I’m running TheBloke/Llama-2-13B-chat-GGUF on my 14 CPU/30GPU 36GB Ram M3 Max via Text generation web UI.

I get max 20 tokens/second. I’ve tried various parameter presets and they all seem to get me around the same 20 toks/sec.

I’ve tried increasing the threads to 14 and n-GPU-layers to 128. I don’t really understand most of the parameters in the model and parameters tab. Just cranking different options to see how I can increase toks/sec. But so far, nothing above 20 toks/sec.

What can I change to crank the performance here? I’m yet to hear the fan go off for this 13B model. I’m trying to push it to the max to see the max toks/sec I can achieve on my machine. Any settings I should try?

  • FlishFlashman@alien.topBEnglish
    1·
    1 year ago

    What quantization are you using? Smaller tends to be faster.

    I get 30 tokens/s with a q4_0 quantization of 13B models on a M1 Max on Ollama (which uses llama.cpp). You should be in the same ballpark with the same software. You aren’t going to do much/any better than that. The M3’s GPU made some significant leaps for graphics, and little to nothing for LLMs.

    Allowing more threads isn’t going to help generation speed, it might improve prompt processing though. Probably best though to keep the number of threads to the number of performance cores.