A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 17, 03:00 UTC
[Misc] [CI/Build] Speed up block manager CPU-only unit tests ~10x by opting-out of GPU cleanup (#3783)
Cade Daniel committed
Apr 1, 2024 at 17:49 UTC
eb69d68804840b1108608316fe643e6a74ae44d0
3 files changed
+25
-18
tests/conftest.py
+12
-2
@@ -55,10 +55,20 @@ def cleanup():
55
torch.cuda.empty_cache()
56
57
58
+@pytest.fixture()
59
+def should_do_global_cleanup_after_test() -> bool:
60
+ """Allow subdirectories to skip global cleanup by overriding this fixture.
61
+ This can provide a ~10x speedup for non-GPU unit tests since they don't need
62
+ to initialize torch.
63
+ """
64
+ return True
65
+
66
+
67
@pytest.fixture(autouse=True)
59
-def cleanup_fixture():
68
+def cleanup_fixture(should_do_global_cleanup_after_test: bool):
69
yield
61
- cleanup()
70
+ if should_do_global_cleanup_after_test:
71
+ cleanup()
72
73
74
@pytest.fixture(scope="session")
tests/core/block/conftest.py
new
+12
@@ -0,0 +1,12 @@
1
+import pytest
2
+
3
+
4
+@pytest.fixture()
5
+def should_do_global_cleanup_after_test() -> bool:
6
+ """Disable the global cleanup fixture for tests in this directory. This
7
+ provides a ~10x speedup for unit tests that don't load a model to GPU.
8
+
9
+ This requires that tests in this directory clean up after themselves if they
10
+ use the GPU.
11
+ """
12
+ return False
tests/core/block/e2e/conftest.py
+1
-16
@@ -1,25 +1,10 @@
1
-import contextlib
2
-import gc
3
-
1
import pytest
5
-import ray
6
-import torch
2
3
+from tests.conftest import cleanup
4
from vllm import LLM
9
-from vllm.model_executor.parallel_utils.parallel_state import (
10
- destroy_model_parallel)
5
from vllm.model_executor.utils import set_random_seed
6
7
14
-def cleanup():
15
- destroy_model_parallel()
16
- with contextlib.suppress(AssertionError):
17
- torch.distributed.destroy_process_group()
18
- gc.collect()
19
- torch.cuda.empty_cache()
20
- ray.shutdown()
21
-
22
-
8
@pytest.fixture
9
def baseline_llm_generator(common_llm_kwargs, per_test_common_llm_kwargs,
10
baseline_llm_kwargs, seed):