• 0 Posts
  • 9 Comments
Joined 2 years ago
cake
Cake day: July 14th, 2023

help-circle
  • Look up “LLM quantization.” The idea is that each parameter is a number; by default they use 16 bits of precision, but if you scale them into smaller sizes, you use less space and have less precision, but you still have the same parameters. There’s not much quality loss going from 16 bits to 8, but it gets more noticeable as you get lower and lower. (That said, there’s are ternary bit models being trained from scratch that use 1.58 bits per parameter and are allegedly just as good as fp16 models of the same parameter count.)

    If you’re using a 4-bit quantization, then you need about half that number in VRAM. Q4_K_M is better than Q4, but also a bit larger. Ollama generally defaults to Q4_K_M. If you can handle a higher quantization, Q6_K is generally best. If you can’t quite fit it, Q5_K_M is generally better than any other option, followed by Q5_K_S.

    For example, Llama3.3 70B, which has 70.6 billion parameters, has the following sizes for some of its quantizations:

    • q4_K_M (the default): 43 GB
    • fp16: 141 GB
    • q8: 75 GB
    • q6_K: 58 GB
    • q5_k_m: 50 GB
    • q4: 40 GB
    • q3_K_M: 34 GB
    • q2_K: 26 GB

    This is why I run a lot of Q4_K_M 70B models on two 3090s.

    Generally speaking, there’s not a perceptible quality drop going to Q6_K from 8 bit quantization (though I have heard this is less true with MoE models). Below Q6, there’s a bit of a drop between it and 5 and then 4, but the model’s still decent. Below 4-bit quantizations you can generally get better results from a smaller parameter model at a higher quantization.

    TheBloke on Huggingface has a lot of GGUF quantization repos, and most, if not all of them, have a blurb about the different quantization types and which are recommended. When Ollama.com doesn’t have a model I want, I’m generally able to find one there.


  • I recommend a used 3090, as that has 24 GB of VRAM and generally can be found for $800ish or less (at least when I last checked, in February). It’s much cheaper than a 4090 and while admittedly more expensive than the inexpensive 24GB Nvidia Tesla card (the P40?) it also has much better performance and CUDA support.

    I have dual 3090s so my performance won’t translate directly to what a single GPU would get, but it’s pretty easy to find stats on 3090 performance.








  • Giphy has a documented API that you could use. There have been bulk downloaders, but I didn’t see any that had recent activity. However you still might be able to use one to model your own script after, like https://github.com/jcpsimmons/giphy-stacks

    There were downloaders for Gfycat - gallery-dl supported it at one point - but it’s down now. However you might be able to find collections that other people downloaded and are now hosting. You could also use the Internet Archive - they have tools and APIs documented

    There’s a Tenor mass downloader that uses the Tenor API and an API key that you provide.

    Imgur has GIFs is supported by gallery-dl, so that’s an option.

    Also, read over https://github.com/simon987/awesome-datahoarding - there may be something useful for you there.

    In terms of hosting, it would depend on my user base and if I want users to be able to upload GIFs, too. If it was just my close friends, then Immich would probably be fine, but if we had people I didn’t know directly using it, I’d want a more refined solution.

    There’s Gifable, which is pretty focused, but looks like it has a pretty small following. I haven’t used it myself to see how suitable it is. If you self-host it (or something else that uses S3), note that you can use MinIO or LocalStack for the S3 container rather than using AWS directly. I’m using MinIO as part of my stack now, though for a completely different app.

    MediaCMS is another option. Less focused on GIFs but more actively developed, and intended to be used for this sort of purpose.