-
Notifications
You must be signed in to change notification settings - Fork 741
Arm backend: Make TOSA backend NCHW-compatible #12994
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
Arm backend: Make TOSA backend NCHW-compatible #12994
Conversation
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/12994
Note: Links to docs will display an error until the docs builds have been completed. ❌ 5 New Failures, 1 Unrelated FailureAs of commit 8c66a7b with merge base d87766e ( 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. |
|
@digantdesai This is a quite big change, do you need to run it internally before merging? |
digantdesai
left a comment
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.
Glad to see so many xfails removed. Left some comments.
| @common.parametrize("module", modules) | ||
| def test_to_tosa_memory_format_tosa_MI_functional(module): | ||
| # Also run the actual pass pipeline to ensure functional correctness. | ||
| pipeline = TosaPipelineMI[input_t](module, module.get_inputs(), []) |
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.
No FVP tests?
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.
My intention here was to only test the logic of the pass with as little other things that could go wrong as possible ( quantization and fvps), we do also test the pass on FVP in other places since the change affects all modules we run tests on. Sounds ok?
|
This will unfortunately need a rebase after the TOSA 0.80.1 removal :( |
- Moves the needed input/output transposes into the delegated graph to run on Ethos-U rather than requiring the EthosUBackend to implement transposes on CPU. - Renames the annotate_channels_last_dim_order_pass to to_tosa_memory_format_pass since to be more descriptive and future proof. This changes additionally enables running muliple batches since the EtohsU transpose supports that natively, whereas the CPU implementation did not. Change-Id: I676e5915b15cbcc370a03d70bfa2ea2fe20b2210 Signed-off-by: Adrian Lundell <[email protected]>
61fa369 to
9c5ef40
Compare
|
@digantdesai has imported this pull request. If you are a Meta employee, you can view this in D79720805. |
@digantdesai Any update on this? |
Change-Id: I98f211d2e44ce68c6a1d03ee6941abf84108b3e5
|
Hey @Sebastian-Larsson Apologies for the delay. Internal tests got broken couple of days back so they are masking this signal once I merge #13221 hopefully I we can rebase this and pull internally again. 🙏 |
|
@digantdesai has imported this pull request. If you are a Meta employee, you can view this in D79720805. |
Change-Id: Ib9e678c9471b861df141becfd2eb9511964fb627 Signed-off-by: Adrian Lundell <[email protected]>
Getting this on the internal CI |
Have you tried the latest version? I have fixed one issue from the moving of the transpose to TOSA dialect change and that passed our internal CI at least. |
|
The sigmoid error can be seen on other PR:s as well so it is not related to this patch. |
|
@digantdesai has imported this pull request. If you are a Meta employee, you can view this in D79720805. |
|
I see that you are getting internal errors, anything I can help with? |
|
Actually its ok. Feel free to merge. |
|
Seems like it broke |
|
Seems like it yes, I thought it was flakyness since we had similar accuracy errors in the sigmoids before but apparently it always fails, sorry about that. It was previously an xfail though so it is just a matter of adding it back, I can do it first thing tomorrow morning if you do not do it. |
Fix issue introduced in #12994 Signed-off-by: Adrian Lundell <[email protected]>
Moves the needed input/output transposes into the delegated graph to run on Ethos-U rather than requiring the EthosUBackend to implement transposes on CPU.
Renames the annotate_channels_last_dim_order_pass to to_tosa_memory_format_pass since to be more descriptive and future proof.
This changes additionally enables running multiple batches since the EthosU transpose supports that natively, whereas the CPU implementation did not.
cc @digantdesai @freddan80 @per @zingo @oscarandersson8218