March 2026: Stabilized CUDA execution paths for dequantization and conversion in core libraries, preventing kernel launch errors and enhancing GPU reliability. Implemented a grid.y cap of 65535 in both llama.cpp and ggml CUDA kernels, with attention to non-contiguous data handling to improve stability during quantized data processing. These fixes reduce runtime errors, improve resilience of GPU workflows, and support more robust inference pipelines across repositories.

May 2025 monthly summary for ggml-org/llama.cpp: Delivered feature enhancements to the Llama Server sampling pipeline by integrating Top-nσ into the main sampling chain and refining top_n_sigma calculations to exclude -infinity values, improving accuracy and stability in production.