAssets`."""
|
||||
|
||||
|
||||
@pytest.fixture(scope="function", autouse=True)
|
||||
def cleanup_VLLM_USE_V1(monkeypatch):
|
||||
"""
|
||||
The V1 oracle sets "VLLM_USE_V1" during loading. This means
|
||||
that each invocation of a test change the env variable.
|
||||
|
||||
If we touch "VLLM_USE_V1" with monkeypatch, then any changes
|
||||
made during the test run by vLLM will be cleaned up.
|
||||
|
||||
This fixture is used by every test.
|
||||
"""
|
||||
|
||||
# If VLLM_USE_V1 is not set, set then delete. This will
|
||||
# cause monkeypatch to clean up VLLM_USE_V1 upon exit
|
||||
# if VLLM modifies the value of envs.VLLM_USE_V1.
|
||||
if "VLLM_USE_V1" not in os.environ:
|
||||
monkeypatch.setenv("VLLM_USE_V1", "")
|
||||
monkeypatch.delenv("VLLM_USE_V1")
|
||||
|
||||
|
||||
@pytest.fixture(autouse=True)
|
||||
def init_test_http_connection():
|
||||
# pytest_asyncio may use a different event loop per test
|
||||
|
||||
@@ -424,15 +424,12 @@ async def test_customize_loggers(monkeypatch):
|
||||
|
||||
|
||||
@pytest.mark.asyncio
|
||||
async def test_customize_aggregated_loggers(monkeypatch):
|
||||
async def test_customize_aggregated_loggers():
|
||||
"""Test that we can customize the aggregated loggers.
|
||||
If a customized logger is provided at the init, it should
|
||||
be added to the default loggers.
|
||||
"""
|
||||
|
||||
with monkeypatch.context() as m, ExitStack() as after:
|
||||
m.setenv("VLLM_USE_V1", "1")
|
||||
|
||||
with ExitStack() as after:
|
||||
with set_default_torch_num_threads(1):
|
||||
engine = AsyncLLM.from_engine_args(
|
||||
TEXT_ENGINE_ARGS,
|
||||
|
||||
@@ -868,11 +868,8 @@ def test_structured_output_batched_with_non_structured_outputs_requests(
|
||||
|
||||
@pytest.mark.parametrize("guided_decoding_backend", ["xgrammar"])
|
||||
def test_structured_output_with_structural_tag(
|
||||
monkeypatch: pytest.MonkeyPatch,
|
||||
guided_decoding_backend: str,
|
||||
):
|
||||
monkeypatch.setenv("VLLM_USE_V1", "1")
|
||||
|
||||
llm = LLM(
|
||||
model="Qwen/Qwen2.5-1.5B-Instruct",
|
||||
guided_decoding_backend=guided_decoding_backend,
|
||||
|
||||
@@ -530,7 +530,6 @@ def test_logprobs_mode(logprobs_mode: LogprobsMode):
|
||||
def test_spec_decode_logprobs(
|
||||
logprobs_mode: LogprobsMode,
|
||||
model_setup: tuple[str, str, str],
|
||||
monkeypatch: pytest.MonkeyPatch,
|
||||
):
|
||||
"""Spec decode logprobs should match those of the base model.
|
||||
|
||||
@@ -541,64 +540,62 @@ def test_spec_decode_logprobs(
|
||||
"""
|
||||
from vllm import LLM
|
||||
|
||||
with monkeypatch.context() as m:
|
||||
m.setenv("VLLM_USE_V1", "1")
|
||||
prompt = "Hello world"
|
||||
sampling_params = SamplingParams(
|
||||
temperature=0, logprobs=3, max_tokens=10, ignore_eos=False
|
||||
)
|
||||
method, model_name, spec_model_name = model_setup
|
||||
max_model_len = 256
|
||||
prompt = "Hello world"
|
||||
sampling_params = SamplingParams(
|
||||
temperature=0, logprobs=3, max_tokens=10, ignore_eos=False
|
||||
)
|
||||
method, model_name, spec_model_name = model_setup
|
||||
max_model_len = 256
|
||||
|
||||
# Run base LLM.
|
||||
ref_llm = LLM(
|
||||
model=model_name,
|
||||
max_logprobs=5,
|
||||
max_model_len=max_model_len,
|
||||
seed=42,
|
||||
logprobs_mode=logprobs_mode,
|
||||
gpu_memory_utilization=0.4,
|
||||
)
|
||||
ref_results = ref_llm.generate([prompt], sampling_params)
