Incubator-mxnet: MXNET for cuda 11.0 Nvidia 450

[i55code]

Is there a version of MXnet for cuda 11.0, Nvidia 450 , Ubuntu 18.04? Something like pip install --pre mxnet_cu102mkl -f ... ?

Thank you!

[DickJC123]

With cuda 11.0 of course comes support for A100 and TensorFloat-32 (TF32). FYI, I'm preparing a PR that refactors how our unittests handle tolerances, so that the proper tolerances are seamlessly applied across dtypes and contexts, including A100 contexts. Stay tuned.

[i55code]

Thank you @DickJC123 !

[chrisroat]

Yes, thank you!

[tadam98]

Hi, I have installed WSL2 on windows 10 and things appear to be working.
WSL2 requites CUDA 11.
Is there a pip install mxnet-cu110 or similar for using mxnet on this new platform?

Thanks,
Mickey

[szha]

@tadam98 not yet. I will open a PR to enable this build

[tadam98]

Please do. As WSL2 is now working not bad at all with my 2080 TI super GPU and some of my stuff uses mxnet.

---

From: Sheng Zha [e-mail]
Sent: Sunday, September 13, 2020 11:04:34 PM
To: apache/incubator-mxnet [e-mail]
Cc: tadam98 [e-mail]; Comment [e-mail]
Subject: Re: [apache/incubator-mxnet] MXNET for cuda 11.0 Nvidia 450 (#18657)

not yet. I will open a PR to enable this build

—
You are receiving this because you commented.
Reply to this email directly, view it on GitHubhttps://github.com/apache/incubator-mxnet/issues/18657#issuecomment-691718222, or unsubscribehttps://github.com/notifications/unsubscribe-auth/AFBMCT3U5DUJL7XEUDLB26TSFUQVFANCNFSM4OPKGTAA.

[tadam98]

Would cuda 9.2 or 10.x work in WSL2 ?

---

From: Sheng Zha [e-mail]
Sent: Sunday, September 13, 2020 11:04:34 PM
To: apache/incubator-mxnet [e-mail]
Cc: tadam98 [e-mail]; Comment [e-mail]
Subject: Re: [apache/incubator-mxnet] MXNET for cuda 11.0 Nvidia 450 (#18657)

not yet. I will open a PR to enable this build

—
You are receiving this because you commented.
Reply to this email directly, view it on GitHubhttps://github.com/apache/incubator-mxnet/issues/18657#issuecomment-691718222, or unsubscribehttps://github.com/notifications/unsubscribe-auth/AFBMCT3U5DUJL7XEUDLB26TSFUQVFANCNFSM4OPKGTAA.

[tadam98]

Is there a way I can assist with the PR?

[szha]

@tadam98 that would be really helpful. Thanks for offering to help. The changes needed are basically the same as other cuda build variants. For example, here are all the mentions of cu102 build: https://github.com/apache/incubator-mxnet/search?q=cu102. These occurrences include changes to build logic and build config, CI/CD, and package information.

Let me know if you have any question or otherwise feel free to ping me when you open a PR.

[RuRo]

@szha Sorry to bug you, but it's been 4 months. Are there any updates?

[szha]

@RuRo sorry for lack of update on this issue. The nightly build version now has the cu110 package (for development purpose only): https://dist.mxnet.io/python

For releasing on PyPI, there are some ongoing license issue that are being sorted out.

[Joaquin-aliaga]

@szha Hi, I've installed the latest mxnet_cu11 ("mxnet_cu110-2.0.0b20201214-py3-none-manylinux2014_x86_64.whl") and the following error has happened:
AttributeError: module 'mxnet' has no attribute 'mod'

Working with:
Python 3.7.8
Debian 10
GPU: NVIDIA Tesla P100
CUDA 11

How can I solve this?
Thanks in advance

[szha]

@Joaquin-aliaga it's probably due to unclean installation. try to:

• remove all previous installation
• (if you have a local copy of mxnet source code) invoke python in a location without mxnet