The current ebuild for sci-libs/pytorch-2.0.0 does not offer all the build functionality that the ebuild for sci-libs/caffe2-2.0.0-r3 does or that can be enabled if built directly from source code. Can the ebuild be updated to provide the ability to enable the same options as if it were built directly from the Source Code? This would enable it to honor our use flags and emable the different backends such as opencl, vulkan, opencv etc.
The pytorch and caffe2 configuration are strictly related: caffe2 save the configuration that pytorch is using. So, apart from error, the functionality of caffe2 ebuild are the same of pytorch. If you find some that are not, please raise a bug or, better, provide patches. For the things that are not enabled, please provide patch or a bug report telling me what dependencies they need. As I don't have most of the backends, a patch / pull request of a tested ebuild is good for me.
The caffe2 ebuild has the following listed below as it builds from the C code using cmake whereas the pytorch has none of these options enabled in the ebuild and does not seem to use cmake. ---- local mycmakeargs=( -DBUILD_CUSTOM_PROTOBUF=OFF -DBUILD_SHARED_LIBS=ON -DUSE_CCACHE=OFF -DUSE_CUDA=$(usex cuda) -DUSE_CUDNN=$(usex cuda) -DUSE_FAST_NVCC=$(usex cuda) -DTORCH_CUDA_ARCH_LIST="${TORCH_CUDA_ARCH_LIST:-3.5 7.0}" -DBUILD_NVFUSER=OFF -DUSE_DISTRIBUTED=$(usex distributed) -DUSE_MPI=$(usex mpi) -DUSE_FAKELOWP=OFF -DUSE_FBGEMM=$(usex fbgemm) -DUSE_FFMPEG=$(usex ffmpeg) -DUSE_GFLAGS=ON -DUSE_GLOG=ON -DUSE_GLOO=$(usex gloo) -DUSE_KINETO=OFF # TODO -DUSE_LEVELDB=OFF -DUSE_MAGMA=OFF # TODO: In GURU as sci-libs/magma -DUSE_MKLDNN=OFF -DUSE_NCCL=OFF # TODO: NVIDIA Collective Communication Library -DUSE_NNPACK=$(usex nnpack) -DUSE_QNNPACK=$(usex qnnpack) -DUSE_XNNPACK=$(usex xnnpack) -DUSE_SYSTEM_XNNPACK=$(usex xnnpack) -DUSE_TENSORPIPE=$(usex tensorpipe) -DUSE_PYTORCH_QNNPACK=OFF -DUSE_NUMPY=$(usex numpy) -DUSE_OPENCL=$(usex opencl) -DUSE_OPENCV=$(usex opencv) -DUSE_OPENMP=$(usex openmp) -DUSE_ROCM=OFF # TODO -DUSE_SYSTEM_CPUINFO=ON -DUSE_SYSTEM_PYBIND11=ON -DUSE_UCC=OFF -DUSE_VALGRIND=OFF -DPYBIND11_PYTHON_VERSION="${EPYTHON#python}" -DPYTHON_EXECUTABLE="${PYTHON}" -DUSE_ITT=OFF -DBLAS=Eigen # avoid the use of MKL, if found on the system -DUSE_SYSTEM_EIGEN_INSTALL=ON -DUSE_SYSTEM_PTHREADPOOL=ON -DUSE_SYSTEM_FXDIV=ON -DUSE_SYSTEM_FP16=ON -DUSE_SYSTEM_GLOO=ON -DUSE_SYSTEM_ONNX=ON -DUSE_SYSTEM_SLEEF=ON
If you look at both the ebuild I save a /var/lib/caffe2/CMakeCache.txt and reuse it in pytorch. The intention for this is to share the setting When I did I thought this could work, at least with 1.12. Maybe this is not working anymore for 2.00 ?
Can you check that it is actually using the settings from /var/lib/caffe2/CMakeCache.txt as settings that are enabled for caffe2 do not seem to be enabled in pytorch once I build it.
