If the input is contiguous, short-circuit infer_size_dv in reshape

[ezyang]

Stack from ghstack (oldest at bottom):

• -> If the input is contiguous, short-circuit infer_size_dv in reshape #95216

The main improvement is that this avoids guards from infer_size_dv,
although this also counts as a minor perf improvement too.

Signed-off-by: Edward Z. Yang ezyang@meta.com

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/95216

• 📄 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 0f5279e:
💚 Looks good so far! There are no failures yet. 💚

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

[albanD]

I guess that's ok as long as you're not doing anything funky with dispatch_sizes_strides_policy.

You do need to add the same if (!self.is_xla() && !self.is_lazy() && !self.is_ipu() && !at::isTensorSubclassLike(self)) { as below though.

[ezyang]

I think it is sound to omit the checks. Let us enumerate the cases:

• What about mkldnn? MKLDNN tensors are never contiguous, thus they will never hit this case.
• What about xla/lazy/ipu/tensor subclass short circuit? In fact, the whole point of the condition there is to ensure we hit a proper view. When we short circuit is contiguous, we always go to view. So we apply this behavior in that case
• What about the non short circuit case? It's always safe to go the slower route. In fact, it is required, because we didn't actually compute the stride

[ezyang]

What about mkldnn? MKLDNN tensors are never contiguous, thus they will never hit this case.

lol but CI HAS PROVED ME WRONGGGG

[ezyang]

@pytorchbot merge

[pytorchmergebot]

Merge started

Your change will be merged once all checks pass (ETA 0-4 Hours).

Learn more about merging in the wiki.

Questions? Feedback? Please reach out to the PyTorch DevX Team

Advanced Debugging

Check the merge workflow status
here

[albanD]

Sounds good!

[laithsakka]

one recent problem is that unbacked hit this case, and when that happens we loose the option to copy as a default last resort behaviour.