Skip to content

llama.cpp b10295: fixes quantized reshaped tensor strides in model loader

llama.cpp release b10295 ships a single fix for quantized reshaped tensor strides in the model loader (PR #26672), with builds across macOS, Linux, Windows, Android, and iOS.

news · llama-cpp · local-llm · release-notes
· 1 min

llama.cpp b10295 is a single-fix release: the model loader now correctly computes strides for quantized reshaped tensors (PR #26672).

The build matrix still spans everything:

  • macOS: Apple Silicon arm64, Intel x64 (KleidiAI variant disabled)
  • Linux: Ubuntu CPU (x64, arm64, s390x), Vulkan, ROCm 7.2, OpenVINO, SYCL FP32/FP16
  • Windows: x64/arm64 CPU, CUDA 12.4 DLLs, CUDA 13.3 DLLs, Vulkan, OpenVINO, SYCL, HIP, OpenCL Adreno
  • Mobile: Android arm64, iOS XCFramework
  • openEuler: Ascend 310p and 910b (ACL Graph) on x86 + aarch64, general build disabled

Strides define how tensor indices map to memory offsets. When a quantized tensor gets reshaped and the loader computes strides incorrectly, it reads from wrong byte positions — producing anything from subtle quality loss to complete output corruption. This is a correctness patch, not a performance change.

NOTE

Two builds are disabled in this release: macOS Apple Silicon with KleidiAI, and the general openEuler target. If you depend on either, hold on the previous build.

Single-fix release, no API or CLI changes. Binaries are up on the release page for all enabled platforms.

Read next

Aug 8llama.cpp b10328: Docker tool isolation for local LLM serversAug 8llama.cpp b10331: server get_info reports correct working directory for tool isolatesAug 7llama.cpp b10306: SYCL GLU fast path, +14% f16 SWIGLU on Arc Pro B70