Conversation
|
cc @orionr |
|
Thanks for kicking off this conversation, @syed-ahmed! As you alluded to, we're looking to merge / refactor Caffe2 and ATen tensors, registry, and operators under a new c10 directory. Effectively aten/ will become c10/ (living at pytorch/c10/ or pytorch/caffe2/c10/ - still TBD) and support both PyTorch and Caffe2 frontends. @syed-ahmed, what is the use case where you need a separate ATen (and hopefully soon c10) build? We're definitely open to this (especially when c10 is mature), but I want to make sure what you're thinking is compatible with this new architectural split. In the meantime, we've stopped syncing to https://github.com/zdevito/ATen/ so it isn't impacted by the large changes coming. cc @dzhulgakov @ezyang and @smessmer as well. |
|
Thank you for the reply! I was just going by the readme and build ATen using cmake, since it mentioned of it being standalone. I like that ATen is standalone and I can build and test features without having to build from the root directory, but it's more of a personal choice. I don't currently have a use case for ATen/c10 to be standalone, but I have been poking around Tensor Comprehensions and it seems to have ATen as a dependency. |
facebook-github-bot
left a comment
There was a problem hiding this comment.
@orionr has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator.
Summary: This PR changes the ATen `CMakeLists.txt` slightly, to enable standalone build of ATen inside PyTorch. Currently, the tests in ATen gets linked to `libcaffe.so libcaffe2.so`. As a result, ATen can't be built standalone without building from the root pytorch directory. I know that there is a big merge happening between caffe2 and pytorch and hence, the purpose of this PR is to really start a conversation on what would be the proper way of migrating the CMakeLists to enable clean builds. We should also follow up on this PR: pytorch/pytorch#7275. For your reference, that PR has the explanation for why `-Wl --no-as-need` is needed. Moreover, without `set(ATen_CUDA_SRCS ${all_cuda_cpp})`, the standalone build will throw unresolved references. Pull Request resolved: pytorch/pytorch#9377 Reviewed By: smessmer Differential Revision: D8825921 Pulled By: orionr fbshipit-source-id: c521159b4885639fc7990a9819202051455d07db
Summary: This PR changes the ATen `CMakeLists.txt` slightly, to enable standalone build of ATen inside PyTorch. Currently, the tests in ATen gets linked to `libcaffe.so libcaffe2.so`. As a result, ATen can't be built standalone without building from the root pytorch directory. I know that there is a big merge happening between caffe2 and pytorch and hence, the purpose of this PR is to really start a conversation on what would be the proper way of migrating the CMakeLists to enable clean builds. We should also follow up on this PR: pytorch/pytorch#7275. For your reference, that PR has the explanation for why `-Wl --no-as-need` is needed. Moreover, without `set(ATen_CUDA_SRCS ${all_cuda_cpp})`, the standalone build will throw unresolved references. Pull Request resolved: pytorch/pytorch#9377 Reviewed By: smessmer Differential Revision: D8825921 Pulled By: orionr fbshipit-source-id: c521159b4885639fc7990a9819202051455d07db
Summary: This PR changes the ATen `CMakeLists.txt` slightly, to enable standalone build of ATen inside PyTorch. Currently, the tests in ATen gets linked to `libcaffe.so libcaffe2.so`. As a result, ATen can't be built standalone without building from the root pytorch directory. I know that there is a big merge happening between caffe2 and pytorch and hence, the purpose of this PR is to really start a conversation on what would be the proper way of migrating the CMakeLists to enable clean builds. We should also follow up on this PR: pytorch#7275. For your reference, that PR has the explanation for why `-Wl --no-as-need` is needed. Moreover, without `set(ATen_CUDA_SRCS ${all_cuda_cpp})`, the standalone build will throw unresolved references. Pull Request resolved: pytorch#9377 Reviewed By: smessmer Differential Revision: D8825921 Pulled By: orionr fbshipit-source-id: c521159b4885639fc7990a9819202051455d07db
This PR changes the ATen
CMakeLists.txtslightly, to enable standalone build of ATen inside PyTorch. Currently, the tests in ATen gets linked tolibcaffe.so libcaffe2.so. As a result, ATen can't be built standalone without building from the root pytorch directory. I know that there is a big merge happening between caffe2 and pytorch and hence, the purpose of this PR is to really start a conversation on what would be the proper way of migrating the CMakeLists to enable clean builds. We should also follow up on this PR: #7275. For your reference, that PR has the explanation for why-Wl --no-as-needis needed. Moreover, withoutset(ATen_CUDA_SRCS ${all_cuda_cpp}), the standalone build will throw unresolved references.