(Experimental) Make nvFuser executor treat ltorch.copy_ differently from prims.copy_
#1177
+29
−4
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
…ciated fd.ops.set only for ltorch.copy_
|
Would it not be cleaner to have a separate prim for the inplace-resolving copy? |
|
Yes, it must be possible to prepare a separate primitive or give |
|
Yeah, I totally agree that we should not mess with prims too much, but to my mind, having a specialized, separate prim for the inplace is something very straightforward and likely better than playing tricks with the levels of decomposition. |
|
That sounds nice. I will close this PR for now and try the way you suggest. Thank you! |
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Outdated suggestions cannot be applied.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.
Fixes #1173. I implemented the idea mentioned in #1173 (comment). This PR makes the nvFuser executor absorb
ltorch.copy_separately fromprims.copy_(which appears within in-place ops).This PR is experimental. It works (if the tests pass), but it is not a good solution in that some transform may walk in and replace
ltorch.copy_with something equivalent. A complete solution would be to track allfd.add_outputgenerated so far and insertfd.ops.setwhere needed.