[dispatcher] avoid autograd fixup step on non-backend keys #46135
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
[ghstack-poisoned]
bhosmer
added a commit
that referenced
this pull request
Oct 10, 2020
ghstack-source-id: c0139a203e9684dea35c9dd22487d06f641a325a Pull Request resolved: #46135
💊 CI failures summary and remediationsAs of commit cb02402 (more details on the Dr. CI page):
This comment was automatically generated by Dr. CI (expand for details).Follow this link to opt-out of these comments for your Pull Requests.Please report bugs/suggestions on the GitHub issue tracker or post in the (internal) Dr. CI Users group. This comment has been revised 16 times. |
Previously we were unconditionally running the autograd fixup post-processing step in `updateDispatchTable_()`, even when the key being updated wasn't a backend key. This wasn't failing because `getAutogradKeyFromBackend()` would default to `AutogradOther` for non-backend keys. (So the extra call wasn't doing any damage, just resynching AutogradOther a bit more than necessary :P). Reviewer notes: here's everything in the diff: 1. Main change is just the `isBackendDispatchKey()` guard on the fixup step 2. ...along with that new helper function. 3. With this change I was able to tighten the definition of `getAutogradKeyFromBackend()` to avoid leaking `AutogradOther`. 4. This actually led me to another such leak: `UndefinedTensorImpl` was adding `AutogradOther` to its `key_set_` as a result of calling `getAutogradKeyFromBackend(DispatchKey::Undefined)`. Avoided this with a guard in `TensorImpl`'s constructor. The new overhead seems tolerable; I think the only hygienic alternative would be to refactor the key_set initialization so TensorImpl and UndefinedTensorImpl have different paths. Didn't seem worth it? 5. The rest is some minor code movement to get things in the right header files for the new predicate functions, and autofixing some whitespace errors from the comments PR. Differential Revision: [D24235974](https://our.internmc.facebook.com/intern/diff/D24235974) [ghstack-poisoned]
Previously we were unconditionally running the autograd fixup post-processing step in `updateDispatchTable_()`, even when the key being updated wasn't a backend key. This wasn't failing because `getAutogradKeyFromBackend()` would default to `AutogradOther` for non-backend keys. (So the extra call wasn't doing any damage, just resynching AutogradOther a bit more than necessary :P). Reviewer notes: here's everything in the diff: 1. Main change is just the `isBackendDispatchKey()` guard on the fixup step 2. ...along with that new helper function. 3. With this change I was able to tighten the definition of `getAutogradKeyFromBackend()` to avoid leaking `AutogradOther`. 4. This actually led me to another such leak: `UndefinedTensorImpl` was adding `AutogradOther` to its `key_set_` as a result of calling `getAutogradKeyFromBackend(DispatchKey::Undefined)`. Avoided this with a guard in `TensorImpl`'s constructor. The new overhead seems tolerable; I think the only hygienic alternative would be to refactor the key_set initialization so TensorImpl and UndefinedTensorImpl have different paths. Didn't seem worth it? 5. The rest is some minor code movement to get things in the right header files for the new predicate functions, and autofixing some whitespace errors from the comments PR. Differential Revision: [D24235974](https://our.internmc.facebook.com/intern/diff/D24235974) [ghstack-poisoned]
bhosmer
added a commit
that referenced
this pull request
Oct 13, 2020
ghstack-source-id: e5ba3f793a9f885b0df82c7d8259eb2dc089a269 Pull Request resolved: #46135
Codecov Report
@@ Coverage Diff @@
## gh/bhosmer/53/base #46135 +/- ##
=====================================================
Coverage ? 68.26%
=====================================================
Files ? 410
Lines ? 53608
Branches ? 0
=====================================================
Hits ? 36593
Misses ? 17015
Partials ? 0 Continue to review full report at Codecov.
|
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Stack from ghstack:
Previously we were unconditionally running the autograd fixup post-processing step in
updateDispatchTable_(), even when the key being updated wasn't a backend key. This wasn't failing becausegetAutogradKeyFromBackend()would default toAutogradOtherfor non-backend keys. (So the extra call wasn't doing any damage, just resynching AutogradOther a bit more than necessary :P).Reviewer notes: here's everything in the diff:
isBackendDispatchKey()guard on the fixup stepDifferential Revision: D24235974