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 14, 05:30 UTC
[Compiler] Disable Inductor standalone compile by default (#25391)
Signed-off-by: ElizaWszola <ewszola@redhat.com>
ElizaWszola committed
Sep 22, 2025 at 19:37 UTC
1d7f95b85c9b1bdaa576a5275851995b75a9420d
1 file changed
+3
-2
vllm/envs.py
+3
-2
@@ -126,6 +126,7 @@ if TYPE_CHECKING:
126
VLLM_DP_RANK: int = 0
127
VLLM_DP_RANK_LOCAL: int = -1
128
VLLM_DP_SIZE: int = 1
129
+ VLLM_USE_STANDALONE_COMPILE: bool = False
130
VLLM_DP_MASTER_IP: str = ""
131
VLLM_DP_MASTER_PORT: int = 0
132
VLLM_MOE_DP_CHUNK_SIZE: int = 256
@@ -437,9 +438,9 @@ environment_variables: dict[str, Callable[[], Any]] = {
438
439
# Feature flag to enable/disable Inductor standalone compile.
440
# In torch <= 2.7 we ignore this flag; in torch >= 2.8 this is
440
- # enabled by default.
441
+ # disabled by default.
442
"VLLM_USE_STANDALONE_COMPILE":
442
- lambda: os.environ.get("VLLM_USE_STANDALONE_COMPILE", "1") == "1",
443
+ lambda: os.environ.get("VLLM_USE_STANDALONE_COMPILE", "0") == "1",
444
445
# local rank of the process in the distributed setting, used to determine
446
# the GPU device id