onnxruntime

Refactoring of attention cuda kernel: move prepare qkv and concat_past_to_present

#17559

Merged