4 ms·

Ah, that's a shame to hear. FWIW, ChatGPT did also suggest that there was a lot of room for improvement in the MPS backend of PyTorch that would likely make it

by buu700 2y ago

Ah, that's a shame to hear. FWIW, ChatGPT did also suggest that there was a lot of room for improvement in the MPS backend of PyTorch that would likely make it more efficient on Apple hardware in time.

Klaus23 2y ago
You fundamentally misunderstand the bottleneck of large LLMs. It is not really possible to make gains that way. A 405B LLM has 405 billion parameters. If you run it at full "prescision", each parameter takes up 2 bytes, which means you need 810GB of memory. If it does not fit in RAM or GPU memory it will swap to disc and be unusably slow. You can run the model at reduced prescision to save memory, called quantisation, but this will degrade the quality of the response. The exact amount of degradation depends on the task, the specific model and its size. Larger models seem to suffer slightly less. 1 byte per parameter is pretty much as good as full precision. 4 bits per parameter is still good quality, 3 bits is noticeably worse and 2 bits is often bad to unusable. With 128GB of RAM, zero overhead and a 405B model, you would have to quantize to about 2.5 bits, which would noticeably degrade the response quality. There is also model pruning, which removes parameters completely, but this is much more experimental than quantisation, also degrades response quality, and I have not seen it used that widely.
- buu700 2y ago
  I appreciate the additional information, but I'm not sure what you're claiming is a fundamental misunderstanding on my part. I was referring to running the model with quantization, and was clear that I hadn't verified the accuracy of the claims. The comment about the MPS PyTorch backend was related to performance, not whether the model would fit at all. I can't say whether it's accurate that the MPS backend has significant room for optimization, but it is still publicly listed as in beta.
  - Klaus23 2y ago
    Yes my mistake, I read your answer to mean that you think that the model could fit into the memory with the help of efficiency gains. I would be sceptical about increasing efficiency. I'm not that familiar with the subject, but as far as I know, LLMs for single users (i.e. with batch size 1) are practically always limited by the memory bandwidth. The whole LLM (if it is monolytic) has to be completely loaded from memory once for each new token (which is about 4 characters). With 400GB per second memory bandwidth and 4-bit quantisation, you are limited to 2 tokens per second, no matter how efficiently the software works. This is not unusable, but still quite slow compared to online services.
    - buu700 2y ago
      Got it, thanks, that makes sense. I was aware that memory was the primary bottleneck, but wasn't clear on the specifics of how model sizes mapped to memory requirements or the exact implications of quantization in practice. It sounds like we're pretty far from a model of this size running on any halfway common consumer hardware in a useful way, even if some high-end hardware might technically be able to initialize it in one form or another.
      - Klaus23 2y ago
        GPU memory costs about $2.5/GB on the spot market, so that is $500 for 200GB. I would speculate that it might be possible to build such a LLM card for $1-2k, but I suspect that the market for running larger LLMs locally is just too small to consider, especially now that the datacentre is so lucrative. Maybe we'll get really good LLMs on local hardware when the hype has died down a bit, memory is cheaper and the models are more efficient.
nl 2y ago
Most "local model runners" (Llama.CPP, Llama-file etc) don't use Pytorch and instead implement the neural network directly themselves optimized for whatever hardware they are supporting. For example here's the list of backends for Llama.cpp: https://github.com/ggerganov/llama.cpp?tab=readme-ov-file#supported-backends https://github.com/ggerganov/llama.cpp?tab=readme-ov-file#su...
- buu700 2y ago
  Ah okay, interesting, thanks.

by buu700 2y ago

Klaus23 2y ago

You fundamentally misunderstand the bottleneck of large LLMs. It is not really possible to make gains that way. A 405B LLM has 405 billion parameters. If you run it at full "prescision", each parameter takes up 2 bytes, which means you need 810GB of memory. If it does not fit in RAM or GPU memory it will swap to disc and be unusably slow. You can run the model at reduced prescision to save memory, called quantisation, but this will degrade the quality of the response. The exact amount of degradation depends on the task, the specific model and its size. Larger models seem to suffer slightly less. 1 byte per parameter is pretty much as good as full precision. 4 bits per parameter is still good quality, 3 bits is noticeably worse and 2 bits is often bad to unusable. With 128GB of RAM, zero overhead and a 405B model, you would have to quantize to about 2.5 bits, which would noticeably degrade the response quality. There is also model pruning, which removes parameters completely, but this is much more experimental than quantisation, also degrades response quality, and I have not seen it used that widely.

buu700 2y ago
I appreciate the additional information, but I'm not sure what you're claiming is a fundamental misunderstanding on my part. I was referring to running the model with quantization, and was clear that I hadn't verified the accuracy of the claims. The comment about the MPS PyTorch backend was related to performance, not whether the model would fit at all. I can't say whether it's accurate that the MPS backend has significant room for optimization, but it is still publicly listed as in beta.
- Klaus23 2y ago
  Yes my mistake, I read your answer to mean that you think that the model could fit into the memory with the help of efficiency gains. I would be sceptical about increasing efficiency. I'm not that familiar with the subject, but as far as I know, LLMs for single users (i.e. with batch size 1) are practically always limited by the memory bandwidth. The whole LLM (if it is monolytic) has to be completely loaded from memory once for each new token (which is about 4 characters). With 400GB per second memory bandwidth and 4-bit quantisation, you are limited to 2 tokens per second, no matter how efficiently the software works. This is not unusable, but still quite slow compared to online services.
  - buu700 2y ago
    Got it, thanks, that makes sense. I was aware that memory was the primary bottleneck, but wasn't clear on the specifics of how model sizes mapped to memory requirements or the exact implications of quantization in practice. It sounds like we're pretty far from a model of this size running on any halfway common consumer hardware in a useful way, even if some high-end hardware might technically be able to initialize it in one form or another.
    - Klaus23 2y ago
      GPU memory costs about $2.5/GB on the spot market, so that is $500 for 200GB. I would speculate that it might be possible to build such a LLM card for $1-2k, but I suspect that the market for running larger LLMs locally is just too small to consider, especially now that the datacentre is so lucrative. Maybe we'll get really good LLMs on local hardware when the hype has died down a bit, memory is cheaper and the models are more efficient.

nl 2y ago

Most "local model runners" (Llama.CPP, Llama-file etc) don't use Pytorch and instead implement the neural network directly themselves optimized for whatever hardware they are supporting. For example here's the list of backends for Llama.cpp: https://github.com/ggerganov/llama.cpp?tab=readme-ov-file#supported-backends https://github.com/ggerganov/llama.cpp?tab=readme-ov-file#su...

buu700 2y ago
Ah okay, interesting, thanks.