Add meta impl for grid_sampler_2d_backward #88745
Conversation
[ghstack-poisoned]
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/88745
Note: Links to docs will display an error until the docs builds have been completed. ✅ No FailuresAs of commit d1f664b: This comment was automatically generated by Dr. CI and updates every 15 minutes. |
github-actions
bot
requested review from
Chillee,
bdhirsh,
ezyang,
miladm and
wconstab
[ghstack-poisoned]
|
Test? |
[ghstack-poisoned]
TODO: add an OpInfo [ghstack-poisoned]
TODO: add an OpInfo [ghstack-poisoned]
TODO: add an OpInfo [ghstack-poisoned]
TODO: add an OpInfo [ghstack-poisoned]
TODO: add an OpInfo [ghstack-poisoned]
TODO: add an OpInfo [ghstack-poisoned]
|
@pytorchbot merge -g |
Merge startedYour change will be merged once all checks on your PR pass since you used the green (-g) flag (ETA: 0-4 Hours). Learn more about merging in the wiki. Questions? Feedback? Please reach out to the PyTorch DevX Team |
Merge failedReason: The following mandatory check(s) failed (Rule Dig deeper by viewing the failures on hud Details for Dev Infra teamRaised by workflow job |
TODO: add an OpInfo [ghstack-poisoned]
TODO: add an OpInfo [ghstack-poisoned]
torch/_meta_registrations.py
Outdated
|
|
||
| @register_meta([aten.round.default, aten.round.decimals]) | ||
| def meta_round(self, **kwargs): | ||
| return _elementwise_meta(self, type_promotion=ELEMENTWISE_PRIM_TYPE_PROMOTION_KIND.DEFAULT) |
There was a problem hiding this comment.
That is the intention, yes, but that meta function is intended for prims.round:
pytorch/torch/_prims/__init__.py
Line 810 in 5faa279
I'm assuming this is adding a meta function for refs.round, torch.round, or aten.round? In that case we need to look at what refs.round does that prims.round does not:
pytorch/torch/_refs/__init__.py
Line 778 in 5faa279
But luckily in this case it just performs "default" type promotion, which is a computation detail and doesn't effect the output's metadata.
Anyway, in this case I think you can directly use the meta function for elementwise prims. Note that if you wanted to write a meta function for an operation like refs.add then the meta function for prims.add would not be correct. A helper for writing meta functions for elementwise torch/aten/reference operations could probably be written without too much trouble, though.
There was a problem hiding this comment.
Note that if you wanted to write a meta function for an operation like refs.add then the meta function for prims.add would not be correct.
how come?
There was a problem hiding this comment.
The prims don't support features like type promotion, and torch.add/refs.add in particular has unique behavior for handling its alpha parameter
TODO: add an OpInfo [ghstack-poisoned]
TODO: add an OpInfo [ghstack-poisoned]
|
@pytorchbot merge -g |
Merge startedYour change will be merged once all checks on your PR pass since you used the green (-g) flag (ETA: 0-4 Hours). Learn more about merging in the wiki. Questions? Feedback? Please reach out to the PyTorch DevX Team |
TODO: add an OpInfo cc mlazos soumith voznesenskym yanboliang penguinwu anijain2305 EikanWang jgong5 Guobing-Chen chunyuan-w XiaobingSuper zhuhaozhe blzheng Xia-Weiwen wenzhe-nrv jiayisunx peterbell10 desertfire [ghstack-poisoned]
Merge failedReason: New commits were pushed while merging. Please rerun the merge command. Details for Dev Infra teamRaised by workflow job |
|
@pytorchbot merge -g |
Merge startedYour change will be merged once all checks on your PR pass since you used the green (-g) flag (ETA: 0-4 Hours). Learn more about merging in the wiki. Questions? Feedback? Please reach out to the PyTorch DevX Team |
TODO: add an OpInfo Pull Request resolved: pytorch#88745 Approved by: https://github.com/ezyang
Stack from ghstack (oldest at bottom):
TODO: add an OpInfo
cc @mlazos @soumith @voznesenskym @yanboliang @penguinwu @anijain2305 @EikanWang @jgong5 @Guobing-Chen @chunyuan-w @XiaobingSuper @zhuhaozhe @blzheng @Xia-Weiwen @wenzhe-nrv @jiayisunx @peterbell10 @desertfire