blindcrone

If inference engines can run the same model, nodes on those inference engines can now interoperate

This simplifies and refactors some of the orchestration messaging, done in service of enabling evaluation and training Currently this isn't quite ready for primetime as it seems to break exo...

Eval working

2

Added facilities for processing examples (which currently consist of an input, a target, and a length) and a means of evaluating them against a (currently hard-defaulted) loss function in a...

Train working

5

Currently this trains on MLX on llama-3.2-3B, but I had to pull a different version of it because I guess the MLX ops needed to train on quantized models are...