Add generic annotator for data layout ops #5814
Conversation
Data layout ops like unsqueeze are not annotated by the quantizer per default which leads to issues down the line. Therefore we add a generic annotator to explicitly annotate those ops. Signed-off-by: Benjamin Klimczak <[email protected]> Change-Id: Id3919abcb3df0b81159f3cccaab9785f8706b9cd
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/5814
Note: Links to docs will display an error until the docs builds have been completed. ❌ 3 New Failures, 1 Unrelated FailureAs of commit 265cdb7 with merge base 393553c ( NEW FAILURES - The following jobs have failed:
BROKEN TRUNK - The following job failed but were present on the merge base:👉 Rebase onto the `viable/strict` branch to avoid these failures
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
facebook-github-bot
added
the
CLA Signed
benkli01
added
partner: arm
Can you elaborate? what issues? Is it just lowering issues? |
| @@ -271,6 +271,7 @@ class ArmQuantizer(Quantizer): | |||
| "mm", | |||
| "cat", | |||
| "one_to_one", | |||
| "generic", | |||
There was a problem hiding this comment.
nit: too vague? May it's just me.
There was a problem hiding this comment.
No, it is a bit vague (generic even 😅), but naming things is hard...
|
|
||
|
|
||
| class TestGenericAnnotator(unittest.TestCase): | ||
| def check_annotation(self, model): |
|
@digantdesai has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator. |
Yes, it's a lowering issue. If the data layout op is in the beginning of the graph the forward propagation of annotations does not work (unless we annotate the inputs explicitly). We tried implementing a backward propagation as well, but sharing specs in the backward direction seems to be not supported, so we went for explicit annotation. Or are we missing something? |
|
@digantdesai merged this pull request in 98a58e0. |
Data layout ops like unsqueeze are not annotated by the quantizer per default which leads to issues down the line. Therefore we add a generic annotator to explicitly annotate those ops.