|
||||
# Collect logprobs outputs from reference LLM.
|
||||
ref_logprobs = []
|
||||
for output in ref_results[0].outputs:
|
||||
for logprobs in output.logprobs:
|
||||
for token_id in logprobs:
|
||||
ref_logprobs.append(logprobs[token_id])
|
||||
del ref_llm
|
||||
torch.cuda.empty_cache()
|
||||
cleanup_dist_env_and_memory()
|
||||
# Run base LLM.
|
||||
ref_llm = LLM(
|
||||
model=model_name,
|
||||
max_logprobs=5,
|
||||
max_model_len=max_model_len,
|
||||
seed=42,
|
||||
logprobs_mode=logprobs_mode,
|
||||
gpu_memory_utilization=0.4,
|
||||
)
|
||||
ref_results = ref_llm.generate([prompt], sampling_params)
|
||||
# Collect logprobs outputs from reference LLM.
|
||||
ref_logprobs = []
|
||||
for output in ref_results[0].outputs:
|
||||
for logprobs in output.logprobs:
|
||||
for token_id in logprobs:
|
||||
ref_logprobs.append(logprobs[token_id])
|
||||
del ref_llm
|
||||
torch.cuda.empty_cache()
|
||||
cleanup_dist_env_and_memory()
|
||||
|
||||
# Run spec decode LLM.
|
||||
spec_llm = LLM(
|
||||
model_name,
|
||||
speculative_config={
|
||||
"method": method,
|
||||
"model": spec_model_name,
|
||||
"num_speculative_tokens": 3,
|
||||
"max_model_len": max_model_len,
|
||||
},
|
||||
max_logprobs=5,
|
||||
max_model_len=max_model_len,
|
||||
seed=42,
|
||||
logprobs_mode=logprobs_mode,
|
||||
gpu_memory_utilization=0.4,
|
||||
)
|
||||
spec_results = spec_llm.generate([prompt], sampling_params)
|
||||
# Collect logprobs outputs from spec decode LLM.
|
||||
spec_logprobs = []
|
||||
for output in spec_results[0].outputs:
|
||||
for logprobs in output.logprobs:
|
||||
for token_id in logprobs:
|
||||
spec_logprobs.append(logprobs[token_id])
|
||||
del spec_llm
|
||||
torch.cuda.empty_cache()
|
||||
cleanup_dist_env_and_memory()
|
||||
# Run spec decode LLM.
|
||||
spec_llm = LLM(
|
||||
model_name,
|
||||
speculative_config={
|
||||
"method": method,
|
||||
"model": spec_model_name,
|
||||
"num_speculative_tokens": 3,
|
||||
"max_model_len": max_model_len,
|
||||
},
|
||||
max_logprobs=5,
|
||||
max_model_len=max_model_len,
|
||||
seed=42,
|
||||
logprobs_mode=logprobs_mode,
|
||||
gpu_memory_utilization=0.4,
|
||||
)
|
||||
spec_results = spec_llm.generate([prompt], sampling_params)
|
||||
# Collect logprobs outputs from spec decode LLM.
|
||||
spec_logprobs = []
|
||||
for output in spec_results[0].outputs:
|
||||
for logprobs in output.logprobs:
|
||||
for token_id in logprobs:
|
||||
spec_logprobs.append(logprobs[token_id])
|
||||
del spec_llm
|
||||
torch.cuda.empty_cache()
|
||||
cleanup_dist_env_and_memory()
|
||||
|
||||
# Per-token logprobs are expected to be the same.
|
||||
assert len(ref_logprobs) == len(spec_logprobs)
|
||||
for ref_logprob, spec_logprob in zip(ref_logprobs, spec_logprobs):
|
||||
assert math.isclose(ref_logprob.logprob, spec_logprob.logprob, abs_tol=1e-3)
|
||||
assert ref_logprob.rank == spec_logprob.rank
|
||||
assert ref_logprob.decoded_token == spec_logprob.decoded_token
|
||||
# Per-token logprobs are expected to be the same.
|
||||
assert len(ref_logprobs) == len(spec_logprobs)
|
||||
for ref_logprob, spec_logprob in zip(ref_logprobs, spec_logprobs):
|
||||
assert math.isclose(ref_logprob.logprob, spec_logprob.logprob, abs_tol=1e-3)
|
||||
assert ref_logprob.rank == spec_logprob.rank
|
||||
assert ref_logprob.decoded_token == spec_logprob.decoded_token
|
||||
|
||||
Reference in New Issue
Block a user