A brief history of LLaMA models (opens in new tab)

What ever happened to the crazy notion of Dockerfiles that simply build successfully?

Isn’t half the point of containerisation that it papers over the madness of the Python module ecosystem?

Astonishing.

I've heard several people say that it is easy, but then surely it ought to be trivial to set script the build so that it works reliable in a container!

The only recourse is using the -bin flavors of PyTorch, etc. which will just download the precompiled upstream versions. Sadly, the result will still be much slower than other distributions. First because Python isn't compiled with optimizations and LTO in nixpkgs by default, because it is not reproducible. So, you override the Python derivation to enable optimizations and LTO. Python builds fine, but to get the machine learning ecosystem on you machine, Nix needs to build a gazillion Python packages, since the derivation hash of Python changed. Turns out that many derivations don't actually build. They build with the little amount of parallelism available on Hydra builders, but many Python packages will fail to build because of concurrency issues in tests that do manifest on your nice 16 core machine.

So, you spend hours fixing derivations so that they build on many core machines and upstream all the diffs. Or YOLO and you disable unit tests altogether. A few hours/days later (depending on your knowledge of Nix), you finally have a built of all packages that you want, you launch whatever you are doing on your CUDA-capable GPU. Turns out that it is 30-50% slower. Finding out why is another multi-day expedition in profiling and tinkering.

In the end pyenv (or a Docker container) on a boring distribution doesn't look so bad.

(Disclaimer: I initially added the PyTorch/libtorch bin packages to nixpkgs and was co-maintainer of the PyTorch derivation for a while.)

Literally every example I've seen so far is completely unversioned and mere weeks after being written simply doesn't work as a direct consequence.

E.g: https://github.com/oobabooga/text-generation-webui/blob/ee68...

Take this line:

    pip3 install torch torchvision torchaudio

    FROM nvidia/cuda:11.8.0-runtime-ubuntu22.04

Check out "requirements.txt":

    accelerate==0.18.0
    colorama
    datasets
    flexgen==0.1.7
    gradio==3.25.0
    markdown
    numpy
    pandas
    Pillow>=9.5.0
    pyyaml
    requests
    rwkv==0.7.3
    safetensors==0.3.0
    sentencepiece
    tqdm

Then this beauty:

    git+https://github.com/huggingface/peft

The NPM ecosystem gets regularly excoriated for the exact same mistakes, which by now are so well known, so often warned against, so often exploited, so regularly broken that it's getting boring.

It's like SQL injection. If you're still doing it in 2023, if your site is still getting hacked because of it, then you absolutely deserve to be labelled immature and even childish.

It was discovered though, that while models may need this level of precision when creating them ("training"), they don't need it nearly as much after the fact, when simply running them to get results ("inference").

So quantisation is the process of getting that big set of, say, 32-bit floats, and "mapping" them to a much smaller number type. Eg, an 8-bit integer ("INT8"). This is a number in the range 0-255 (or -128 to +127).

So, to quantise a list of 32-bit floats, you could go through the list and analyse. Maybe they're all in the range -1.0 to +1.0. Maybe there are many around the value of 0.99999 and 0.998 etc, so you decide to assign those the value "255" instead.

Repeat this until you've squashed that bunch of 32-bit values into 8-bits each. (Eg, maybe 0.750000 could be 192, etc.)

This could give a saving in memory footprint for the model of 4x smaller, and also makes it able to be run faster. So while you needed 16GB to run it before, now you might only need 4GB.

The expense is the model won't be as accurate. But, typically this is on the order of values like 90%, versus the memory savings of 4x. So it's deemed worth it.

It's through this process folks can run models that would normally require a 5-figure GPU to run, on their home machine, or even on the CPU, as it might be able to process integers easier and faster than floating point.

https://huggingface.co/docs/optimum/concept_guides/quantizat...

I'm not sure how many of these models are actively taking advantage of that architecture yet though.

The tough thing to find is something affordable that will run the unquantized 65B model at an acceptable speed. You can put 128GB of RAM in affordable hardware but ordinary desktops aren't fast. The things that are fast are expensive (e.g. I bet Epyc 9000 series would do great). And that's the thing Apple doesn't get you either, because Apple Silicon isn't available with that much RAM, and if it was it wouldn't be affordable (the 96GB Macbook Pro, which isn't enough to run the full model, is >$4000).