alt Hacker NewsThat isn’t right. Use oMLX or something similar to serve the model with internal MTP enabled. You should get at least 40 tok/sec, but I’m not sure what hardware you are using. Dflash will mess up batching, not really worth it. You need to sift through hugging face for the right model though, and some of the MTP models are meant for rapid MLX and not oMLX.
Replies
Well, yes and no.
The author has an M3.
Here's reality, MLX on the software layer will not magically place hardware matrix multiplication units in your GPU cores.
Newer Macs are always just gonna smoke anything earlier than an M5.
Not even sure why this person is trying to get this stuff to run on hardware that wasn't designed for AI?
Everyone is pointing him to newer hardware precisely because you need the newer stuff to get models to be performant. You can go with AMD, NVidia or Apple, but you're gonna be using stuff designed well after the M3 if you want to push >100tok/s.
I was just trying that yesterday on my M4 Max with the 6bit quant. It started at 40 tps but dropped to 10 once the context loaded up.