3 ms·

But it's likely to be much slower than what you'd get with a backend like llama.cpp on CPU (particularly if you're running on a Mac, but I think on Linux as wel

But it's likely to be much slower than what you'd get with a backend like llama.cpp on CPU (particularly if you're running on a Mac, but I think on Linux as well), as well as not supporting features like CPU offloading.

p1esk 2y ago
Are there benchmarks? 2x speed up would not be enough for me to return to c++ hell, but 5x might be, in some circumstances.
- SushiHippie 2y ago
  I think the biggest selling point of ollama (llama.cpp) are quantizations, for a slight hit (with q8 or q4) in quality you can get a significant performance boost.
  - p1esk 2y ago
    Does ollama/llama.cpp provide low bit operations (avx or cuda kernels) to speed up inference? Or just model compression with inference still done in fp16? My understanding is the modern quantization algorithms are typically implemented in Pytorch.
    - SushiHippie 2y ago
      Sorry I don't know much about this topic. The only thing I know (from using it) that with quantization I can fit models like llama2 13b, in my 24GB of VRAM when I use q8 (16GB) instead of fp16 (26GB). This means I can get nearly the full quality of llama2 13b's output while still being able to use only my GPU, without the need to do very slow inference on only CPU+RAM. And the models are quantized before inference, so I'd only download 16GB for the llama2 13b q8 instead of the full 26GB, which means it's not done on the fly.
      - p1esk 2y ago
        As an aside, even gpt4 level quality does not feel satisfactory to me lately. I can’t imagine willingly using models as dumb as llama2-13b. What do you do with it?
        SushiHippie 2y ago
        Yeah I agree, everytime a new model releases I download the highest quantization or fp16, that fits into my VRAM, test it out with a few prompts, and then realize that downloadable models are still not as good as the closed ones (except speed wise). I don't know why I still do it, but everytime I read so many comments how good model X is, and how it outperforms anything else, and then I want to see it for myself.
- simonw 2y ago
  There's a Python binding for llama.cpp which is actively maintained and has worked well for me: https://github.com/abetlen/llama-cpp-python https://github.com/abetlen/llama-cpp-python

p1esk 2y ago

Are there benchmarks? 2x speed up would not be enough for me to return to c++ hell, but 5x might be, in some circumstances.

SushiHippie 2y ago
I think the biggest selling point of ollama (llama.cpp) are quantizations, for a slight hit (with q8 or q4) in quality you can get a significant performance boost.
- p1esk 2y ago
  Does ollama/llama.cpp provide low bit operations (avx or cuda kernels) to speed up inference? Or just model compression with inference still done in fp16? My understanding is the modern quantization algorithms are typically implemented in Pytorch.
  - SushiHippie 2y ago
    Sorry I don't know much about this topic. The only thing I know (from using it) that with quantization I can fit models like llama2 13b, in my 24GB of VRAM when I use q8 (16GB) instead of fp16 (26GB). This means I can get nearly the full quality of llama2 13b's output while still being able to use only my GPU, without the need to do very slow inference on only CPU+RAM. And the models are quantized before inference, so I'd only download 16GB for the llama2 13b q8 instead of the full 26GB, which means it's not done on the fly.
    - p1esk 2y ago
      As an aside, even gpt4 level quality does not feel satisfactory to me lately. I can’t imagine willingly using models as dumb as llama2-13b. What do you do with it?
      - SushiHippie 2y ago
        Yeah I agree, everytime a new model releases I download the highest quantization or fp16, that fits into my VRAM, test it out with a few prompts, and then realize that downloadable models are still not as good as the closed ones (except speed wise). I don't know why I still do it, but everytime I read so many comments how good model X is, and how it outperforms anything else, and then I want to see it for myself.
simonw 2y ago
There's a Python binding for llama.cpp which is actively maintained and has worked well for me: https://github.com/abetlen/llama-cpp-python https://github.com/abetlen/llama-cpp-python