openvino
e4d09f99 - [CPU] Add PagedAttention Support for Sink Input and decode phases Sliding Window (#33035)

Commit
124 days ago
[CPU] Add PagedAttention Support for Sink Input and decode phases Sliding Window (#33035) ### Details: - *adds support for sink input parameter and sliding window attention mechanism in PagedAttention implementation with related test coverage.* Key Changes Sink Input Support: Extended PagedAttentionExtension to accept optional sink input as the 21st parameter, updated executor to handle sink tokens in attention computation Sliding Window Attention: Implemented sliding window mechanism for both prefill and decode phases Test Coverage: Added comprehensive test suite comparing PA vs SDPA across multiple configurations (f32/bf16, with/without sink, normal/sliding window) ### Tickets: - *CVS-173535 CVS-176456*
Author
Parents
Loading