ggml/llama.cpp releases: b10796

TL;DR

A new function, `n_expert_used_max`, was added to the ggml/llama.cpp project. This update allows each layer in the model to have a specific number of experts, enhancing flexibility and potentially improving performance in certain configurations.

Detailed Summary

The ggml/llama.cpp project has released updates including the addition of a `n_expert_used_max` function. This update, part of a larger commit that introduced support for NVIDIA Nemotron-3-Puzzle-75B-A9B (NemotronHPuzzle), allows each layer to have a specific number of experts used, enhancing model flexibility and performance. These changes impact the broader machine learning community by providing more advanced capabilities in model architecture design.

Key Points

• • src : add n_expert_used_max function
• • Support added for NVIDIA Nemotron-3-Puzzle-75B-A9B (NemotronHPuzzle)
• • Each layer can now have a specific number of experts