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] Fix `returned_lse` not Defined issue (#25106)
Signed-off-by: yewentao256 <zhyanwentao@126.com> Co-authored-by: Tyler Michael Smith <tyler@neuralmagic.com>
Wentao Ye committed
Sep 18, 2025 at 15:32 UTC
75fb112d80f680624dc99a00e02be6a45661f948
1 file changed
+3
-4
vllm/v1/attention/backends/mla/cutlass_mla.py
+3
-4
@@ -206,12 +206,11 @@ class CutlassMLAImpl(MLACommonImpl[MLACommonMetadata]):
206
)
207
208
if H < MAX_HEADS:
209
- # Extract the subsets of the outputs
210
- returned_lse = lse[:, :H].contiguous(
211
- ) if self.need_to_return_lse_for_decode else lse
209
out = out[:, :H]
210
+ if self.need_to_return_lse_for_decode:
211
+ lse = lse[:, :H].contiguous()
212
214
- return out, returned_lse
213
+ return out, lse
214
215
def _forward_decode(
216
self,