NXP backend: NXP backend: Add option to run test reference quantized in Python#17733
Conversation
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/17733
Note: Links to docs will display an error until the docs builds have been completed. ❌ 2 New FailuresAs of commit af07191 with merge base 8f527e1 ( NEW FAILURES - The following jobs have failed:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
|
@roman-janik-nxp please review this PR. There was some error and my previous PR is corrupted. The changes here should be identical to what you have approved in the other PR. |
|
The corresponding Executorch Integration PR is here: https://bitbucket.sw.nxp.com/projects/AITEC/repos/executorch-integration/pull-requests/127/diff#test%2Fpython%2Ftests%2Ftest_mobilenet_v2.py |
Please rebase on recent main. |
6b4e679 to
5b6299a
Compare
5b6299a to
af07191
Compare
|
The failing checks seem unrelated and the internal build is passing (https://bamboo3.sw.nxp.com/browse/MLTECE-EXIGH93-11). Merging |
Summary
NXP tests run models delegated to Neutron using the NSYS simulator. To determine correct output, a reference model is run on the CPU. Before, there were 2 choices for the reference, either non-delegated .pte file running in c++, or the original non-quantized float32 PyTorch model running in Python. This PR adds a 3rd option (to run in quantized edge dialect in Python), as well as easy extension to 2 more options in the future.
Test plan
Unit-tests provided.
cc @robert-kalmar @JakeStevens @digantdesai