Comment on Smaug-72B-v0.1: The New Open-Source LLM Roaring to the Top of the Leaderboard
simple@lemm.ee 4 months ago
I’m afraid to even ask for the minimum specs on this thing, open source models have gotten so big lately
CUDA 11.4 and above are recommended (this is for GPU users, flash-attention users, etc.) To run Qwen-72B-Chat in bf16/fp16, at least 144GB GPU memory is required (e.g., 2xA100-80G or 5xV100-32G). To run it in int4, at least 48GB GPU memory is requred (e.g., 1xA100-80G or 2xV100-32G).
It’s derived from Qwen-72B, so same specs. Q2 clocks it in at only ~30GB.
Just a data center or two. Easy peasy dirt cheapy.
I think I read somewhere that you’ll basically need 130 GB of RAM to load this model. You could probably get some used server hardware for less than $600 to run this.
Oh if only it were so simple lmao, you need ~130GB of VRAM, aka the graphics card RAM. So you would need about 9 consumer grade 16GB graphics cards and you’ll probably need Nvidia because of fucking CUDA so we’re talking about thousands of dollars. Probably approaching 10k
Ofc you can get cards with more VRAM per card, but not in the consumer segment so even more $$$$$$
Afaik you can substitute VRAM with RAM at the cost of speed. Not exactly sure how that speed loss correlates to the sheer size of these models, though. I have to imagine it would run insanely slow on a CPU.
I tested it with a 16GB model and barely got 1 token per second. I don't want to imagine how long it would take if I used swap instead.
I’m pretty sure you can load the model using RAM like another poster said. Here’s a used server under $600 that could theoretically run it: ebay.
You would want to look for an R730, which can be had for not too much more. The 20 series was the “end of an era” and the 30 series was the beginning of the next era. Most importantly for this application, R30s use DDR4 whereas R20s use DDR3.
RAM speed matters a lot for ML applications and DDR4 is about 2x as fast as DDR3 in all relevant measurements.
If you’re going to offload any part of these models to CPU, which you 99.99% will have to do for a model of this size with this class of hardware, skip the 20s and go to the 30s.
Unless you’re getting used datacenter grade hardware for next to free, I doubt this. You need 130 gb of VRAM on your GPUs
So can I run it on my Radeon RX 5700? I overclocked it some and am running it as a 5700 XT, if that helps.
To run this model locally at gpt4 writing speed you need at least 2 x 3090 or 2 x 7900xtx. VRAM is the limiting factor in 99% of cases for interference. You could try a smaller model like mistral-instruct or SOLAR with your hardware though.
Around 48gb of VRAM if you want to run it in 4bits
TheChurn@kbin.social 4 months ago
Every billion parameters needs about 2 GB of VRAM - if using bfloat16 representation. 16 bits per parameter, 8 bits per byte -> 2 bytes per parameter.
1 billion parameters ~ 2 Billion bytes ~ 2 GB.
From the name, this model has 72 Billion parameters, so ~144 GB of VRAM
nicetriangle@kbin.social 4 months ago
Ok but will this run on my TI-83? It's a + model.
Rai@lemmy.dbzer0.com 4 months ago
Only if it’s silver.
01189998819991197253@infosec.pub 4 months ago
Dang. So close.
whoelectroplateuntil@sh.itjust.works 4 months ago
no. but put this clustering software i wrote in ti-basic on 40 million of them? still no
FaceDeer@kbin.social 4 months ago
It's been discovered that you can reduce the bits per parameter down to 4 or 5 and still get good results. Just saw a paper this morning describing a technique to get down to 2.5 bits per parameter, even, and apparently it 's fine. We'll see if that works out in practice I guess
Corngood@lemmy.ml 4 months ago
I’m more experienced with graphics than LM, but wouldn’t that cause a significant increase in computation time, since those aren’t native types for arithmetic? Maybe that’s not a big problem?
If you have a link for the paper I’d like to check it out.
FaceDeer@kbin.social 4 months ago
My understanding is that the bottleneck for the GPU is moving data into and out of it, not the processing of the data once it's in there. So if you can get the whole model crammed into VRAM it's still faster even if you have to do some extra work unpacking and repacking it during processing time.
The paper was posted on /r/localLLaMA.
L_Acacia@lemmy.one 4 months ago
You can take a look at exllama and llama.cpp source code on github if you want to see how it is implemented.
Amaltheamannen@lemmy.ml 4 months ago
Though with quantisatiom you can get it down to like 30GB of vram or less.
rs137@lemmy.world 4 months ago
Llama 2 70B with 8b quantization takes around 80GB VRAM if I remember correctly. I’ve tested it a while ago.
OutrageousUmpire@lemmy.world 4 months ago
Any idea what 8Q requirements would be? Or 4 or 5?
General_Effort@lemmy.world 4 months ago
huggingface.co/senseable/…/main
About 44GB and 50GB for the Q4 and 5. You’d need quite some extra to fully use the 32k context length.