llama.cpp b9986 patches reasoning leak and Nemotron regression
New release fixes template issue that swallowed reasoning content in chat models.
The ggml-org/llama.cpp project released version b9986 on July 13, focusing on two important chat-related fixes. The first patch addresses a "reasoning leak" bug that occurred with force-opened bare templates. In previous versions, the reasoning start tag inferred from prior turns could carry trailing whitespace (e.g., a newline), causing a bare prefill template (like a simple `</s>`) to be mismatched and swallowed into the content rather than triggering the intended reasoning behavior. The fix trims that whitespace so the prefill is correctly recognized.
The second patch fixes a regression in Nemotron Nano v2 support. This model, based on NVIDIA's Nemotron architecture, had broken behavior in earlier releases. The commit also includes cross-platform builds for macOS (Apple Silicon and Intel), Linux (x64/arm64/s390x with CPU, Vulkan, ROCm, OpenVINO, SYCL), Windows (CPU, CUDA 12/13, Vulkan, OpenVINO, SYCL, HIP), and Android (arm64), ensuring broad usability for local LLM inference.
- Fixes reasoning leak in force-opened bare templates by trimming trailing whitespace from inferred tags.
- Restores compatibility with Nemotron Nano v2, a key model for resource-constrained environments.
- Available on 20+ platform/backend configurations including CUDA 12/13, ROCm, Vulkan, and Apple Silicon.
Why It Matters
Critical patch for local LLM users running chat models with customizable templates and NVIDIA's Nemotron family.