Package Base Details: caffe2

Git Clone URL: https://aur.archlinux.org/caffe2.git (read-only, click to copy)
Keywords: ai artificial cuda intelligence nvidia
Submitter: dbermond
Maintainer: dbermond
Last Packager: dbermond
Votes: 0
Popularity: 0.000000
First Submitted: 2017-04-29 16:21
Last Updated: 2019-06-29 19:14

Packages (2)

Pinned Comments

dbermond commented on 2018-08-22 18:15

Important notice:

This package now provides the non-cuda version (as known as the 'cpu only' build).

If you want caffe2 with cuda support, use package caffe2-cuda. This new package naming schema will better reflect the package contents. This conforms to tensorflow package naming from the official repositories.

Latest Comments

« First ‹ Previous 1 2 3 4 Next › Last »

dbermond commented on 2017-09-09 17:09

It seems that CUDA/NVCC 8.0.61 is not compatible with glibc 2.26. As a result, currently this package cannot be compiled with gpu support as it should be.
Use caffe2-cpu instead (cpu-only, without gpu support) until there is a repository or upstream fix.
Please leave a reply if you find a workaround.

dbermond commented on 2017-09-09 12:40

@petronny
Hi. Thank you for reporting this.
I could reproduce the issue and identified the cause as being the gloo compilation, and not nccl directly. Compiling without gloo support (-DUSE_GLOO:BOOL='OFF') solves the issue. Updating the gloo commit to latest git master also solves the issue, since upstream gloo removed nccl from its third party dependencies.

But another issue has just appeared: both caffe2 and gloo fails to compile with the newly released glibc 2.26. Since this glibc 2.26 issue is an unsolved upstream bug, I will update the package only after it's being solved, since it will not compile anyway for now.

petronny commented on 2017-09-08 14:58

I get
-- Found OpenMP_C: -fopenmp (found version "4.0")
-- Found OpenMP_CXX: -fopenmp (found version "4.0")
-- Adding -fopenmp
-- CUDA detected: 8.0
-- Automatic GPU detection failed. Building for all known architectures.
-- Added CUDA NVCC flags for: sm_20 sm_21 sm_30 sm_35 sm_50 sm_52 sm_60 sm_61
-- Found libcuda: /opt/cuda/lib64/stubs/libcuda.so
-- Found libnvrtc: /opt/cuda/lib64/libnvrtc.so
-- Found CUDNN: /opt/cuda/include
-- Found cuDNN: v7.0.1 (include: /opt/cuda/include, library: /opt/cuda/lib64/libcudnn.so)
-- Could NOT find CUB (missing: CUB_INCLUDE_DIR)
-- Could NOT find Gloo (missing: Gloo_INCLUDE_DIR Gloo_LIBRARY)
-- Found hiredis: /usr/include/hiredis
-- Found hiredis (include: /usr/include/hiredis, library: /lib/libhiredis.so)
-- MPI include path: /usr/include
-- MPI libraries: /usr/lib/openmpi/libmpi_cxx.so/usr/lib/openmpi/libmpi.so
-- Found CUDA: /opt/cuda (found suitable version "8.0", minimum required is "7.0")
-- CUDA detected: 8.0
-- Found libcuda: /opt/cuda/lib64/stubs/libcuda.so
-- Found libnvrtc: /opt/cuda/lib64/libnvrtc.so
CMake Error at /usr/share/cmake-3.9/Modules/ExternalProject.cmake:2010 (message):
No download info given for 'nccl_external' and its source directory:

/build/caffe2/src/caffe2-0.8.1/third_party/gloo/third-party/nccl

is not an existing non-empty directory. Please specify one of:

* SOURCE_DIR with an existing non-empty directory
* URL
* GIT_REPOSITORY
* HG_REPOSITORY
* CVS_REPOSITORY and CVS_MODULE
* SVN_REVISION
* DOWNLOAD_COMMAND
Call Stack (most recent call first):
/usr/share/cmake-3.9/Modules/ExternalProject.cmake:2565 (_ep_add_download_command)
third_party/gloo/cmake/External/nccl.cmake:16 (ExternalProject_Add)
third_party/gloo/cmake/Dependencies.cmake:53 (include)
third_party/gloo/CMakeLists.txt:44 (include)

without nccl installed.

Adding nccl back to depends works.

dbermond commented on 2017-08-22 20:20

@wangqr
Thank you for reporting this. I can reproduce the issue.
It seems that the caffe2 python dependencies have slightly changed from when I firstly checked it.
I will be updating this package and the other caffe2 packages on the AUR to fix this issue.

wangqr commented on 2017-08-22 08:34

After installed this package I tried to import it, but get the following error:

Python 2.7.13 (default, Jul 21 2017, 03:24:34)
[GCC 7.1.1 20170630] on linux2
Type "help", "copyright", "credits" or "license" for more information.
>>> from caffe2.python import workspace, model_helper
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
File "/usr/lib/python2.7/site-packages/caffe2/python/workspace.py", line 14, i n <module>
from past.builtins import basestring
ImportError: No module named past.builtins
>>>

Installing python2-future resolves the issue.

dbermond commented on 2017-08-09 17:25

caffe2 0.8.1 adds support for the newly released cudnn 7.0.

dbermond commented on 2017-08-08 10:31

@petronny
I have already "done something":
https://aur.archlinux.org/cgit/aur.git/commit/?h=caffe2&id=894fb0890375500b0629e13ff2ad457244eaf68f

That's the workaround that I could manage to do while a proper patch cannot be applied. This will require downgrade of cudnn to 6.0.21.
In the meanwhile you can use caffe2-cpu, which does not have cudnn dependency.

petronny commented on 2017-08-08 07:09

Hi, the cudnn in [community] has been updated to 7.0.1
Please do something. The package won't be built now.

pdrocaldeira commented on 2017-07-24 22:05

@dbermond

Thank YOU!

Worked just fine. :)

dbermond commented on 2017-07-21 01:57

@pdrocaldeira
Thank you for reporting it. This should be fixed as of version 0.7.0-13.
All other caffe2 packages were fixed too.