Update Custom C++ and CUDA Operators to use PyTorch stable API/ABI #3701
Conversation
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/tutorials/3701
Note: Links to docs will display an error until the docs builds have been completed. ❌ 1 New FailureAs of commit 803be71 with merge base 10eefc3 ( NEW FAILURE - The following job has failed:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
mikaylagawarecki
force-pushed
the
stable_custom_op
branch
from
479f1c2 to
a8c5294
Compare
| Use ``torch.library.register_autograd`` to add training support for an operator. Prefer | ||
| this over directly using Python ``torch.autograd.Function`` or C++ ``torch::autograd::Function``; | ||
| you must use those in a very specific way to avoid silent incorrectness (see | ||
| this over directly using Python ``torch.autograd.Function`` (see |
There was a problem hiding this comment.
remove reference to torch::autograd::Function since we don't support that in stable API/ABI
| Defining the custom op and adding backend implementations | ||
| --------------------------------------------------------- | ||
| First, let's write a C++ function that computes ``mymuladd``: | ||
| First, let's write a C++ function that computes ``mymuladd`` using the LibTorch Stable ABI: |
There was a problem hiding this comment.
Hmm same question here as on the other PR. Do we have enough in 2.10 to always recommend the libtorch stable APIs? If not, then we're going to want a section with libtorch stable apis and a section with the regular libtorch APIs.
There was a problem hiding this comment.
I updated extension-cpp to have both a libtorch stable and an unstable extension and refer the user to them at the top. The snippets in this PR are from the stable variant but I beefed up the section on "Reverting to the Non-Stable LibTorch API" and included 1 code snippet there, referring the user to the non stable extension in extension-cpp otherwise.
Does this sound good to you?
(btw the link check is failing because the extension-cpp pr needs to land first)
2 participants
Description
Migrate
advanced_source/cpp_custom_ops.rstto libtorch stable ABI/API. Accompanying change toextension-cppChecklist