Consumer grade GPUs like NVidia's 3090 and 4090 max out at 24 GB VRAM, and those cost $1000-2000 each. You can get higher VRAM but need enterprise GPUs which are in the five figures, easily starting at $30K a pop.
Per this calculator, for training, only gpt2-large and gpt2-medium would work with those two top-of-the-line GPUs.
For inference it's certainly a bit better, only the Llama-2-70b-hf and Llama-2-13b-hf don't fit in that much VRAM, all the other models do.
Nvidia’s workstation cards are available with more RAM than the consumer cards, at a lower price than the datacenter cards. RTX 6000 Ada has 48 GB VRAM and retails for $6800, and RTX 5000 Ada has 32 GB VRAM and retails for $4000[1].
Very large models have to be distributed across multiple GPUs though, even if you’re using datacenter chips like H100s.
You can only fit 1-2 graphics cards in a “normal” ATX case (each card takes 2-3 “slots”). If you want 4 cards on one machine, you need a bigger/more expensive motherboard, case, PSU, etc. I haven’t personally seen anyone put 6 cards in a workstation.
In a water cooled config the cards only take 1 slot. I’ve got 2 3090s and am buying another two shortly. Preemtively upgraded the power to 220v, found a 2kw PSU, and installed a dedicated mini split. I’m also undervolting the cards to keep power and heat down, because even 2000w is not enough to run 4 and a server grade CPU without tripping. When you start accumulating GPUs you also run into all kinds of thermal and power problems for the room, too.
I was fortunate enough to scoop up a bunch of Gigabyte RTX 3090 Turbos. Cheap used eight slot SuperMicro (or whatever), a cabling kit, four 3090s, boot.
Sincere question: Is installing and running a mini split actually cheaper than racking them in a colo, or paying for time on one of the GPU cloud providers?
Regardless, I can understand the hobby value of running that kind of rig at home.
I personally haven’t done the calculation. I have rented colo space before and they are usually quite stingy on power. The other issue is, there’s a certain element to having GPUs around 24/7/365 to play with that I feel is fundamentally different than running it on a Cloud Provider. You’re not stressing out about every hour it’s running. I think in the long run (2yr+) it will be cheaper, and then you can swap in the latest and greatest GPU without any additional infrastructure cost.
You have to pass the context between GPUs for large models that don't fit in VRAM. Often ends up slower. Also, tooling around AMD GPUs is still poor in comparison.
Used 3090 are going for ~600usd these days (at least in Europe) thanks to crypto mining crash - building a workstation with 2 of these is fairly easy for 48GB of vram, with 4 a bit more tricky but still doable and affordable IMO
Recently bought 24GB 3090 for $700 in US. Used but never tweaked, runs stable for 6 months despite heavy workloads.
nVidias play seems obvious. Game graphics don’t move that fast these days. Used market flush with 3090s and down is fine to them while they focus on extracting top dollar from fast moving AI researchers/VCs
They seem to be broken when I try any HF ids besides what came preconfigured. e.g. just tried brucethemoose/Yi-34B-200K-DARE-merge-v5-3.1bpw-exl2-fiction or LoneStriker/shisa-7b-v1-3.0bpw-h6-exl2
I noticed the default parameter count value is 1.418 billion but if you erase it you can't actually enter it back because you can't type a decimal point in the input area. Also, you can't enter parameter counts smaller than 1 billion
Are people still rawdoggin' 16-bit models? I almost exclusively use 5-bit inference quants (or 8-bit natives like Yi-34b) on my MacBook Pro. Tiny accuracy loss, runs fast, and leave plenty of (V)RAM on the table. Mixtral 8x7 is my new daily driver, and only takes like 40GB to run! I wonder if I could run two of them talking to each other...
Mixed precision is a default method to pretrain and full fine tune right now. It is especially good in transformers, because they have memory bottleneck in activations (outputs of intermediate layers stored for backprop), and running forward pass in fp16/bf16 reduces VRAM by almost half (speeds up forward pass as well).
I wonder about that too. With the small precision, parameter updates might be too small to have an effect (is it possible to use some sort of probabilistic update in that case?) Unfortunately, I haven’t found any resources describing the feasibility of full fp16 or bf16 training.
You are correct, training sorely in fp16/bf16 can lead to imprecise weight updates or even gradients turning to zero. Because of that, mixed precision is used. In mixed precision training, we keep a copy of the weights in fp32 (master model) and the training loop looks like this:
compute the output with the fp16 model, then the loss
-> back-propagate the gradients in half-precision
-> copy the gradients in fp32 precision
-> do the update on the master model (in fp32 precision)
-> copy the master model in the fp16 model.
We also do loss scaling which means multiplying the output of the loss function by some scalar number before backprop (necessary in fp16 but not required in bf16).
Usually, for efficiency, you use quantized models. Quantized models reduce the number of bits available for each parameter, saving space and reduce RAM usage.
