pytorch-0.3 with cuda and cudnn #32438

akamaus · 2017-12-08T07:32:47Z

Motivation for this change

Currently pytorch derivation lacks cuda support. After introducing cudatoolkit9 supporting gcc6 it became easy to add cuda-related flags.

Notes

Unfortunately cuda-enabled pytorch tests fail to detect driver if run by nix builder. Still they can be run manually after package installation. Just run unpackPhase and execute the script.

Had to provide a fixed commit fix hash for revision, without it getFetchFromGithub doesn't download required submodules.

FRidh · 2017-12-09T09:00:06Z

pkgs/development/python-modules/pytorch/default.nix

-  checkPhase = ''
-    ${stdenv.shell} test/run_test.sh
-  '';
+  preConfigure = if cudnn != null then


lib.optionalString

FRidh · 2017-12-09T09:00:28Z

pkgs/development/python-modules/pytorch/default.nix

+     utillinux
+     which
+  ] ++ lib.optionals cudaSupport
+        (lib.remove null [cudatoolkit cudnn]);


no need to remove null

FRidh · 2017-12-09T09:02:05Z

pkgs/top-level/python-packages.nix

+    cudnn = pkgs.cudnn;
+  };
+
+  pytorchWithCuda = self.pytorch.override {


why add extra attributes when its a matter of overriding and setting cudaSupport=true;?

akamaus · 2017-12-19T21:19:50Z

done

Fixed.

orivej · 2017-12-23T23:20:14Z

pkgs/development/python-modules/pytorch/default.nix

  '';

+  doCheck = false; # for some unknown reason doesn't detect cuda if run from builder user


Do the tests have to be disabled for pytorchWithoutCuda too?

actually, not.

tcosmo · 2018-01-13T10:43:17Z

Hey, sorry I am a huge noob with NixOs, how can I use your fork in my nixpkgs so that I can finally run pytorch with cuda like calling something like : nix-env -iA nixos.python35Packages.pytorchWithCuda
Thank you so much,

akamaus · 2018-01-14T13:51:11Z

@tcosmo, personally I just cherry-picked top two commits to channels/nixos-17.09. Now I can run pytorch like this:

nix-shell -p python3Packages.ipython -p python3Packages.pytorchWithCuda

You may find this link useful: http://anderspapitto.com/posts/2015-11-01-nixos-with-local-nixpkgs-checkout.html

FRidh · 2018-01-14T13:53:21Z

cc maintainer @teh

tcosmo · 2018-01-14T14:33:02Z

@akamaus Thank you very much for your answer!!

teh

LGTM. Thanks!

teh · 2018-01-14T17:59:29Z

pkgs/development/python-modules/pytorch/default.nix

  '';

+  doCheck = !cudaSupport; # for some unknown reason doesn't detect cuda if run from builder user


Could the check failure be caused by group permissions? I think to use cuda you need to be in the video group? In any case disabling for the cuda case with comment is OK.

Well, I'm not sure. Users working with Cuda on my machine don't belong to video group. DISPLAY variable doesn't make any difference too.
Permissions for nvidia device are as follows:

% ls /dev/nvidia0 -l
crw-rw-rw- 1 root root 195, 0 jan 14 18:35 /dev/nvidia0

andersk · 2018-04-03T15:54:34Z

This doesn’t build when merged to current master (or channels/nixos-18.03).

In file included from /nix/store/gv7w3c71jg627cpcff04yi6kwzpzjyap-cudatoolkit-9.1.85.1/include/host_config.h:50:0,
                 from /nix/store/gv7w3c71jg627cpcff04yi6kwzpzjyap-cudatoolkit-9.1.85.1/include/cuda_runtime.h:78,
                 from <command-line>:0:
/nix/store/gv7w3c71jg627cpcff04yi6kwzpzjyap-cudatoolkit-9.1.85.1/include/crt/host_config.h:121:2: error: #error -- unsupported GNU version! gcc versions later than 6 are not supported!
 #error -- unsupported GNU version! gcc versions later than 6 are not supported!
  ^~~~~

teh · 2018-04-03T21:12:43Z

Another issue: The non-cuda build seems to require cuda. I think it's down to this line:

preConfigure = lib.optionalString (cudnn != null) "export CUDNN_INCLUDE_DIR=${cudnn}/include";

teh · 2018-04-03T21:30:07Z

Another one: multiprocessig tests fail with

RuntimeError: refcounted file mapping not supported on your system at /tmp/nix-build-python3.6-pytorch-0.3.0.drv-0/source/torch/lib/TH/THAllocator.c:525

I think the optionalString has to be sth like this:

preConfigure = lib.optionalString (cudaSupport && cudnn != null) "export CUDNN_INCLUDE_DIR=${cudnn}/include";

@akamaus you still interested in updating this PR? Might be worth updating to 0.3.1 as well.

andersk · 2018-04-06T21:29:45Z

I opened a new PR #38530 based on this, addressing all the above problems.

