MLX builds for Apple Silicon (8/6/4-bit + bf16)

#1
by Paris99 - opened

Hi, and thank you for releasing Laplace-1-Vector with open weights.

We converted it to MLX for Apple Silicon, from revision 06cea7d. The builds run in stock mlx-lm (0.32 and newer support the spark2_5 architecture natively), and each card includes your score_candidates helper ported to MLX with your prompt:

Against your model code in float32, the bf16 build picks the same next token on 99.6% of test positions, and every build picks the same candidate on our 16 scoring cases. On the 231 public JevBench items in candidate-scoring mode, on a MacBook Pro M5 Max: bf16 78.8%, 8-bit 79.7%, 6-bit 79.2%, 4-bit 78.4%. These come from our own harness, so they are not directly comparable with the figure on your card.

The builds are Apache-2.0 with your LICENSE and a NOTICE crediting Juneo Lab and XHToken. They are published by NetProjects, and we are not affiliated with Juneo Lab or XHToken. If you would like anything changed on the cards, please tell us here.

Paris Michalas, NetProjects

Sign up or log in to comment