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

[Misc] unify variable for LLM instance (#20996)

Browse Source 
Signed-off-by: Andy Xie <andy.xning@gmail.com>
This commit is contained in:
Ning Xie
2025-07-21 19:18:33 +08:00
committed by GitHub
parent e6b90a2805
commit d97841078b
53 changed files with 237 additions and 236 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
tests/basic_correctness/test_basic_correctness.py
View File

@@ -236,13 +236,13 @@ def test_failed_model_execution(vllm_runner, monkeypatch) -> None:
monkeypatch.setenv('VLLM_ENABLE_V1_MULTIPROCESSING', '0')
with vllm_runner('facebook/opt-125m', enforce_eager=True) as vllm_model:
if isinstance(vllm_model.model.llm_engine, LLMEngineV1):
if isinstance(vllm_model.llm.llm_engine, LLMEngineV1):
v1_test_failed_model_execution(vllm_model)
def v1_test_failed_model_execution(vllm_model):
engine = vllm_model.model.llm_engine
engine = vllm_model.llm.llm_engine
mocked_execute_model = Mock(
side_effect=RuntimeError("Mocked Critical Error"))
engine.engine_core.engine_core.model_executor.execute_model =\

10
tests/basic_correctness/test_preemption.py
View File

@@ -81,7 +81,7 @@ def test_chunked_prefill_recompute(
disable_log_stats=False,
) as vllm_model:
vllm_outputs = vllm_model.generate_greedy(example_prompts, max_tokens)
assert (vllm_model.model.llm_engine.scheduler[0].artificial_preempt_cnt
assert (vllm_model.llm.llm_engine.scheduler[0].artificial_preempt_cnt
< ARTIFICIAL_PREEMPTION_MAX_CNT)
for i in range(len(example_prompts)):
@@ -118,10 +118,10 @@ def test_preemption(
distributed_executor_backend=distributed_executor_backend,
) as vllm_model:
vllm_outputs = vllm_model.generate_greedy(example_prompts, max_tokens)
assert (vllm_model.model.llm_engine.scheduler[0].artificial_preempt_cnt
assert (vllm_model.llm.llm_engine.scheduler[0].artificial_preempt_cnt
< ARTIFICIAL_PREEMPTION_MAX_CNT)
total_preemption = (
vllm_model.model.llm_engine.scheduler[0].num_cumulative_preemption)
vllm_model.llm.llm_engine.scheduler[0].num_cumulative_preemption)
check_outputs_equal(
outputs_0_lst=hf_outputs,
@@ -174,12 +174,12 @@ def test_preemption_infeasible(
) as vllm_model:
sampling_params = SamplingParams(max_tokens=max_tokens,
ignore_eos=True)
req_outputs = vllm_model.model.generate(
req_outputs = vllm_model.llm.generate(
example_prompts,
sampling_params=sampling_params,
)
assert (vllm_model.model.llm_engine.scheduler[0].artificial_preempt_cnt
assert (vllm_model.llm.llm_engine.scheduler[0].artificial_preempt_cnt
< ARTIFICIAL_PREEMPTION_MAX_CNT)
# Verify the request is ignored and not hang.