NXP backend: Make the flow robust against input/output swapping. #12890
Conversation
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/12890
Note: Links to docs will display an error until the docs builds have been completed. ❗ 1 Active SEVsThere are 1 currently active SEVs. If your PR is affected, please view them below: ❌ 2 New FailuresAs of commit 3c5e971 with merge base 4197fc1 ( NEW FAILURES - The following jobs have failed:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
facebook-github-bot
added
the
CLA Signed
StrycekSimon
force-pushed
the
upstream/main-nxp/EIEX-400-upstream-robustness-against-input-output-swapping
branch
from
299fc73 to
03abf61
Compare
|
@pytorchbot label "release notes: nxp" |
pytorch-bot
bot
added
the
release notes: nxp
|
NXP build failing |
|
Thanks, I noticed it earlier but I was already out of office. I will fix it. |
StrycekSimon
force-pushed
the
upstream/main-nxp/EIEX-400-upstream-robustness-against-input-output-swapping
branch
from
03abf61 to
d0eb3c4
Compare
robert-kalmar
added
the
module: nxp
| assert ( | ||
| len(outputs) < 256 | ||
| len(neutron_artifacts.input_indices) < 256 | ||
| ), "Models with more than 255 inputs are not supported." |
There was a problem hiding this comment.
NIT: This is actually a valid case, not a "programming error". We just use 8bit field to encode input indices. Consider using ValueError instead of assert.
| len(outputs) < 256 | ||
| len(neutron_artifacts.input_indices) < 256 | ||
| ), "Models with more than 255 inputs are not supported." | ||
| assert ( |
There was a problem hiding this comment.
NIT: This is actually a valid case, not a "programming error". We just use 8bit field to encode input indices. Consider using ValueError instead of assert.
robert-kalmar
force-pushed
the
upstream/main-nxp/EIEX-400-upstream-robustness-against-input-output-swapping
branch
from
d0eb3c4 to
3c5e971
Compare
| #define OUTPUT_TENSOR_FORMAT_LEN_POS 1 | ||
| #define INPUT_TENSOR_FORMAT_ARRAY_ADDR(base) (base + 2 * ITEM_SIZE) | ||
| #define INPUT_ARGS_LEN_POS 2 | ||
| #define INPUT_TENSOR_FORMAT_ARRAY_ADDR(base) (base + 3 * ITEM_SIZE) |
There was a problem hiding this comment.
This is ok for now but in the future consider FC/BC issues when updating serialization formats.
There was a problem hiding this comment.
We hope that the format is now general enough to cover the future scenarios.
There was a problem hiding this comment.
Says like a true developer, until a second before breaking BC :-p
robert-kalmar
deleted the
upstream/main-nxp/EIEX-400-upstream-robustness-against-input-output-swapping
branch
…orch#12890) ### Summary The NXP backend is now robust against the swapping the order of inputs/outputs in the model converter. Release notes: NXP ### Test plan test_neutron_backend.py tests this feature
Summary
The NXP backend is now robust against the swapping the order of inputs/outputs in the model converter.
Release notes: NXP
Test plan
test_neutron_backend.py tests this feature
cc @digantdesai @JakeStevens @robert-kalmar