xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2025-12-21 12:45:01 +08:00
Code Issues Packages Projects Releases Wiki Activity

[UX] Raise error for attn backend of batch invariant (#29348)

Browse Source 
Signed-off-by: yewentao256 <zhyanwentao@126.com>
This commit is contained in:
Wentao Ye 2025-11-24 19:49:01 -05:00 committed by GitHub
parent c17610e2ba
commit 699bca76c0
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 7 additions and 7 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

14
vllm/model_executor/layers/batch_invariant.py
View File

@ -812,19 +812,19 @@ def override_envs_for_invariance():
# "TRITON_MLA",
]
if curr_attn_backend not in supported_backends:
warning = (
"Forcibly updating attention backend to"
f" {supported_backends[0]} for batch_invariant. "
f" Supported backends: {supported_backends}."
error = (
"VLLM batch_invariant mode requires an attention backend in "
f"{supported_backends}, but got '{curr_attn_backend}'. "
"Please set the 'VLLM_ATTENTION_BACKEND' environment variable "
"to one of the supported backends before enabling batch_invariant."
)
logger.warning_once(warning)
os.environ["VLLM_ATTENTION_BACKEND"] = supported_backends[0]
raise RuntimeError(error)
if os.environ["VLLM_ATTENTION_BACKEND"] != supported_backends[0]:
warning = (
"You are using a decode-invariant form of batch invariance. "
"This will not be invariant between prefill and decode."
)
logger.warning_once(warning)
logger.warning_once(warning, scope="local")
os.environ["VLLM_ALLREDUCE_USE_SYMM_MEM"] = "0"
os.environ["CUBLAS_WORKSPACE_CONFIG"] = ":4096:8"