TL;DR

This release consolidates several updates including performance improvements for Vulkan and CUDA, support for DeepSeek V3.2 models, and HTTP ETags for the server.

Breaking

No breaking changes were introduced in these releases.

New

• Added support for DeepSeek V3.2 models with a new sparse attention implementation.
• Implemented HTTP ETags in llama-server for improved caching.
• Optimized performance for Vulkan with a fast path for Walsh-Hadamard transforms.
• Improved CUDA performance with MMVQ optimizations for AMD MFMA hardware.

Fixes Worth Knowing

• Fixed an integer overflow in the CUDA FATN kernel.
• Corrected a memory access issue in the Vulkan memory logger.
• Addressed a bug in Arm SVE usage.
• Fixed format specifier errors in logging.

Before You Upgrade

No specific upgrade steps are required for this release.