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 12, 14:45 UTC
[UX] Raise error for attn backend of batch invariant (#29348)
Signed-off-by: yewentao256 <zhyanwentao@126.com>
Wentao Ye committed
Nov 24, 2025 at 19:49 UTC
699bca76c00b81ba6c7ead38fed01712f5f56aa1
1 file changed
+7
-7
vllm/model_executor/layers/batch_invariant.py
+7
-7
@@ -812,19 +812,19 @@ def override_envs_for_invariance():
812
# "TRITON_MLA",
813
]
814
if curr_attn_backend not in supported_backends:
815
- warning = (
816
- "Forcibly updating attention backend to"
817
- f" {supported_backends[0]} for batch_invariant. "
818
- f" Supported backends: {supported_backends}."
815
+ error = (
816
+ "VLLM batch_invariant mode requires an attention backend in "
817
+ f"{supported_backends}, but got '{curr_attn_backend}'. "
818
+ "Please set the 'VLLM_ATTENTION_BACKEND' environment variable "
819
+ "to one of the supported backends before enabling batch_invariant."
820
)
820
- logger.warning_once(warning)
821
- os.environ["VLLM_ATTENTION_BACKEND"] = supported_backends[0]
821
+ raise RuntimeError(error)
822
if os.environ["VLLM_ATTENTION_BACKEND"] != supported_backends[0]:
823
warning = (
824
"You are using a decode-invariant form of batch invariance. "
825
"This will not be invariant between prefill and decode."
826
)
827
- logger.warning_once(warning)
827
+ logger.warning_once(warning, scope="local")
828
os.environ["VLLM_ALLREDUCE_USE_SYMM_MEM"] = "0"
829
830
os.environ["CUBLAS_WORKSPACE_CONFIG"] = ":4096:8"