llama.cpp has merged a pull request (#29243) adding Fast Walsh-Hadamard Transform (FWHT) kernels for SYCL block widths above 512. Previously, larger blocks fell back to O(n^2) dense GEMM operations; the new implementation supports widths of 640, 768, 1280, and powers of two up to 8192 using an efficient butterfly network with work-group local memory shuffling. The code was verified against a recursive-doubling Hadamard reference on Intel oneAPI DPC++ 2026.1, showing negligible error margins across single- and multi-row configurations. This update improves inference efficiency for Intel GPU users without requiring new hardware, addressing a specific performance bottleneck in the SYCL backend.
Read originalEarlier coverage that leads up to this article, and what followed. Lines connect each piece to the closest one after it, converging here.
llama.cpp Releases · June 28, 2026 · Same story
llama.cpp Releases · September 19, 2026 · Same story
This release targets the persistent performance gap on AMD hardware by optimizing Vulkan Flash Attention specifically for RDNA3 and RDNA4 architectures. The changes include reducing shared memory usage through buffer reuse and fixing critical stability issues like NVIDIA hangs and sparse FA underflow bugs. While there are no new model architectures, these low-level shader optimizations directly translate to faster inference speeds for local users on modern AMD GPUs. It is a quiet but necessary engineering update that tightens the ecosystem for non-NVIDIA accelerators.
This release quietly extends llama.cpp’s hardware support to the latest NVIDIA and AMD stacks. By shipping binaries for CUDA 13.5 and ROCm 10.0, it ensures users on cutting-edge GPUs don't have to compile from source to get performance gains. The Vulkan build also fixes header inclusion issues that previously broke compilation on some Linux distributions. It’s a maintenance update rather than a feature leap, but it keeps the local inference runtime compatible with the fastest consumer and datacenter hardware available today.
This release prioritizes stability over new features, addressing a critical zstd decompression vulnerability that could have exposed response data. Beyond security, it significantly hardens the reliability of Model Context Protocol (MCP) integrations, fixing persistent connection drops and authentication loops that plagued background tasks. The update also resolves context compaction failures on third-party gateways, ensuring sessions no longer crash when hitting token limits. For developers relying on Claude Code for complex workflows, this patch removes several friction points that previously interrupted automated coding sessions.
© GitHub ChangelogvLLM Releases · November 11, 2026 · Related
llama.cpp b11059 optimizes Metal FWHT for Apple Silicon
2 developments
GitHub is quietly removing Claude Haiku 4.5 from its entire Copilot ecosystem by December 2026, a move that forces enterprise and business users to migrate away from Anthropic’s fastest model. While individual annual subscribers retain access, the deprecation signals a strategic shift in GitHub’s default model stack, likely prioritizing higher-margin or more integrated options like GPT-4o or newer Claude variants. This isn't just a sunset notice; it's a mandate for workflow updates, requiring administrators to actively manage model policies rather than relying on passive defaults. The move highlights the fragility of relying on third-party models within closed IDE ecosystems, where access can be revoked unilaterally. Developers must now audit their settings to ensure compliance with new default stacks. This transition marks a clear boundary between individual flexibility and enterprise control in AI tooling.