I'd prefer to be able to compile on my Mac with 'Apple Silicon' / GPU / 'metal'!

[DXXS]

Looks like a potentially very interesting project to me!, but I'd prefer to be able to compile it on my Mac laptop with 'Apple Silicon'... I have pytorch set up to be able to use its GPU w/'metal' for other projects just fine, but your project seems to me to be overly dependent on one GPU vendor (and particularly so when attempting to build the associated 'flash-attention' module dependency).

[adityapandey9]

Hi @DXXS, I implemented the metal-based solution. Train Sudoku example. But it is super slow, here:

https://github.com/adityapandey9/hrm-comp

I train with a NVIDIA GPU with a similar config. It is faster may be because of flash attn which metal lacks.