[PT2][Quant] Move add/add relu pattern via module partitioner #102397
Conversation
This diff uses module partitioners to find add and add + relu patterns. Differential Revision: [D46095330](https://our.internmc.facebook.com/intern/diff/D46095330/) [ghstack-poisoned]
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/102397
Note: Links to docs will display an error until the docs builds have been completed. ❌ 1 New Failure, 3 Unrelated FailuresAs of commit dfb2f2c: NEW FAILURE - The following job has failed:
BROKEN TRUNK - The following jobs failed but were present on the merge base ae5606b:👉 Rebase onto the `viable/strict` branch to avoid these failures
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
…ner" This diff uses module partitioners to find add and add + relu patterns. Differential Revision: [D46095330](https://our.internmc.facebook.com/intern/diff/D46095330/) [ghstack-poisoned]
…ner" This diff uses module partitioners to find add and add + relu patterns. Differential Revision: [D46095330](https://our.internmc.facebook.com/intern/diff/D46095330/) [ghstack-poisoned]
| @@ -13,6 +14,7 @@ | |||
| {torch.nn.Conv2d, torch.nn.functional.conv2d}, | |||
| {torch.nn.ReLU, torch.nn.functional.relu, torch.nn.functional.relu_}, | |||
| {torch.nn.BatchNorm2d, torch.nn.functional.batch_norm}, | |||
| {torch.add, operator.add}, | |||
There was a problem hiding this comment.
Thanks for providing this helper function? How about the operator.iadd? Should it also been added here?
There was a problem hiding this comment.
yeah I think this is needed, thanks!
There was a problem hiding this comment.
Thanks for the reply @jerryzh168 @kimishpatel. BTW, I think current _EQUIVALENT_TYPES list has been hardcoded. So PyTorch Extension Libraries such as intel-extension-for-pytorch can't use Module Partition API with customized _EQUIVALENT_TYPES. Do you think is it reasonable to make it customizable? I have drafted a PR here #102516.
There was a problem hiding this comment.
yeah I think so, we haven't finalized our APIs though
There was a problem hiding this comment.
Thanks for providing this helper function? How about the operator.iadd? Should it also been added here?
Yes, adding iadd makes sense.
BTW, I think current _EQUIVALENT_TYPES list has been hardcoded. So PyTorch Extension Libraries such as intel-extension-for-pytorch can't use Module Partition API with customized _EQUIVALENT_TYPES.
For this can you say more. I briefly looked at the extension but couldnt quite figure what it is doing. Does it introduce custom nn modules for fusion etc? If so I would ask if fusion is something that can be done via graph rewrite? Any other examples might be helpful as well.
Regarding _EQUIVALENT_TYPES, yeah it can be something tht is extensible, but I am curious to know your usecase.
…ner" This diff uses module partitioners to find add and add + relu patterns. Differential Revision: [D46095330](https://our.internmc.facebook.com/intern/diff/D46095330/) [ghstack-poisoned]
|
@pytorchbot merge -f 'Landed internally' (Initiating merge automatically since Phabricator Diff has merged, using force because this PR might not pass merge_rules.json but landed internally) |
Merge startedYour change will be merged immediately since you used the force (-f) flag, bypassing any CI checks (ETA: 1-5 minutes). Learn more about merging in the wiki. Questions? Feedback? Please reach out to the PyTorch DevX Team |
Stack from ghstack (oldest at bottom):
This diff uses module partitioners to find add and add + relu patterns.
Differential Revision: D46095330