Flash Attention

This release introduces several new models including Meta Muse Glimmer, GraniteMoeSWA, A.X-K1/K2, and Cosmos3 Edge. It also includes significant updates to attention mechanisms, vision processing, and generation capabilities, alongside important breaking changes in kernel opt-in and cache cropping.

Introduces SYCL Flash Attention with XMX engine via oneDNN, improving prefill speeds for Qwen3.6-27b-Q8_0. Includes optimizations for multi-device synchronization and addresses review feedback on the oneDNN path.

This release introduces significant performance enhancements for Vulkan backends by adding support for the `v_dot2_f32_f16` extension in matrix multiplication and Flash Attention kernels. Several platform builds were disabled pending further updates.

This release introduces a CUDA optimization to reduce memory overhead by conditionally allocating the Flash Attention temporary buffer. It includes a wide range of pre-built binaries for multiple operating systems and hardware architectures.

This release introduces a Vulkan optimization that reduces Flash Attention rows for small cache runs to improve efficiency.

This release introduces support for Nemotron 3 Nano and Olmo 3 models, enables Flash Attention by default, and provides critical fixes for Gemma 3 model architectures.

This release focuses on performance optimizations for qwen3-vl, including default Flash Attention support, and fixes several issues related to model thinking modes and image processing.