TL;DR: llama.cpp introduced n_expert_used_max to correctly handle models with varying numbers of experts per layer, resolving loading errors for complex Mixture-of-Experts (MoE) architectures.
Summary: The latest llama.cpp update (b10796) includes the n_expert_used_max function. This addition addresses issues encountered when loading models, such as NVIDIA's Nemotron-3-Puzzle-75B-A9B, which feature a specific, potentially varying, number of experts per layer. The function ensures that model loading checks correctly account for these advanced Mixture-of-Experts (MoE) configurations.
Why it matters: This improves compatibility and stability for running sophisticated MoE models on local hardware via llama.cpp. AI builders should update to leverage broader model support and avoid loading errors with newer MoE architectures.
Source: github_releases