Package: wnpp Severity: wishlist Owner: Mathieu Baudier <[email protected]> X-Debbugs-Cc: [email protected], [email protected]
* Package name : python-mistral-common Version : 1.11.7 Upstream Contact: Mistral AI <[email protected]> * URL : https://github.com/mistralai/mistral-common * License : Apache-2.0 Programming Lang: Python Description : Official inference library for pre-processing of Mistral models Mistral-common is a library of common utilities for Mistral AI. It is notably used to convert LLMs in Mistral-specific format to the GGUF file format used for inference by llama.cpp. ## Relevance of the package ## As part of the llama.cpp packaging effort, I am working on the conversion of LLM models from the safetensors format to the GGUF format which is used by llama.cpp, and is generally more efficient for LLM inference deployment. Thanks to the recent addition of required machine-learning Python modules, these conversions now work in unstable for the vast majority of models, using only proper Debian packages. However, the conversion of the Mistral family of LLMs requires a specific Python module (mistral-common), which this package will provide. A working draft is available here: https://salsa.debian.org/mbaudier/python-mistral-common ## Maintenance ## I have little experience with Python, so this package would probably be maintained by 'Debian Deep Learning Team <[email protected]>' along with the other ML-related Python packages, with whom it shares most dependencies. The module looked rather clean, as far as I could see, and it was straightforward to package with --buildsystem=pybuild. My goal was just to save time to more experienced Python-packagers and make sure it would work without requiring further dependencies, etc.
