Add custom class pt2 tutorial #3431
Conversation
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/tutorials/3431
Note: Links to docs will display an error until the docs builds have been completed. ✅ No FailuresAs of commit ba48f42 with merge base 8476a99 ( This comment was automatically generated by Dr. CI and updates every 15 minutes. |
angelayi
force-pushed
the
angelayi/torchbind
branch
from
06880b9 to
be8b159
Compare
|
|
||
| import torch | ||
|
|
||
| torch.ops.load_library("//caffe2/test:test_torchbind_cpp_impl") |
There was a problem hiding this comment.
probably, I hope the OSS buck/bazel target for torch isn't caffe2 lol
There was a problem hiding this comment.
So it would be nice if we had some sample code someone could download that had the C++ extension set up with it. Then it would be clearer exactly what needs to be loaded (via load_library, or an import statement). But I'm happy to ship some version of this as-is without that because it already helps alot
| However, it may be difficult for users to write a fake class: the original class | ||
| uses some third-party library that determines the output shape of the methods, | ||
| or is complicated and written by others. Besides, users may not care about the | ||
| limitations listed above. In this case, please reach out to us! |
There was a problem hiding this comment.
Can also mention users can define x.tracing_mode(), which returns "real" so as to skip fakification.
|
Thanks for porting this tutorial! Looks good to me. |
| This tutorial is a follow-on to the | ||
| :doc:`custom C++ classes <torch_script_custom_classes>` tutorial, and | ||
| introduces additional steps that are needed to support custom C++ classes in | ||
| PyTorch 2. |
There was a problem hiding this comment.
torch.compile/torch.export. I don't think users understand the difference between 2 being a version and pt2 being the codename for torch.compile/torch.export.
ditto for the title.
| :doc:`custom C++ classes <torch_script_custom_classes>` tutorial, and | ||
| introduces additional steps that are needed to support custom C++ classes in | ||
| PyTorch 2. | ||
|
|
There was a problem hiding this comment.
Do we have a sense of the stability of this? Does it always work now? If not, some warning about the rough edges around this (this is a "prototype" feature, right? or "unstable" in the current classification) would help
| Why do we need to make a Fake Class? | ||
| ------------------------------------ |
There was a problem hiding this comment.
We should try to list out all the caveats around tensor aliasing. Namely, we're gonna assume that tensors inside and outside a torchbind object are different. If you do things like mutate one of them, there will be undefined behavior.
Porting over part of https://docs.google.com/document/d/1xUfRsqyQp3CUqpHaPjG3APxx9gHEKjQ3ISYi5jtKpuk/edit?tab=t.0
preview: https://docs-preview.pytorch.org/pytorch/tutorials/3431/advanced/custom_class_pt2.html
cc @williamwen42 @msaroufim @anijain2305 @avikchaudhuri @gmagogsfm @zhxchen17 @tugsbayasgalan @suo @ydwu4