🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/159639

• 📄 Preview Python docs built from this PR
• 📄 Preview C++ docs built from this PR
• ❓ Need help or want to give feedback on the CI? Visit the bot commands wiki or our office hours

Note: Links to docs will display an error until the docs builds have been completed.

✅ No Failures

As of commit 52c5b69 with merge base ff0d56d ():
💚 Looks good so far! There are no failures yet. 💚

This comment was automatically generated by Dr. CI and updates every 15 minutes.

@pytorchbot label "topic: not user facing"

@pytorchbot label "topic: not user facing"

Will let @soulitzer review this one

Thank you.

So sorry to you @soulitzer, could you help to take a look at this?

Sorry for the delay @FFFrog, I just returned from PTO
Could you share a code snippet of how this fallback will be used for out-of-tree backends?

Anyhow we probably don't want to expose the macro in the header, and instead just the function declaration

Thank you for your reply :D

Could you share a code snippet of how this fallback will be used for out-of-tree backends?

Okay, here's the snippet:

New accelerators maybe want to add forward functions for new operators, so it would be convenient if we could expose autograd_fallback in the header file. However, AUTOGRAD_FALLBACK is not necessary.

Thanks for the context!

btw, we do already have fallback APIs exposed in torch/csrc/autograd/autograd_not_implemented_fallback.h.

• The "autogradNotImplementedFallback" version attaches a node that errors upon backward.
• There is also existing basicAutogradNotImplementedFallback that is a more "gentle" to maintain bc: warn instead of error, etc.

What this PR exposes is a wrapper on top of the latter API, that has another of layer of indirection handling the case where the particular build of PyTorch does not have autograd.

It's probably fine to expose the version (with better naming, and after collocating the wrapper next to the other APIs), but I wonder if the existing APIs already work for your case @FFFrog?