WIP: Add CUDA support (proposal) #1
Conversation
|
Hi! This is the friendly conda-forge-admin automated user. I just wanted to let you know that I linted all conda-recipes in your PR ( |
|
Given the tests pass, this must mean there is an NVIDA GPU on the test machine. I tried the same thing on a machine without an NVIDA GPU and it failed to run the tests even though it was able to build successfully. |
| export CUDA_RPM_MD5="01fd3415b197f9ddc4e58f5a4f2a9af6" | ||
| curl "http://developer.download.nvidia.com/compute/cuda/repos/rhel6/x86_64/cuda-repo-rhel6-${CUDA_RPM_VERSION}.x86_64.rpm" > cuda.rpm | ||
| md5sum cuda.rpm | grep "${CUDA_RPM_MD5}" | ||
| yum install -y cuda.rpm |
There was a problem hiding this comment.
So should we be doing this as a conda package?
There was a problem hiding this comment.
That is an open question for sure. There are three options that come to mind. Some could be mixed. There could also be other variants.
- Package it and make it a dependency.
- Create a pseudo package that merely verifies the system version is present and works.
- Use the system version and document the requirement.
A related point is that GPU support probably should be an optional feature so that users can install this and get basic CPU support without GPU. However, users wanting GPU support (and possibly CUDA in particular) could have a feature package they install to enable this support.
There was a problem hiding this comment.
If we went for 3, we could use yum requirements (of course we need to add the repo though).
|
Closing as stale (2016). |
PLEASE DO NOT MERGE!
This PR attempts to build Caffe with CUDA support in addition to regular CPU support. It is far from done and will likely be broken up into a CUDA package that Caffe can then build against. Also, it does not currently contain cuDNN. The later requires registration for download. This may require some discussion with NVIDA about the best way for us to package these CUDA components and ship them. Most libraries appear to be ok by the EULA (though IANAL). However, it appears only one header is allowed. I don't know if this presents a problem for building against the libraries or not. Hopefully, this will help the discussion in this issue ( conda-forge/conda-forge.github.io#63 ) proceed by looking at a real world example.