[v0.11.0][Perf] Eliminating the zerolike operator through patch (#3632)

### What this PR does / why we need it?
There is a zero-like operator before the attention operation in each
decoding stage. After analysis, this operator can be eliminated. The
purpose of this PR is to remove this operator and improve performance.

---------

Signed-off-by: ZYang6263 <zy626375@gmail.com>
This commit is contained in:
ZYang6263
2025-10-23 14:49:28 +08:00
committed by GitHub
parent 74903af460
commit 6975d46627
9 changed files with 111 additions and 6 deletions

View File

@@ -160,3 +160,15 @@
# Future Plan:
# Remove this patch when adapted vllm version contains the above PR.
#
# ** File: worker/patch_attention_layer.py **
# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
# 1. `vllm.attention.layer.Attention.forward`
# Why:
# There is a zerolike operator before the attention operation in each decoding stage.
# How
# Replace this zerolike operator with torch.empty
# Related PR (if no, explain why):
# - https://github.com/vllm-project/vllm/pull/26680
# Future Plan:
# Remove this to match the optimization supported in the VLLM version.
#