On building the latest version (1.7.0) of mxnet with cuda 11 from source

The reason I want to build the latest version (1.7.0) of mxnet with cuda 11 from source is simply that I can’t find any official binary mxnet pip package available. However, follow the docs in the official site

(note there is no https://mxnet.apache.org/versions/1.7/get_started/build_from_source yet)
I couldn’t find clear ways or steps to successfully compile the source in a ubuntu 20.04 system with an Nvidia GPU and cuda 11 installed. The building method seems inconsistent too around different versions of mxnet. Anyone has experiences on this issue? Thanks in advance for your help!

As I know, cuda 11 is not supported now!

Am I right ?


There is a way to build mxnet: https://mxnet-tqchen.readthedocs.io/en/latest/how_to/build.html
I can’t garantee it is useful…

From my message from a developer, I knew that version 2.0 will support cuda11.
The quickest support is version 1.8 that maybe release this month…
Don’t be so optimistic about v1.8… Bugs and timeout always happen…

Thanks everyone for your responses! Looks that we’d better waiting for version 2.0 of MXNet to support cuda 11. Nevertheless, I would suggest that the owners of MXNet have a better docs on building MXNet from source. In case of TensorFlow and PyTorch, they also have no official package to support cuda 11 yet. However, follow their related docs on building from source, I can build (or find a way to build) related pip package with cuda 11.

I agree with you.
Already told the developer.