CODEWITHSUNDEEP
mxnet
Aleksandar
Aleksandar

Reputation: 21

What is the version of MXNet for CUDA 12.0?

I am running a neural net of gpu cloud platform with 4 GPUs A100.When I run the code for training I am getting this error:

MX call MXNDArrayCreateEx failed... MXNetError: Compile with USE_CUDA=1 to enable GPU usage

From the support suggested couple of solutions, none of which is helping, except to check the compatibility b/w CUDA and MXNet.So typing nvidia-smi shows CUDA version is 12.0, what is the corresponding version of MXNet for it, and how to install it?

I tried reducing the batch_size and checking the GPU's memory but none of them seems to change the error.

Upvotes: 2

Views: 2282

Answers (2)

Vikas Kumar
Vikas Kumar

Reputation: 1

There is a new fork for MXNet that supports CUDA 12.6. You need the latest CUDA 12.6 SDK installed. It can also work with CUDA 12.2 if you are willing to edit the header files for some namespace mismatches.

The instructions to do this is present in the INSTALL.md file in the repository linked below:

https://github.com/selectiveintellect/modified-mxnet

Upvotes: 0

SickerDude43
SickerDude43

Reputation: 196

there is currently no MXNet version supporting CUDA 12.0 as of now. Looks like the only option is downgrading CUDA or switching over to use the CPU.

For downgrading check out this tutorial:

https://www.youtube.com/watch?v=5eJTzhGe2QE

For using the CPU simply change from mx.gpu() to mx.cpu().

Hope they'll update soon and support CUDA 12.x versions.

Upvotes: 0

Related Questions