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 15, 18:30 UTC
Fix broken tests (#14713)
Signed-off-by: JovanSardinha <jovan.sardinha@gmail.com> Co-authored-by: Luka Govedič <ProExpertProg@users.noreply.github.com>
Jovan Sardinha committed
Mar 19, 2025 at 19:06 UTC
70e500cad925499b7d1999c627e35841c937d50a
3 files changed
+19
-14
.buildkite/test-pipeline.yaml
+1
@@ -295,6 +295,7 @@ steps:
295
# these tests need to be separated, cannot combine
296
- pytest -v -s compile/piecewise/test_simple.py
297
- pytest -v -s compile/piecewise/test_toy_llama.py
298
+ - pytest -v -s compile/test_pass_manager.py
299
300
- label: PyTorch Fullgraph Test # 18min
301
source_file_dependencies:
requirements/test.in
+1
-1
@@ -30,7 +30,7 @@ matplotlib # required for qwen-vl test
30
mistral_common[opencv] >= 1.5.4 # required for pixtral test
31
datamodel_code_generator # required for minicpm3 test
32
lm-eval[api]==0.4.4 # required for model evaluation test
33
-transformers==4.48.2
33
+transformers==4.48.2
34
# quantization
35
bitsandbytes>=0.45.3
36
buildkite-test-collector==0.1.9
tests/compile/test_pass_manager.py
+17
-13
@@ -4,34 +4,38 @@ import pickle
4
5
import pytest
6
import torch
7
-from torch._inductor.codecache import BypassFxGraphCache
7
9
-from vllm.compilation.config import CompilationConfig
10
-from vllm.compilation.inductor_pass import (CallableInductorPass,
11
- as_inductor_pass)
8
+from vllm.compilation.inductor_pass import CallableInductorPass, InductorPass
9
from vllm.compilation.pass_manager import PostGradPassManager
10
+from vllm.config import CompilationConfig
11
12
13
def simple_callable(graph: torch.fx.Graph):
14
pass
15
16
19
-@as_inductor_pass(files=(__file__, ))
20
-def callable_decorated(graph: torch.fx.Graph):
21
- pass
17
+callable_uuid = CallableInductorPass(simple_callable,
18
+ InductorPass.hash_source(__file__))
19
20
21
@pytest.mark.parametrize(
22
"works, callable",
26
- [(False, simple_callable), (True, callable_decorated),
27
- (True, CallableInductorPass(simple_callable, "simple_callable"))])
23
+ [
24
+ (False, simple_callable),
25
+ (True, callable_uuid),
26
+ (True, CallableInductorPass(simple_callable)),
27
+ ],
28
+)
29
def test_pass_manager(works: bool, callable):
30
config = CompilationConfig().pass_config
30
- pass_manager = PostGradPassManager([callable])
31
- pass_manager.configure(config) # Adds default passes
31
32
+ pass_manager = PostGradPassManager()
33
+ pass_manager.configure(config)
34
+
35
+ # Try to add the callable to the pass manager
36
if works:
37
+ pass_manager.add(callable)
38
pickle.dumps(pass_manager)
39
else:
36
- with pytest.raises(BypassFxGraphCache):
37
- pickle.dumps(pass_manager)
40
+ with pytest.raises(AssertionError):
41
+ pass_manager.add(callable)