simply pytorch install by using pytorch-nightly instead of source

[soumith]

simply pytorch install by using pytorch-nightly instead of source

[jmp84]

Patching and trying instructions again (or did you already try?).

[soumith]

I haven't tried.

[soumith]

if you are using cuda9, it is: conda install pytorch-nightly cuda90 -c pytorch

[jmp84]

Is this nightly package also including Caffe2? (@pjh5)
@jhcross, getting seg fault when running export example with this new install:

wget https://repo.continuum.io/miniconda/Miniconda3-latest-Linux-x86_64.sh -O miniconda.sh
chmod +x miniconda.sh
./miniconda.sh -b -p ~/miniconda
rm miniconda.sh
. ~/miniconda/bin/activate

conda install pytorch-nightly -c pytorch

conda install -y cmake

git clone --recursive https://github.com/onnx/onnx.git
yes | pip install ./onnx 2>&1 | tee ONNX_OUT

pip uninstall -y pytorch-translate
python3 setup.py build develop

[theweiho]

You can also try seeing if

conda uninstall -y caffe2 caffe2-cuda8.0-cudnn7
conda install -y -c caffe2 caffe2-cuda8.0-cudnn7

before trying to install ONNX helps

[jmp84]

@theweiho, same issue.

[pjh5]

@jmp84 where's the segfault?
pytorch-nightly does not include Caffe2
It is safe to clone Pytorch and only build Caffe2 from it and get the Pytorch installation from conda; this will save all the time needed to build Pytorch. You can't do this with the Caffe2 conda packages because they are not built with Aten. The pytorch-caffe2_cuda8.0_cudnn7 builds in the caffe2 conda channel are built with Aten, and also include Pytorch, but the Pytorch included from these builds will be slightly different than the one from pytorch-nightly (though these differences should not affect you). jsyk the names of these Pytorch+Caffe2 packages will probably change within the next day.

[theweiho]

@pjh5 - you're suggesting that we should be able to use the new combined pytorch-caffe2_cuda8.0_cudnn7 conda builds now? (Once package names settle) That'd be awesome!

@jhcross (and maybe @jamesr66a ?) may have more info for what the ONNX compatibility/segfault issue we ran into was (so we can confirm whether that'd be resolved with the new combined PyTorch+Caffe2 packages).

[theweiho]

@anderspapitto @bddppq - James Cross mentioned y'all helped him debug the ONNX compatibility/segfault issue - and that it may have been due to differences in gcc compiler versions or something?

[anderspapitto]

we got something to work, but there wasn't full clarity into exactly what the issue had been.

If you're able to make changes and things work, then you're probably fine - the crashes we were running into were very upfront

[jmp84]

Fixed in 7ba9dfc by @theweiho