FRidh · 2018-04-07T05:44:14Z

Closing in favor of #38530.

akamaus requested a review from FRidh as a code owner December 8, 2017 07:32

GrahamcOfBorg added 10.rebuild-darwin: 0 10.rebuild-linux: 0 labels Dec 8, 2017

FRidh previously requested changes Dec 9, 2017

View reviewed changes

FRidh added the 6.topic: python label Dec 9, 2017

akamaus force-pushed the master branch 2 times, most recently from 2ca3be7 to 8e871a1 Compare December 19, 2017 21:16

pytorch-0.3 with optional cuda and cudnn

83592fd

akamaus force-pushed the master branch from 8e871a1 to 83592fd Compare December 19, 2017 21:17

GrahamcOfBorg added 10.rebuild-linux: 1-10 and removed 10.rebuild-linux: 0 labels Dec 19, 2017

orivej reviewed Dec 23, 2017

View reviewed changes

pytorch tests reenabled if compiling without cuda

f1e1655

teh approved these changes Jan 14, 2018

View reviewed changes

andersk mentioned this pull request Apr 6, 2018

pytorch: 0.2.0 → 0.3.1 with CUDA and cuDNN #38530

Merged

8 tasks

FRidh closed this Apr 7, 2018

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

pytorch-0.3 with cuda and cudnn #32438

pytorch-0.3 with cuda and cudnn #32438

akamaus commented Dec 8, 2017 •

edited

FRidh Dec 9, 2017

FRidh Dec 9, 2017

FRidh Dec 9, 2017

akamaus commented Dec 19, 2017

orivej Dec 23, 2017

akamaus Dec 27, 2017

tcosmo commented Jan 13, 2018 •

edited

akamaus commented Jan 14, 2018 •

edited

FRidh commented Jan 14, 2018

tcosmo commented Jan 14, 2018

teh left a comment

teh Jan 14, 2018

akamaus Jan 14, 2018

andersk commented Apr 3, 2018 •

edited

teh commented Apr 3, 2018

teh commented Apr 3, 2018 •

edited

andersk commented Apr 6, 2018

FRidh commented Apr 7, 2018

		'';

		doCheck = false; # for some unknown reason doesn't detect cuda if run from builder user

		'';

		doCheck = !cudaSupport; # for some unknown reason doesn't detect cuda if run from builder user

pytorch-0.3 with cuda and cudnn #32438

pytorch-0.3 with cuda and cudnn #32438

Conversation

akamaus commented Dec 8, 2017 • edited

Motivation for this change

Notes

FRidh Dec 9, 2017

Choose a reason for hiding this comment

FRidh Dec 9, 2017

Choose a reason for hiding this comment

FRidh Dec 9, 2017

Choose a reason for hiding this comment

akamaus commented Dec 19, 2017

orivej Dec 23, 2017

Choose a reason for hiding this comment

akamaus Dec 27, 2017

Choose a reason for hiding this comment

tcosmo commented Jan 13, 2018 • edited

akamaus commented Jan 14, 2018 • edited

FRidh commented Jan 14, 2018

tcosmo commented Jan 14, 2018

teh left a comment

Choose a reason for hiding this comment

teh Jan 14, 2018

Choose a reason for hiding this comment

akamaus Jan 14, 2018

Choose a reason for hiding this comment

andersk commented Apr 3, 2018 • edited

teh commented Apr 3, 2018

teh commented Apr 3, 2018 • edited

andersk commented Apr 6, 2018

FRidh commented Apr 7, 2018

akamaus commented Dec 8, 2017 •

edited

tcosmo commented Jan 13, 2018 •

edited

akamaus commented Jan 14, 2018 •

edited

andersk commented Apr 3, 2018 •

edited

teh commented Apr 3, 2018 •

edited