Closed wohlbier closed 4 years ago
dtrudg
commented
4 years ago You probably want to conda install specifically pytorch-gpu here:
https://anaconda.org/anaconda/pytorch-gpu
From the MKL etc. dependencies coming in it appears you are install a CPU version, not a GPU version.
wohlbier
commented
4 years ago The default pytorch installs both cpu and gpu backends. Yes, it installs MKL, but also cudatoolkit, viz.
The following packages will be downloaded:
package | build
---------------------------|-----------------
_pytorch_select-0.2 | gpu_0 2 KB
blas-1.0 | mkl 6 KB
cudatoolkit-10.0.130 | 0 261.2 MB
cudnn-7.6.4 | cuda10.0_0 226.6 MB
intel-openmp-2019.4 | 243 729 KB
libgfortran-ng-7.3.0 | hdf63c60_0 1006 KB
mkl-2019.4 | 243 131.2 MB
mkl-service-2.3.0 | py37he904b0f_0 218 KB
mkl_fft-1.0.15 | py37ha843d7b_0 172 KB
mkl_random-1.1.0 | py37hd6b4f25_0 321 KB
ninja-1.9.0 | py37hfd86e86_0 1.2 MB
numpy-1.17.4 | py37hc1035e2_0 5 KB
numpy-base-1.17.4 | py37hde5b4d6_0 4.1 MB
pytorch-1.3.1 |cuda100py37h53c1284_0 294.9 MB
------------------------------------------------------------
Total: 921.7 MBAnd note the build of pytorch is the cuda100 build.
cudatoolkit from conda does not provide the nvcc compiler which is one of the things python tests for when looking for cuda. Therefore I installed my own cuda in the container, but this doesn't seem to do the trick.
dtrudg
commented
4 years ago @wohlbier - this is something I'd suggest bringing up in the community Google Group or Slack channel. You are likely to reach people who are already successfully containerizing similar things, and may get an answer faster.
https://sylabs.io/resources/community
We're happy to follow up further here, but I'm afraid it requires replicating your work to follow it through completely, and I won't be able to do that immediately.
wohlbier
commented
4 years ago Ok, I'll post there. Thanks.
wohlbier
commented
4 years ago I managed to get a working recipe. https://github.com/wohlbier/training/blob/master/object_detection/object_detection.def
Hi, I'm trying to build a container with cuda enabled pytorch that comes from conda. The system I'm building on does NOT have a cuda device. I'm building it on a RHEL box with an ubuntu container. I install cuda in the container but when I attempt to build a python package with gpu support it can not find the cuda runtime. I'm not clear on whether this should work since the box does not have a GPU.
My def file is here https://github.com/wohlbier/training/blob/master/object_detection/object_detection.def
Snippet showing cuda support not recognized. Full output below.
Thanks! jgw
Version of Singularity:
What version of Singularity are you using? Run:
What OS/distro are you running