llama.cpp introduces support for the NVIDIA Nemotron-3-Puzzle-75B-A9B model, enabling inference for this Mixture of Experts architecture. The update implements infrastructure for variable-per-layer expert feed-forward sizes and top-k routing, which are required because the model has 40 MoE layers with distinct n_ff_exp and top-k values.

  • Adds per-layer n_ff_exp and n_expert_used arrays with scalar-or-array loading capabilities.
  • Implements G1/G2 infrastructure to handle heterogeneous expert configurations across layers.
  • Updates the converter to parse block_configs and normalize tensor names for official BF16 checkpoints.
  • Distinguishes Nemotron 3 Puzzle from Super variants based on the per-layer expert_used_count array.

This change allows llama.cpp to load and run the Nemotron-3-Puzzle model, which previously lacked support due to its complex, non-uniform MoE structure.