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
[compile] Turn standalone_compile back on (#27460)
Signed-off-by: Richard Zou <zou3519@gmail.com>
Richard Zou committed
Oct 24, 2025 at 09:30 UTC
cd390b609d101265ec5ec5ebb47c4c2c6ddc4a49
1 file changed
+4
-4
vllm/envs.py
+4
-4
@@ -131,7 +131,7 @@ if TYPE_CHECKING:
131
VLLM_DP_RANK: int = 0
132
VLLM_DP_RANK_LOCAL: int = -1
133
VLLM_DP_SIZE: int = 1
134
- VLLM_USE_STANDALONE_COMPILE: bool = False
134
+ VLLM_USE_STANDALONE_COMPILE: bool = True
135
VLLM_DP_MASTER_IP: str = ""
136
VLLM_DP_MASTER_PORT: int = 0
137
VLLM_MOE_DP_CHUNK_SIZE: int = 256
@@ -494,10 +494,10 @@ environment_variables: dict[str, Callable[[], Any]] = {
494
os.environ.get("VLLM_FLASH_ATTN_VERSION", None)
495
),
496
# Feature flag to enable/disable Inductor standalone compile.
497
- # In torch <= 2.7 we ignore this flag; in torch >= 2.8 this is
498
- # disabled by default.
497
+ # In torch <= 2.7 we ignore this flag; in torch >= 2.9 this is
498
+ # enabled by default.
499
"VLLM_USE_STANDALONE_COMPILE": lambda: os.environ.get(
500
- "VLLM_USE_STANDALONE_COMPILE", "0"
500
+ "VLLM_USE_STANDALONE_COMPILE", "1"
501
)
502
== "1",
503
# Debug pattern matching inside custom passes.