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 6, 05:45 UTC
[ROCm][CI] Fix MEGA_AOT_ARTIFACT fallback when PyTorch < 2.10.0 lacks AOT support (#37763)
Signed-off-by: Andreas Karatzas <akaratza@amd.com>
Andreas Karatzas committed
Mar 22, 2026 at 03:02 UTC
3b06c55c78e85b6e71e086db3a77226405550f2d
1 file changed
+5
-1
tests/compile/test_startup.py
+5
-1
@@ -15,6 +15,7 @@ from torch._dynamo.utils import counters
15
import vllm.envs as envs
16
from vllm.compilation.counter import compilation_counter
17
from vllm.config import CompilationConfig, CompilationMode, CUDAGraphMode
18
+from vllm.utils.torch_utils import is_torch_equal_or_newer
19
20
from ..utils import fork_new_process_for_each_test
21
@@ -71,7 +72,10 @@ def test_moe_startup(monkeypatch, vllm_runner, fresh_vllm_cache, mega_aot_artifa
72
num_compiled_artifacts_saved=0,
73
):
74
_run_vllm(vllm_runner)
74
- if envs.VLLM_USE_MEGA_AOT_ARTIFACT:
75
+ mega_aot_active = envs.VLLM_USE_MEGA_AOT_ARTIFACT and is_torch_equal_or_newer(
76
+ "2.10.0"
77
+ )
78
+ if mega_aot_active:
79
# MEGA_AOT_ARTIFACT is enabled, so we expect no aot_autograd running on
80
# subgraphs.
81
assert counters["aot_autograd"]["total"] == 0