openvino
[ONNX FE][Transformation] GroupQueryAttention refine work for static shape, make valid kv left aligned [use select with mask]
#35679

Closed

Commits

initial refine work fro ONNX GQA, insert new kv into the past kv, valid data is in front of the buffer as MLAS does

bopeng1234 committed 88 days ago
add ScatterUpdate method, NPU will hang with that

bopeng1234 committed 88 days ago
remove scatterupdate, keep select with mask

bopeng1234 committed 88 days ago