Typically you need about 1GB graphics RAM for each billion parameters (i.e. one byte per parameter). This is a 405B parameter model. Ouch.
Edit: you can try quantizing it. This reduces the amount of memory required per parameter to 4 bits, 2 bits or even 1 bit. As you reduce the size, the performance of the model can suffer. So in the extreme case you might be able to run this in under 64GB of graphics RAM.
https://www.ebay.com/p/116332559 lga2011 motherboards quite cheap, insert 2 xeon 2696v4 44 threads each totalling at 88 threads and 8 ddr4 32gb sticks, it comes quite cheap actually, you can also install Nvidia p40 with 24gb each, you can max out this build for ai for under 2000$
With 144Gb of total RAM, you should be able to run any CPU intensive software.
The LLMs use GPU vRAM though, so it doesn’t matter how much system RAM you have, since GPU vRAM is what the xformers and tensor scripts prioritize and have been ultimately optimized to use over CPU and RAM.
So does OSM data. Everyone can download the whole earth but to serve it and provide routing/path planning at scale takes a whole other skill and resources. It’s a good thing that they are willing to open source their model in the first place.
Technically correct ™
Before you get your hopes up: Anyone can download it, but very few will be able to actually run it.
What’s the resources requirements for the 405B model? I did some digging but couldn’t find any documentation during my cursory search.
Typically you need about 1GB graphics RAM for each billion parameters (i.e. one byte per parameter). This is a 405B parameter model. Ouch.
Edit: you can try quantizing it. This reduces the amount of memory required per parameter to 4 bits, 2 bits or even 1 bit. As you reduce the size, the performance of the model can suffer. So in the extreme case you might be able to run this in under 64GB of graphics RAM.
Or you could run it via cpu and ram at a much slower rate.
Yeah uh let me just put in my 512GB ram stick…
Samsung do make them.
Goodluck finding 512gb of VRAM.
https://www.ebay.com/p/116332559 lga2011 motherboards quite cheap, insert 2 xeon 2696v4 44 threads each totalling at 88 threads and 8 ddr4 32gb sticks, it comes quite cheap actually, you can also install Nvidia p40 with 24gb each, you can max out this build for ai for under 2000$
Finally! My dumb dumb 1TB ram server (4x E5-4640 + 32x32GB DDR3 ECC) can shine.
At work we habe a small cluster totalling around 4TB of RAM
It has 4 cooling units, a m3 of PSUs and it must take something like 30 m2 of space
When the 8 bit quants hit, you could probably lease a 128GB system on runpod.
Can you run this in a distributed manner, like with kubernetes and lots of smaller machines?
According to huggingface, you can run a 34B model using 22.4GBs of RAM max. That’s a RTX 3090 Ti.
Hmm, I probably have that much distributed across my network… maybe I should look into some way of distributing it across multiple gpu.
Frak, just counted and I only have 270gb installed. Approx 40gb more if I install some of the deprecated cards in any spare pcie slots i can find.
405b ain’t running local unless you got a proepr set up is enterpise grade lol
I think 70b is possible but I haven’t find anyone confirming it yet
Also would like to know specs on whoever did it
I’ve run quantized 70B models on CPU with 32 gigs but it is very slow
I have a home server with 140 gigs of RAM, it was surprisingly cheap. It’s an HP z6 with the 6146 gold xeon processor.
I found a seller who was selling it with a low spec silver and 16 gigs of RAM for like 250 bucks.
Found the processor upgrade for about $120 and spend another $150 on 128gb of second-hand ECC ddr4.
I think the total cost was something like $700 after throwing a couple of 8 TB hard drives in.
I’ve also placed a Nvidia 4070 in it, which I got doing some horse trading.
How close am I on the specs to being able to run the 70b version?
What’s the bus speed of the RAM? You might run it just fine but still bottlenecked there.
It’s clocked at ddr4 2666
With 144Gb of total RAM, you should be able to run any CPU intensive software.
The LLMs use GPU vRAM though, so it doesn’t matter how much system RAM you have, since GPU vRAM is what the xformers and tensor scripts prioritize and have been ultimately optimized to use over CPU and RAM.
I gonna add some RAM with hope I can split original 70b between GPU and RAM. 8b is great what it is as is
Looks like it should be possible, not sure how much performance hit offloading to RAM will do. Fafo
I regularly run llama3 70b unqantized on two P40s and CPU at like 7tokens/s. It’s usable but not very fast.
so there is no way a 24gb and 64gb can run thing?
My specs because you asked:
CPU: Intel(R) Xeon(R) E5-2699 v3 (72) @ 3.60 GHz GPU 1: NVIDIA Tesla P40 [Discrete] GPU 2: NVIDIA Tesla P40 [Discrete] GPU 3: Matrox Electronics Systems Ltd. MGA G200EH Memory: 66.75 GiB / 251.75 GiB (27%) Swap: 75.50 MiB / 40.00 GiB (0%)ok this is a server. 48gb cards and 67gb ram? for model alone?
Each card has 24GB so 48GB vram total. I use ollama it fills whatever vrams is available on both cards and runs the rest on the CPU cores.
What are you asking exactly?
What do you want to run? I assume you have a 24GB GPU and 64GB host RAM?
correct. and how ram speed work in this tbh
My memory sticks are all DDR4 with 32GB@2133MT/s.
As a general rule of thumb, you need about 1 GB per 1B parameters, so you’re looking at about 405 GB for the full size of the model.
Quantization can compress it down to 1/2 or 1/4 that, but “makes it stupider” as a result.
This would probably run on a a6000 right?
Edit: nope I think I’m off by an order of magnitude
“an order of magnitude” still feels like an understatement LOL
My 35b models come out at like Morse code speed on my 7800XT, but at least it does work?
When the RTX 9090 Ti comes, anyone who can afford it will be able to run it.
That doesn’t sound like much of a change from the situation right now.
So does OSM data. Everyone can download the whole earth but to serve it and provide routing/path planning at scale takes a whole other skill and resources. It’s a good thing that they are willing to open source their model in the first place.