llama.cpp b9544 fixes LFM2 reasoning round-trip and memory leak
Critical bug fix for Llama model reasoning round-trip and data leak in chat.
ggml-org has pushed version b9544 of llama.cpp, the widely-used C++ implementation for running large language models locally. This release addresses two critical bugs in the common/chat module: a reasoning round-trip problem in LFM2 (likely a fine-tuned Llama variant) and a stray data leak that could compromise conversation integrity. The fix is gated by reasoning format detection and whether the template supports the feature, ensuring backward compatibility.
The project, which boasts over 115,000 stars and 19,300 forks on GitHub, remains a cornerstone for on-device AI inference. This update comes with extensive build support including macOS (Apple Silicon and Intel), Linux (CPU, Vulkan, ROCm, OpenVINO, SYCL), Windows (CPU, CUDA, Vulkan, HIP), and Android. Developers relying on llama.cpp for production or experimentation should upgrade to b9544 to resolve potential conversation corruption in LFM2-based workflows.
- Fixes LFM2/LFM2.5 reasoning round-trip and stray data leak in chat module
- Gate mechanism ensures the fix only activates when reasoning format and template support it
- Supports 30+ build targets across macOS, Linux, Windows, and Android, including GPU backends
Why It Matters
Essential bugfix for local LLM users relying on llama.cpp for reliable reasoning and conversation integrity.