openvino
0cf1985a - [CPU][x86] add support for sink input of SDPA (#32374)

Commit
222 days ago
[CPU][x86] add support for sink input of SDPA (#32374) ### Details: - *extend the softmax kernel API to make it support sink input* - *in the reference implementation of sink input, there are 3 operations, **cat,** **softmax**, **drop**, because **cat** and **drop** will make attention_weight shape changed, so modify softmax API maybe a good way to reduce memory allocation and movement* - reference1: https://github.com/huggingface/transformers/blob/main/src/transformers/models/gpt_oss/modeling_gpt_oss.py#L258-L264 - reference2: https://github.com/openvinotoolkit/openvino/pull/31786/files ### Tickets: - *CVS-173533* --------- Signed-off-by: HU Yuan2 <yuan2.hu@intel.com>
Parents
Loading