Incubator-mxnet: R loadModule fails (Windows, GPU, mxnet v1.7.x)

[howff]

Description

Whilst building the R package it attempts to perform a test load which fails. The call to loadModule("mxnet") in R-package/R/zzz.R fails, ERROR: loading failed is reported and the R package cannot be loaded into R.

Error Message

installing to C:\blah...
** R
** demo
** inst
** byte-compile and prepare package for lazy loading
** help
** installing vignettes
** testing if installed package can be loaded from temporary location
ERROR: loading failed

To Reproduce

I have checked out the source code to mxnet branch 1.7.x (because that's marked as current release).
The mxnet libraries build successfully (libmxnet.dll, etc).
The R-package builds successfully (mxnet.dll).
The installation fails to load the module.

Steps to reproduce

R CMD INSTALL --no-multiarch R-package

What have you tried to solve it?

1. I have tried both versions of R; R-3.6.3 and R-4.0.2
2. I have inserted print statements into R/zzz.R to determine that the fault is in loadModule("mxnet")
3. I have determined that the DLLs appear to load as library.dynam(libmxnet) and (mxnet) succeed.

Environment

Windows 10, R-4.0.2 and rtools40 AND R-3.6.3 and rtools35, nvidia CUDA 10.2, secondary GPU (GTX1050), openblas

[github-actions[bot]]

Welcome to Apache MXNet (incubating)! We are on a mission to democratize AI, and we are glad that you are contributing to it by opening this issue.
Please make sure to include all the relevant context, and one of the @apache/mxnet-committers will be here shortly.
If you are interested in contributing to our project, let us know! Also, be sure to check out our guide on contributing to MXNet and our development guides wiki.

[szha]

cc @hetong007 for R and @yajiedesign for windows build.

[hetong007]

For R installation on Windows, please refer to this guide: https://github.com/apache/incubator-mxnet/blob/v1.x/docs/static_site/src/pages/get_started/windows_setup.md#install-the-mxnet-package-for-r

[howff]

@hetong007 Thanks for the link.

I've come to github as a last resort having explored all the documentation, blogs and forum posts over the last two weeks trying to get this installed. The link has information which is very out of date. R 3.5 has been superceded twice, but 3.6 and now by 4.0. The pre-built libraries haven't been updated for 2 years. So I'm building everything from source and it's not working.

[jeremiedb]

The pre-built binaries for Windows are no longer generated by https://github.com/yajiedesign/mxnet/releases from what I can see.
After latest built I made was for R 3.6 Mxnet 1.5. if you need a more recent, Mxnet shall be built from source, which I also had issues with at the time, but haven't tried recently. If building from source works, I think the R package component shouldn't be an issue.
More recently, I've been running from Linux so avoided the windows build difficulty.

[szha]

@jeremiedb thanks for elaborating. @yajiedesign is currently working on moving the windows build pipeline to our CD, pending on https://github.com/apache/incubator-mxnet/issues/18793. Once that's completed, we can also add R build pipeline based on it.

[jeremiedb]

@howff In the meantime, although it's not latest mxnet and CUDA version, I've built a GPU version for MXNET 1.5.1 on CUDA 10.1 / cuDNN 7.6.4 for R 4.0.

Can be installed with:
install.packages("https://jeremiedb.s3.ca-central-1.amazonaws.com/share/mxnet/GPU/1.5.1/R4.0/mxnet.zip", repos = NULL)