Out of memory error -- Nvidia A6000 48GB ~2300 aa residues

[rjrich]

Getting out of memory errors with Boltz-1 version 0.3.2 when predicting protein with ~2300 residues. My GPU has 48GB memory; my CPU RAM is 256GB. I tried downloading a fresh cache, but afterward I got the same out of memory error. I have 2 x Nvidia A6000 GPUs -- is it possible to use both of them for the prediction? When I tried setting devices to 2, I got an error saying that the number of devices was greater than the number of predictions. Thanks for any assistance you can provide and thanks for making boltz-1 available.

[gcorso]

Hi @rjrich, no unfortunately at the moment sharding a single prediction across GPUs is not yet supported and unfortunately 2300 residues might be too much for a 48GB GPU at the moment. We will try to improve the memory efficiency further in the future!

[rjrich]

Thanks for your response. Is it possible to designate a specific GPU? I have 2 x Nvidia A6000, and I would like to specify the visible device 1 rather than visible device 0. Thanks again.

On Wed, Dec 18, 2024 at 11:19 AM Gabriele Corso @.***> wrote:

Hi @rjrich https://github.com/rjrich, no unfortunately at the moment sharding a single prediction across GPUs is not yet supported and unfortunately 2300 residues might be too much for a 48GB GPU at the moment. We will try to improve the memory efficiency further in the future!

— Reply to this email directly, view it on GitHub https://github.com/jwohlwend/boltz/issues/106#issuecomment-2551749927, or unsubscribe https://github.com/notifications/unsubscribe-auth/AAP6T7K3Q4WDXVG7AVVODTL2GGOCPAVCNFSM6AAAAABTVAOEFKVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDKNJRG42DSOJSG4 . You are receiving this because you were mentioned.Message ID: @.***>

[jwohlwend]

You can just do CUDA_VISIBLE_DEVICES=1 boltz predict ...

[rjrich]

Thanks!

On Wed, Dec 18, 2024 at 12:12 PM Jeremy Wohlwend @.***> wrote:

You can just do CUDA_VISIBLE_DEVICES=1 boltz predict ...

— Reply to this email directly, view it on GitHub https://github.com/jwohlwend/boltz/issues/106#issuecomment-2551863744, or unsubscribe https://github.com/notifications/unsubscribe-auth/AAP6T7LC3URYGLJCT4SB7ZT2GGUGDAVCNFSM6AAAAABTVAOEFKVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDKNJRHA3DGNZUGQ . You are receiving this because you were mentioned.Message ID: @.***>

[ghost]

Hi, I am having the same out of memory problem. How much memory do I need for a homotrimer with 500 AA per monomer?

[gcorso]

I would say probably something between 40 and 50 GB

[paloha]

Hi all, thanks for Boltz! I have protein sequences for 17 subunits of my structure that I want to fold, totaling around 6300 amino acids. I’m considering using Boltz-1 on CPU with RAM, as even GPUs with 80GB VRAM are not sufficient to handle this scale. I have around 512GB of RAM available. Is it feasible to run this on CPU, naturally sacrificing speed, and would this much memory be enough for folding a structure of this size? Is there any table where one can look up the scaling? Thanks a lot.