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
[Bug] Temporally Disable `VLLM_ALLREDUCE_USE_SYMM_MEM` by Default (#26925)
Signed-off-by: yewentao256 <zhyanwentao@126.com>
Wentao Ye committed
Oct 15, 2025 at 16:18 UTC
e5b438a24786decebd0b69340f024f18ea184d44
1 file changed
+2
-2
vllm/envs.py
+2
-2
@@ -198,7 +198,7 @@ if TYPE_CHECKING:
198
VLLM_USE_FLASHINFER_MOE_MXFP4_BF16: bool = False
199
VLLM_ROCM_FP8_MFMA_PAGE_ATTN: bool = False
200
VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8_CUTLASS: bool = False
201
- VLLM_ALLREDUCE_USE_SYMM_MEM: bool = True
201
+ VLLM_ALLREDUCE_USE_SYMM_MEM: bool = False
202
VLLM_TUNED_CONFIG_FOLDER: str | None = None
203
VLLM_GPT_OSS_HARMONY_SYSTEM_INSTRUCTIONS: bool = False
204
VLLM_CUSTOM_SCOPES_FOR_PROFILING: bool = False
@@ -1343,7 +1343,7 @@ environment_variables: dict[str, Callable[[], Any]] = {
1343
),
1344
# Whether to use pytorch symmetric memory for allreduce
1345
"VLLM_ALLREDUCE_USE_SYMM_MEM": lambda: bool(
1346
- int(os.getenv("VLLM_ALLREDUCE_USE_SYMM_MEM", "1"))
1346
+ int(os.getenv("VLLM_ALLREDUCE_USE_SYMM_MEM", "0"))
1347
),
1348
# Allows vllm to find tuned config under customized folder
1349
"VLLM_TUNED_CONFIG_FOLDER": lambda: os.getenv("VLLM_TUNED_CONFIG_FOLDER", None),