Mxnet 1.2.1 on windows


#1

a = mx.nd.ones((3,4), ctx=mx.gpu())

this line can run on windows with mxnet 1.2.0 without any problem.
But it can NOT run on windows with mxnet 1.2.1. The program just quit without any response.

I guess the mxnet 1.2.1 does NOT support GPU on window. Both cuda 9.0 and cuda 9.2 are tested.

21-Aug-2018


#2

@zzy380 can you specify how you installed MXNet 1.2.1 ? Did you build from source or from pip install?

FYI we build and run test on every commit to master on windows and linux.

See for example this master build from today:
http://jenkins.mxnet-ci.amazon-ml.com/blue/organizations/jenkins/incubator-mxnet/detail/master/1470/pipeline


#3

pip install mxnet-cu92
If no gpu involved, the code can run.
If mx.gpu() is used, the code cannot run.

I tested the same codes. For previous versions, it is no problem. But for 1.2.1, the above mentioned problem occurs.

this line: a = mx.nd.ones((3,4), ctx=mx.gpu())
with pip install mxnet-cu92==1.2.1 (CANNOT RUN)
with pip install mxnet-cu92==1.2.0 (CAN RUN)


#4

Thanks for the details,

What version of cudnn and cuda have you installed on your machine? What is your GPU?

Can you run this script:

1. Download the diagnosis script from https://raw.githubusercontent.com/apache/incubator-mxnet/master/tools/diagnose.py
2. Run the script using `python diagnose.py` and paste its output here.

I would recommend creating an issue on https://github.com/apache/incubator-mxnet and reporting the diagnostic there.

Likely it is rare combination of your environment, hardware and libraries that create this bug. Also make sure you wipe clean previous MXNet installs before installing the latest.

Could you also try pip install mxnet-cu92 --pre and see if you get the same issue?


#5

@ThomasDelteil Thanks so much, Thomas!
pip install mxnet-cu92 --pre solves the problem. Now, the version mx.version is 1.3.0, and it can run without any problem.