Upgrade from CUDA 10.2 to 11.0 ?

[lowener]

Is it possible to upgrade the notebook Cuda version to 11.0 ? This version is fixing bugs in svd:
cupy/cupy#2351

Thanks

[rosbo]

Hi @lowener,

This is on our radar. We are waiting until all our GPU-enabled packages support cuda 11.

Thank you

[claverru]

Any updates on this?

[rosbo]

@claverru we are waiting on mxnet to release their GA version for CUDA 11.0: apache/mxnet#18657

[toshi-k]

I think it's the time to install CUDA 11 (by removing mxnet at once and add it again after the update for CUDA11).
The number of users of RTX 30X0 GPUs / A6000 / A100 would be more than the users of mxnet in Kaggle.

[djherbis]

@toshi-k Are there a lot of users using this image with those GPUs?

I was under the assumption that the primary consumer of these images is Kaggle Notebooks, and that very few people actually use this image outside Kaggle.

[toshi-k]

@djherbis Basically, there are two ways of submission in Kaggle, CSV based competition, and code competition (2stage competition).

In code competition, all participants have to submit inference notebook through Kaggle Notebooks. RTX 30X0 users train the model on own machine (with CUDA11), and submit their model by Kaggle Notebooks (with CUDA10). The difference of CUDA version sometimes causes issues, which seems more crucial than the availability of mxnet.

[djherbis]

I see thanks for the clarification!

[rosbo]

Hi @toshi-k,

The mxnet wheel for CUDA 11 was released last week. I will start working on upgrading our image to CUDA 11.

[toshi-k]

@rosbo That's good news ! I'm looking forward to the upgrade of GPU image.

[rosbo]

I upgraded CUDA to version 11.0 in #985.

We will release a new Docker image to Kaggle notebooks next week.

Thank you for your patience