b8533
The latest commit patches a division-by-zero error that caused floating point exceptions during statistics generation.
The open-source project llama.cpp, maintained by ggml-org, has released a significant bug fix in commit b8533. This update specifically addresses issue #19190, which caused the software to crash with floating point exceptions when users attempted to generate statistics using the --show-statistics flag on imatrix files containing zero-count values. The crash occurred due to a division-by-zero error during processing.
The fix implements proper checks to skip zero-count entries entirely and handles empty activation vectors gracefully, preventing the crash scenario. The commit also lowers the log level for zero-count skip messages to DBG to reduce noise in standard operation. This update is particularly important because imatrix (importance matrix) files are used for quantization and model optimization in llama.cpp, making this a stability improvement for a core feature.
What makes this release notable is its comprehensive platform coverage. The fix is distributed across 24 different builds including macOS Apple Silicon (arm64), Windows with CUDA 12.4 and 13.1 support, Linux with Vulkan and ROCm 7.2 backends, iOS frameworks, and specialized builds for openEuler with Huawei Ascend NPU support. This demonstrates the project's commitment to maintaining stability across the diverse hardware ecosystem that llama.cpp supports.
- Fixes critical crash bug #19190 in imatrix feature when using --show-statistics with zero counts
- Prevents division-by-zero floating point exceptions by adding proper zero-count checks and empty vector handling
- Distributed across 24 platform-specific builds including CUDA, ROCm, Vulkan, and specialized NPU versions
Why It Matters
Ensures stable quantization workflows for developers running Llama models across diverse hardware, from consumer GPUs to enterprise NPUs.