-
Notifications
You must be signed in to change notification settings - Fork 646
[Backend Tester] Clean up a few test issues #13258
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
base: gh/GregoryComer/113/head
Are you sure you want to change the base?
Conversation
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/13258
Note: Links to docs will display an error until the docs builds have been completed. ❌ 3 New FailuresAs of commit 5e92884 with merge base 3a02146 ( NEW FAILURES - The following jobs have failed:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
@@ -9,6 +9,14 @@ | |||
|
|||
import torch | |||
|
|||
# Set of unsupported ops that should cause tests to be skipped | |||
UNSUPPORTED_PORTABLE_OPS = { |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
If we are adding a Portable flow, how would these show up there? Some PTE_FAIL
?
@@ -142,12 +159,15 @@ def build_result( | |||
tester.run_method_and_compare_outputs( | |||
inputs=None if generate_random_test_inputs else inputs, | |||
statistics_callback=lambda stats: error_statistics.append(stats), | |||
atol=1e-1, |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
atol seems pretty high for a general default, no?
The fp16 and quantized tests can generate occasional high element-wise error but still have decent SNR (~60+).
Do you know how this is tested on PyTorch/PyTorch side? >60 SNR is good but outliers are not great esp for individual ops, if they are expected I would prefer if we set them per test basis which would allow us to reason about the math being done on that specific test warrenting high ATOL/RTOL.
There are a few broken tests that need cleaning up. Some are failing due to missing portable kernels. These tests are now skipped if any unsupported portable ops remain post-delegation. I also fixed a few other small issues and bumped the element-wise tolerance to reduce false positives. SNR should hopefully catch most blatant correctness issues. The fp16 and quantized tests can generate occasional high element-wise error but still have decent SNR (~60+).