onnxruntime
Add Memory Efficient Attention decode support and tests for ONNX Attention
#27851
Open
Open
Add Memory Efficient Attention decode support and tests for ONNX Attention #27851
titaiwangms
wants to merge
8 commits into
main
from
feature/mea-decode-support
Add MEA+decode test cases for ONNX Attention LLM op
719964ae
Add MEA+decode support in ONNX Attention LLM op
6ab008b7
Fix MEA eligibility: skip decode when head_size != v_head_size
f5befa80
Fix review findings: use v_head_size for V ops, add safety comment
53bddd27
Fix test review findings for MEA+decode tests
7318e243
Zero present buffers before concat to prevent NaN propagation
c2da4b12
Add asymmetric head_size regression test for MEA fallback
aadf5da1
Fix FURB110 lint: use `or` instead of ternary for v_head_size
0bdde29d
titaiwangms
changed the title Reviewers
No reviews
Assignees
No one assigned
Labels
None yet
Milestone
No milestone
Login to write a write a comment.
Login via GitHub