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, 06:00 UTC
[BugFix] Allow use_cudagraph to work with dynamic VLLM_USE_V1 (#19390)
Signed-off-by: rzou <zou3519@gmail.com>
Richard Zou committed
Jun 10, 2025 at 19:54 UTC
77f0d465d0a666b65dd877ec462f024a980dd55c
2 files changed
+11
-1
tests/compile/test_config.py
+10
@@ -11,6 +11,16 @@ from vllm.config import (CompilationConfig, CompilationLevel, VllmConfig,
11
from .piecewise.test_simple import SillyModel
12
13
14
+def test_use_cudagraphs_dynamic(monkeypatch):
15
+ assert vllm.envs.VLLM_USE_V1
16
+ vllm_config = VllmConfig()
17
+ assert vllm_config.compilation_config.use_cudagraph
18
+
19
+ monkeypatch.setenv('VLLM_USE_V1', '0')
20
+ vllm_config = VllmConfig()
21
+ assert not vllm_config.compilation_config.use_cudagraph
22
+
23
+
24
@pytest.mark.parametrize("enabled", [True, False])
25
def test_use_cudagraphs(enabled):
26
assert vllm.envs.VLLM_USE_V1
vllm/config.py
+1
-1
@@ -3931,7 +3931,7 @@ class CompilationConfig:
3931
constructor, e.g. `CompilationConfig(inductor_passes={"a": func})`."""
3932
3933
# CudaGraph compilation
3934
- use_cudagraph: bool = envs.VLLM_USE_V1
3934
+ use_cudagraph: bool = field(default_factory=lambda: envs.VLLM_USE_V1)
3935
"""Whether to use cudagraph inside compilation.
3936
- False: cudagraph inside compilation is not used.
3937
- True: cudagraph inside compilation is used. It requires