Incubator-mxnet: CUDA 10 Support and Testing

[sbodenstein]

Now that CUDA 10 and Turing GPUs are out, what are the plans for adding testing for CUDA 10 builds?

For example, we've had breakages with building MXNet with CUDA 10 on OSX (eg MXImperativeInvoke seems to cause corruption of NDArrays).

[lanking520]

@sbodenstein thanks for your issue.
@szha @marcoabreu @larroy @lebeg to help you out.
@mxnet-label-bot please add [CUDA, CI, test] here

[szha]

The code for cuda 10 pip release is ready. We need to request to increase the size limit of mxnet-cu100/mxnet-cu100mkl pip packages to 600mb. @dstufft used to help us with these requests but I haven't heard from him yet. We may need to request this through github issues on pypa repo this time but I don't have enough time to drive this right now.

[universewill]

any update about this?
Can I install mxnet from GitHub source code with cuda10?
@szha @dstufft

[tribbloid]

looks like this pypi is slow to respond, is there another release channel we can use?

[szha]

Unfortunately, no...

[larroy]

s3?

[lebeg]

https://github.com/apache/incubator-mxnet/pull/12850

[gerkulesov]

Are there any news about mxnet-cu100 pip release?

[szha]

@yajiedesign opened an issue and it's currently sitting in https://github.com/pypa/warehouse/issues/4986
Given that, I'm going to start publishing the build artifacts also to s3 bucket, so that at least people can get access while we resolve the size limit.

[UnclaimedPants]

@szha Can you please post the link to the artifacts in s3 when they're ready?

[szha]

@UnclaimedPants will do

[szha]

I've been working towards an event recently so unfortunately, I haven't got to this yet. I will try to find some cycles on this around 12/3

[gerkulesov]

Upload limit for mxnet-cu100 and mxnet-cu100mkl was increased on PyPI!

[szha]

CUDA10 versions are being published now.