search

google-research / l2p

Learning to Prompt (L2P) for Continual Learning @ CVPR22 and DualPrompt: Complementary Prompting for Rehearsal-free Continual Learning @ ECCV22
https://arxiv.org/pdf/2112.08654.pdf
Apache License 2.0
416 stars 42 forks source link

This code can not run in requirement.txt, please check again! #1

Closed miss-rain closed 2 years ago

miss-rain commented 2 years ago

This code can not run in requirement.txt,

jaxlib and cuda and cudd is fine in my ubuntu.

please check again!

KingSpencer commented 2 years ago

Hi,

Thanks for your message. Could you please provide more detailed error message so that we can figure it out easier?

miss-rain commented 2 years ago

I have follow your code,but it does not work on my GPU it only run in CPU

hh23333 commented 2 years ago

I have follow your code,but it does not work on my GPU it only run in CPU @miss-rain, hi, I meet the same problem. Have you solved it?

KingSpencer commented 2 years ago

I just checked this issue again. I think the root problem might be JAX could not find your GPU due to version mismatch. I have updated the README about how to download a correct JAX version that corresponds to your CUDA version, please have a try, thanks!

miss-rain commented 2 years ago

Thanks for update and replay, its long time(^.^)

I have solved the problem of jax for Nvidia GPU.

but now

I don't run completely in my 4 Telsa T4 (GPU), the error call 'out of memory', even i change batchsize=1

could you release a lite version?

Thanks again.

JH-LEE-KR commented 2 years ago

@miss-rain I found in the official document that when JAX executes the first JAX command, it pre-allocates 90% of the available GPU memory. As described in the document, I can either disable the pre-allocation or reduce the pre-allocation ratio to run the ViT-Base model. As a result, I was able to run on my GPUs 8 RTX 3090 and 8 A5000.