These were converted and quantized from source safetensors using llama.cpp on April 3, 2024.
This matters because there are several GGUF files on HF which were created before llama.cpp's support for MoE quantization was fully debugged,
even though it looked like it was producing working files at the time.
I'll be uploading the quantized .gguf sources I created as well if anyone wants them as a reference or for further work.
Llamafiles are a standalone executable that run an LLM server locally on a variety of operating systems including FreeBSD, Windows, Windows via WSL, Linux, and Mac.
The same file works everywhere, I've tested several of these on FreeBSD, Windows, Windows via WSL, and Linux.
You just download the .llamafile, (chmod +x or rename to .exe as needed), run it, open the chat interface in a browser, and interact.
Options can be passed in to expose the api etc. See their
docs for details.
-
Windows note: If it's over 4gb and you want to use it on Windows, you'll have to run it from WSL.
-
WSL note: If you get the error about APE, and the recommended command
sudo sh -c 'echo -1 > /proc/sys/fs/binfmt_misc/WSLInterop'
doesn't work, the WSLInterop file might be named something else. I had success with
sudo sh -c 'echo -1 > /proc/sys/fs/binfmt_misc/WSLInterop-late'
If that fails too, just navigate to /proc/sys/fs/binfmt_msc and see what files look like WSLInterop
and echo a -1 to whatever they're called by changing that part of the recommended command.
-
FreeBSD note: Yes, it actually works on a fresh install of FreeBSD.
