llama.cpp b10182 moves suppress_tokens to common/sampling, fixes security
Critical security fix and refactor for local LLM inference in llama.cpp b10182
The latest release of llama.cpp, version b10182, introduces an important refactor by relocating `suppress_tokens` handling from core internals to the `common/sampling` module. This change consolidates sampling logic and directly addresses security vulnerabilities tied to the previous implementation. Additionally, the release removes the `has_logit_bias` flag, streamlining the codebase.
llama.cpp remains the go-to library for running LLMs locally on consumer hardware, and this update further enhances its modularity and safety. The release is available across a wide range of platforms: macOS (Apple Silicon with and without KleidiAI, Intel), Linux (x64, arm64, s390x) with CPU, Vulkan, ROCm 7.2, OpenVINO, and SYCL support; Windows (x64, arm64) with CUDA 12/13, Vulkan, OpenCL, and HIP; Android arm64 (CPU); and iOS via XCFramework. Users updating to b10182 will benefit from a cleaner, more secure sampling pipeline.
- Moved `suppress_tokens` handling to `common/sampling` for better modularity
- Addressed security issues related to the previous sampling implementation
- Removed `has_logit_bias` flag; supports macOS, Linux, Windows, Android, and iOS
Why It Matters
For developers running local LLMs, this security fix and refactor improves stability and safety in llama.cpp.