So far, the main consumer platform capable of running it without 'ruining' the quality of its output (with high levels of quantization) is the newer Apple Silicon Macs with unified memory - generally >=48GB. It can apparently be done on 32 or 36GB, but there's not much headroom.
Edit: As coder543 points out, yes - you can run it without more lossy levels of quantization on multi-GPU setups providing those have enough combined vram.
For the amount of money you're talking about, you could also buy two 3090s (~$750 each on eBay) and have 48GB of VRAM to run with less quantization at full speed.
M-series Macs are surprisingly flexible platforms, but they're not "the only" consumer platform that can do Mixtral.
And you're right, you can run it on a multi-GPU setup if you're so inclined.
My 14" M1 Max does around 30t/s on Mixtral Q4_K_M.
That was my experience as well - 3-bit version is pretty good.
I also tried 2-bit version, which was disappointing.
However, there is a new 2-bit approach in the works[1] (merged yesterday) which performs surprisingly well for Mixtral 8x7B Instruct with 2.10 bits per weight (12.3 GB model size).
After trying the various options for running locally, I have settled on just using Ollama - really convenient and easy, and the serve APIs let me use various LLMs in several different (mostly Lisp) programming languages.
With excellent resources from Hugging Face, tool providers, etc., I hope that the user facing interface for running LLMs is simplified even further: enter your hardware specs and get available models filtered by what runs on a user’s setup. Really, we are close to being there.
Off topic: I hope I don’t sound too lazy, but I am retired (in the last 12 years before retirement I managed a deep learning team at Capital One, worked for a while at Google and three other AI companies) and I only allocate about 2 hours a day to experiment with LLMs so I like to be efficient with my time.
Mixtral has been MLXd already? Write ups, if any?
However, there's one curious thing: llama.cpp _always_ leads to GPU throttling on Apple Silicon (e.g. the M1Max GPU will go from 1200MHz to around 700MHz), and then fully saturates it. In the rare cases I could get MLX to stay on the GPU, it was able to keep it at the maximum clock rate. However the unpredictable pauses and seemingly unoptimized prompt processing makes it hard to pick a winner in end-to-end tokens/s
https://simonwillison.net/2023/Dec/18/mistral/
I liked "Using Llamafile’s OpenAI API endpoint" described there, using Justine Tunney's llamafiles for Mixtral, but the article link is out of date, as the models have been replaced with newer: https://huggingface.co/jartine
https://github.com/ml-explore/mlx-examples/tree/main/llms/mi...
We'll have a tutorial soon (next week) on combining/composing with Reexpress to add uncertainty quantification (and to use it for semantic search). A link will be here: Tutorial 6: https://re.express/guide.html
1) He doesn't have an ax to grind / an LLM to pimp out, so he's relatively even-handed
2) He uses the same (secret) test data for each model, so his testing is resistant to cherry-picking/finetuning on tests
3) He likes weirdo role-play prompting, so he has a very good sense of the edges of refusal and alignment tuning
4) He picks up stuff well before it hits the only other fair testing I know of, the chat arena
5) I think asking stuff in German is at worst neutral, and at best useful for testing capacity in edge cases.
Practically speaking, his 'preferred' non-giant models, Nous-Capybara-34B and Mixtral both are excellent in comparison with some of the others he looks at, and good recommendations.
That said, I'd like to see a test suite that GPT-4 fails at, or struggles at, at least. And, it would save him a lot of time if he could get something automated together, it's clearly a lot of effort to hand test all those models.
For example I'm pretty sure Mistral models are better at french, so doing a benchmark using french only would be advantageous for them.
If you want to compare all models, better use english. Because now his benchmark just show which models is better at german.
That being said, it's still a very welcomed benchmark.
I agree that better ways to evaluate models would be super super useful, and benchmarks like MMLU and whatever's next will continue to be helpful ("real" science). And, it seems like there may even be some benefits for models to training to 'ace the test' more broadly, which is interesting, and ties to some educational theories in teaching humans.
However, one area that open tests can't excel in is in this "fair" evaluation arena -- and I do think that private tests have some value there, to the extent that they can show utility and maintain trust. I don't make any claims about German sex role-play being a good or bad start for these, though.
Something that has come to light since the release of the weights, and not mentioned in this paper is that it looks like fairly likely that the 8 experts were all seeded by Mistral 7B and subsequently diverged. This has generated a lot of experimentation in the local LLM community with cloning models as a way to cheaply generate experts.
It was generally thought likely that training an 8x7B network would be as much work as training 8 7B networks, but this seems not to have been true for Mistral, which is super interesting.
There's still a lot of rapid innovation happening in this space, with papers like Calm from DeepMind this week, and a lot of the adhoc experimental layer combining happening in the wild, (see, e.g. Goliath-120b), I think we're likely to see some pretty interesting architectural improvements this year in the LLM space.
Calm seems to point the way to a next step after MoE, and models like Goliath seem to indicate that even a really really lazy version of Calm (no Linear layer combination, just literally alternating layers at full weights) can be very impactful. Overall I think we will see really, really strong models that are performant on consumer hardware in 2024, likely first half of this year.