(In reply to Andrew Cameron from comment #4) > Can you check that it is actually using the settings from > /var/lib/caffe2/CMakeCache.txt as settings that are enabled for caffe2 do > not seem to be enabled in pytorch once I build it. Please provide the build log of both caffe2 and pytorch. Build pytorch after any change of USE for caffe2
Created attachment 861667 [details] Caffe2 Build log with opencl enabled
Created attachment 861668 [details] pytorch build log The pytorch build seems to ignore the settings from caffe2
Created attachment 861669 [details] /var/lib/caffe2/CmakeCache.txt
Created attachment 861670 [details] Sample pytorch run showing opencl was not included in the build
#python Python 3.11.3 (main, May 5 2023, 10:18:46) [GCC 12.2.1 20230304] on linux Type "help", "copyright", "credits" or "license" for more information. >>> import torch >>> t2=torch.randn(1,10).to('opencl') Error in cpuinfo: processor architecture is not supported in cpuinfo Traceback (most recent call last): File "<stdin>", line 1, in <module> RuntimeError: PyTorch is not linked with support for opencl devices >>>
from a quick grep in the source, it seems that opencl is not available. I don't see a device registration for the openCL. The only registration I see are for: CUDA, Metal, Vulkan, CPU, Meta, MPS, Lazy (some I don't know what are ) And from the web https://discuss.pytorch.org/t/runtimeerror-pytorch-is-not-linked-with-support-for-opencl-devices/164999 it seems the same. I added support in the ebuild, as I found the options in the CMakeList but, if that is confirmed, I should drop it
You can see the list defined here https://pytorch.org/cppdocs/api/program_listing_file_c10_core_DeviceType.h.html torch itself lists these as valid backends in the current version RuntimeError: Expected one of cpu, cuda, ipu, xpu, mkldnn, opengl, opencl, ideep, hip, ve, fpga, ort, xla, lazy, vulkan, mps, meta, hpu, mtia, privateuseone device type at start of device string: try python import torch t2=torch.randn(1,10).to('cpu') print(t2) Each of the devices listed above can be linked in to pytorch. Here is an ebuild that includes some of the options https://github.com/aclex/pytorch-ebuild/blob/master/sci-libs/pytorch/pytorch-1.13.1.ebuild On mine I want to enable the opengl opencl and vulkan backends but none of them work EG These should all work python Python 3.11.3 (main, May 5 2023, 10:18:46) [GCC 12.2.1 20230304] on linux Type "help", "copyright", "credits" or "license" for more information. >>> import torch >>> t2=torch.randn(1,10).to('cpu') Error in cpuinfo: processor architecture is not supported in cpuinfo >>> print(t2) tensor([[-0.4604, 2.6746, 0.7875, 1.1237, 0.8364, 1.4360, 0.8594, -1.1803, 0.6411, 0.0886]]) >>> t2=torch.randn(1,10).to('opengl') Traceback (most recent call last): File "<stdin>", line 1, in <module> RuntimeError: PyTorch is not linked with support for opengl devices >>> t2=torch.randn(1,10).to('opencl') Traceback (most recent call last): File "<stdin>", line 1, in <module> RuntimeError: PyTorch is not linked with support for opencl devices >>> t2=torch.randn(1,10).to('vulkan') Traceback (most recent call last): File "<stdin>", line 1, in <module> RuntimeError: PyTorch is not linked with support for vulkan devices >>>
I am emerging your updated versions. I will let you know tomorrow if it makes a difference.
(In reply to Andrew Cameron from comment #12) > You can see the list defined here > https://pytorch.org/cppdocs/api/program_listing_file_c10_core_DeviceType.h. > html > > torch itself lists these as valid backends in the current version > RuntimeError: Expected one of cpu, cuda, ipu, xpu, mkldnn, opengl, opencl, > ideep, hip, ve, fpga, ort, xla, lazy, vulkan, mps, meta, hpu, mtia, > privateuseone device type at start of device string: > > try > python > import torch > t2=torch.randn(1,10).to('cpu') > print(t2) > > Each of the devices listed above can be linked in to pytorch. > > > Here is an ebuild that includes some of the options > https://github.com/aclex/pytorch-ebuild/blob/master/sci-libs/pytorch/pytorch- > 1.13.1.ebuild > > On mine I want to enable the opengl opencl and vulkan backends but none of > them work > > EG > These should all work > python > Python 3.11.3 (main, May 5 2023, 10:18:46) [GCC 12.2.1 20230304] on linux > Type "help", "copyright", "credits" or "license" for more information. > >>> import torch > >>> t2=torch.randn(1,10).to('cpu') > Error in cpuinfo: processor architecture is not supported in cpuinfo > >>> print(t2) > tensor([[-0.4604, 2.6746, 0.7875, 1.1237, 0.8364, 1.4360, 0.8594, > -1.1803, > 0.6411, 0.0886]]) > >>> t2=torch.randn(1,10).to('opengl') > Traceback (most recent call last): > File "<stdin>", line 1, in <module> > RuntimeError: PyTorch is not linked with support for opengl devices > >>> t2=torch.randn(1,10).to('opencl') > Traceback (most recent call last): > File "<stdin>", line 1, in <module> > RuntimeError: PyTorch is not linked with support for opencl devices > >>> t2=torch.randn(1,10).to('vulkan') > Traceback (most recent call last): > File "<stdin>", line 1, in <module> > RuntimeError: PyTorch is not linked with support for vulkan devices > >>> I understand, but looking at the error, its source is libtorch_cpu.so (in caffe2), and a grep of C10_REGISTER_GUARD_IMPL, that seems the registration procedure of the backend, you don't have opencl. Whatever is, being the backtrace in caffe2, I don't think the problem is in the python side. Or opencl is not ready, or I was not able to support in caffe2
> On mine I want to enable the opengl opencl and vulkan backends but none of vulkan seems available only for android opencl seems only a starting point for development opengl is not selectable by cmake
I have created the following question on the pytorch discussions to see what they say https://discuss.pytorch.org/t/compile-pytorch-2-0-with-backend-support-for-opencl-opengl-and-vulkan/179996
(In reply to Andrew Cameron from comment #16) > I have created the following question on the pytorch discussions to see what > they say > https://discuss.pytorch.org/t/compile-pytorch-2-0-with-backend-support-for- > opencl-opengl-and-vulkan/179996 thanks
Please can you add the riscv keyword to the latest ebuilds for both caffe2 and pytorch and their dependencies as you saw from my ebuild logs they do compile on my riscv system. Once you have done that please close this BUG for now.
I am closing this as the behavior is the same when I build it from source from github.
