Change8

b9860

📦 llama-cppView on GitHub →
3 features🐛 1 fixes🔧 3 symbols

Summary

This release introduces the public C API function llama_ftype_name() to expose model quantization type names and refactors it to be allocation-free. It also provides numerous pre-compiled binaries for different operating systems and hardware configurations.

Migration Steps

  1. If you were using the previous name for the function, rename calls from llama_model_ftype_name() to llama_ftype_name().

✨ New Features

  • Exposed the model file type (quantization) name via a new public C API function, now named llama_ftype_name().
  • The llama_ftype_name() function now prepends the "(guessed)" label instead of appending it, enabling allocation-free operation by removing a non-thread-safe static std::string.
  • Added LLAMA_FTYPE_PREFIX.

🐛 Bug Fixes

  • The implementation of llama_ftype_name() was modified to not check for a valid model, and the returned pointer is valid for the lifetime of the model or nullptr if invalid/unknown type.

Affected Symbols