llamafile v0.10.0 Alpha: The Christmas update 🎄

[aittalam]

Hi all, here's an update about our ongoing work on the new_build_wip branch of llamafile.

We successfully achieved our first goal of replicating a cosmopolitan llama.cpp build from scratch to get the best of two worlds:

• Classic llamafile portability: we managed to preserve portability across different systems and architectures and the possibility of bundling model weights within llamafile executables.
• Modern features: we now also have access to the features and model support made available by recent versions of llama.cpp.

We realise that what makes a llamafile is not just an APE executable, so we are now bringing back more of its favorite features. We have started with the llamafile TUI and just merged a PR which introduces dynamic GPU support (starting with Metal).

⚠️ A Note on Stability ⚠️
While we are adding llamafile features back, some parts of the build might fail (for instance, the current build does not support whisperfile and stable diffusion yet) and not all the features you were accustomed to might be present.
If you want the more stable build with all the features you are used to, please use the code (and instructions) on the main branch. In our work branch, instead, you'll find the most recent, less stable stuff. The README_0_10_0.md file will always contain the most up-to-date installation instructions.
Last but not least, we want to be mindful of the fact that there are recent issues and PRs that we have not addressed yet. We want you to know that we are aware of them, we will get back to them one by one as soon as our code is more stable, and we will try to integrate your suggestions directly into the new code if they are still pertinent.

🎁 Christmas Presents 🎅
Here is a quick video showing how to build from scratch from our work branch.
Here is a page from which you can download some pre-compiled binaries to play with! We chose three models we like to experiment with, but we are happy to hear suggestions from you for the future alpha binaries.