NXP backend: Add prelu support using new neutron flow#20301
NXP backend: Add prelu support using new neutron flow#20301roman-janik-nxp wants to merge 4 commits into
prelu support using new neutron flow#20301Conversation
roman-janik-nxp
added
module: nxp
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/20301
Note: Links to docs will display an error until the docs builds have been completed. ❌ 3 New Failures, 1 PendingAs of commit 278b617 with merge base 06143cb ( NEW FAILURES - The following jobs have failed:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
meta-cla
Bot
added
the
CLA Signed
|
roman-janik-nxp
had a problem deploying
to
cadence
GitHub Actions
Failure
roman-janik-nxp
had a problem deploying
to
cadence
GitHub Actions
Failure
| ), | ||
| ], | ||
| ) | ||
| def test__w_conv_unsupported(self, input_spec): |
There was a problem hiding this comment.
It seems odd to me to handle the broadcasting stuff in a prelu PR (since the input alpha should always have the shape [1] or [<num_channels>], neither of which should be relevant), but so be it.
I assume you are working on EIEX-908. But unless I'm missing something, you didn't address the related EIEX-932 (which is apparently also in progress in JIRA). The second task specifically focuses on removing the restrictions related channels last broadcasting, which sometimes requires the insertion of Transpose ops. Now that all transpositions are supported, these tests should be enabled as well.
Are you planning on including EIEX-932 here as well? If not, why is EIEX-908 included in this prelu PR?
| pytest.param((1,), id="1D."), | ||
| pytest.param( | ||
| (36, 487), | ||
| id="2D incorrect.", |
There was a problem hiding this comment.
Nit: The id incorrect doesn't provide much context to someone hypothetically seeing this test fail in the console. Consider id="2D incorrect results.".
| # Check `prelu` was not decomposed into simpler edge operators | ||
| assert not graph_contains_any_of_ops( | ||
| exported_program.graph, | ||
| class TestPreluConverter: |
There was a problem hiding this comment.
I don't see any tests using the num_parameters torch.nn.PReLU parameter. Is it intentional?
2 participants
Summary
Only last 2 commits apply. Requires NXP backend: Improve test result gathering#20024 PR to be merged. Add tests verifying correct support for prelu by the Neutron backend using the new Neutron MLIR flow
Test plan
Unit tests provided.
cc @robert-kalmar