b8999
Critical fix for --tensor-type when default qtype is overridden.
The open-source llama.cpp project, known for running large language models locally, has released version b8999. This patch addresses a regression introduced by the developer (my fault) where the `--tensor-type` flag did not properly apply when the default quantization type was overridden. The fix was contributed by Anai-Guo (re-submitted under a new policy) and closes issue #22544.
Alongside the bug fix, b8999 provides pre-built binaries for a wide range of platforms: macOS (Apple Silicon and Intel, with optional KleidiAI acceleration), Linux (CPU, GPU via Vulkan, ROCm 7.2, OpenVINO, SYCL FP32/FP16, and s390x), Windows (CPU, CUDA 12/13, Vulkan, SYCL, HIP), Android (arm64 CPU), and openEuler with Ascend NPU support. This ensures users can quickly deploy the fix across diverse hardware.
- Fixes a bug where --tensor-type was ignored when default qtype was overridden
- Patch submitted by Anai-Guo, re-submitted under new contributor policy
- Pre-built binaries for 20+ platform variants including CUDA, Vulkan, ROCm, and Ascend
Why It Matters
Maintains reliability for local LLM deployments with custom quantization settings across all major platforms.