llama.cpp Disables Sliding Window Attention for Phi Models
May 21, 2025
A new update to llama.cpp disables Sliding Window Attention (SWA) specifically for Phi models. This change modifies how these models handle context and attention mechanisms within the ggml inference framework.
HOW THIS AFFECTS YOU
●
builderYou should check your Phi-based deployments to ensure performance and context handling remain as expected after this change.