[Bugfix] Fix cuda profiler missing bug (#55237)
This commit is contained in:
@@ -4,10 +4,11 @@
|
||||
import asyncio
|
||||
import time
|
||||
from contextlib import ExitStack
|
||||
from unittest.mock import MagicMock
|
||||
from unittest.mock import AsyncMock, MagicMock, call
|
||||
|
||||
import pytest
|
||||
|
||||
import vllm.v1.engine.async_llm as async_llm_module
|
||||
from vllm import SamplingParams
|
||||
from vllm.assets.image import ImageAsset
|
||||
from vllm.config import VllmConfig
|
||||
@@ -59,6 +60,49 @@ VISION_PROMPT = {
|
||||
}
|
||||
|
||||
|
||||
def test_cuda_profiler_requests_reach_engine_core(monkeypatch: pytest.MonkeyPatch):
|
||||
vllm_config = MagicMock()
|
||||
vllm_config.observability_config.otlp_traces_endpoint = None
|
||||
vllm_config.scheduler_config.stream_interval = 1
|
||||
vllm_config.profiler_config.profiler = "cuda"
|
||||
vllm_config.profiler_config.ignore_frontend = False
|
||||
|
||||
renderer = MagicMock()
|
||||
engine_core = MagicMock()
|
||||
engine_core.profile_async = AsyncMock()
|
||||
monkeypatch.setattr(
|
||||
async_llm_module, "maybe_register_config_serialize_by_value", MagicMock()
|
||||
)
|
||||
monkeypatch.setattr(
|
||||
async_llm_module,
|
||||
"load_stat_logger_plugin_factories",
|
||||
MagicMock(return_value=[]),
|
||||
)
|
||||
monkeypatch.setattr(
|
||||
async_llm_module,
|
||||
"renderer_from_config",
|
||||
MagicMock(return_value=renderer),
|
||||
)
|
||||
monkeypatch.setattr(async_llm_module, "InputProcessor", MagicMock())
|
||||
monkeypatch.setattr(async_llm_module, "OutputProcessor", MagicMock())
|
||||
monkeypatch.setattr(
|
||||
async_llm_module.EngineCoreClient,
|
||||
"make_async_mp_client",
|
||||
MagicMock(return_value=engine_core),
|
||||
)
|
||||
|
||||
engine = AsyncLLM(vllm_config, MagicMock(), log_stats=False)
|
||||
|
||||
async def profile():
|
||||
await engine.start_profile()
|
||||
await engine.stop_profile()
|
||||
|
||||
asyncio.run(profile())
|
||||
|
||||
assert engine.profiler is None
|
||||
engine_core.profile_async.assert_has_awaits([call(True, None), call(False)])
|
||||
|
||||
|
||||
async def generate(
|
||||
engine: AsyncLLM,
|
||||
request_id: str,
|
||||
|
||||
Reference in New Issue
Block a user