LocalLLaMA

1 readers

1 users here now

Community to discuss about Llama, the family of large language models created by Meta AI.

founded 10 months ago

MODERATORS

communick@poweruser.forum

How to choose 3080 20g x 8 or 4090 x 2 to use qlora to finetune 34B models? (alien.top)

submitted 10 months ago by WitchSayo@alien.top to c/localllama@poweruser.forum

6 comments fedilink hide all child comments

Currently I have a 4090x2 and I am looking to upgrade my machine. Due to well known reasons, the price of 4090 and 3090 is insanely high right now, and I see another option: magic modding a 3080 with 20g of vram.

My aim is to use qlora to fine tune a 34B model, and I see that the requirement for fine tuning a 34B model using a single card from qlora is 24g vram, and the price of 4090x2 is about equal to 3080 20g x8. so what would be a better choice for a multi-card?

4090x2 or 3080 20g x8?

you are viewing a single comment's thread
view the rest of the comments

[–] WitchSayo@alien.top 1 points 10 months ago (1 children)

emmmm, I'm not sure, all I can do is just plug them into the motherboard.

Do you have a link to the post please? I'd like to check it out.

[–] CKtalon@alien.top 1 points 10 months ago (1 children)

Does your motherboard even have 8 PCIe slots? You’ll be needing server boards. Even workstation boards typically support only 7 (and can’t fit them all due to size)

[–] WitchSayo@alien.top 1 points 10 months ago (1 children)

Using a PCIe splitter cable, split the PCIe 4.0x16 x1 into PCIe 4.0x8 x2. And all gpu use a PCIe extension cable.

[–] The_Last_Monte@alien.top 1 points 10 months ago

I would recommend against this, not to say you are incapable o lr wouldn't figure it out, but it's a head ache. Just the power consumption on something like this alone would drive you to a sizable power bill, not to mention lack of any kind of Nvlink, your 4090 setup is a much better option in my opinion.