darkolorin

Basically “faster” means better performance e.g. tokens/s without loosing quality (benchmarks scores for models). So when we say faster we provide more tokens per second than llama cpp. That means we effectively utilize hardware API available (for example we wrote our own kernels) to perform better.

I made it! 90 t/s on my iPhone with llama1b fp16

We completely rewrite the inference engine and did some tricks. This is a summarization with llama 3.2 1b float16. So most of the times we do much faster than MLX. lmk in comments if you wanna test the inference and I’ll post a link.