NXP backend: Add preprocessing pass to split multilayer GRU.
#13757
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/13757
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: ❌ 1 New FailureAs of commit b0f6ecb with merge base f8a422c ( NEW FAILURE - The following job has failed:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
meta-cla
bot
added
the
CLA Signed
|
@pytorchbot label "module: nxp" "release notes: nxp" |
pytorch-bot
bot
added
module: nxp
MartinPavella
force-pushed
the
upstream/main-nxp/EIEX-501-upstream-pass-to-decompose-multilayer-gru
branch
from
ffc75b4 to
325ee60
Compare
robert-kalmar
approved these changes
Aug 29, 2025
digantdesai
reviewed
Aug 29, 2025
| passes: list[PassType] = passes or [ | ||
| FuseBatchNormWithConvPass(), | ||
| FuseBatchNormWithLinearPass(), | ||
| SplitGRUBasedOnNumLayers(), |
There was a problem hiding this comment.
we do it here in aten is because we want to support quantization?
There was a problem hiding this comment.
We do it in the aten dialect because it allows the pass to focus only of the transformation, and not worry about the quantization. This way, the implementation is simpler, and if the quantization requirements of GRU (or the other ops) ever change, we only need to update the quantizer (otherwise we would also have to update this pass).
There was a problem hiding this comment.
We do it here for 2 reasons:
- Neutron NPU supports GRU as primitive operation Neutron IR ==> for Neutron NPU we do not need and want to decompose it to primitive ops by the ExecuTorch.
- Neutron NPU supports GRU, but only single layer. Therefore we transform the multilayer GRU to a sequence of single layer GRU, perform the quantization and obtain a graph what can be represented in Neutron IR, including the proper quantization parameters for inputs and outputs of individual GRU layers.
Note: we will preserve the GRU op in to_edge tranformation - coming change
MartinPavella
force-pushed
the
upstream/main-nxp/EIEX-501-upstream-pass-to-decompose-multilayer-gru
branch
2 times, most recently
from
5d6e466 to
fbb061c
Compare
…ltiple operators. The `aten.gru.input` has a parameter `num_layers`. For values > 1, it represents multiple `aten.gru.input` operators chained together. The introduced pass can split the original GRU into a chain of simpler GRU nodes.
MartinPavella
force-pushed
the
upstream/main-nxp/EIEX-501-upstream-pass-to-decompose-multilayer-gru
branch
from
fbb061c to
b0f6ecb
Compare
robert-kalmar
deleted the
upstream/main-nxp/EIEX-501-upstream-pass-to-decompose-multilayer-gru
branch
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Summary
This PR introduces a pre-processing pass on the aten dialect level, which splits
grunodes withnum_layers > 1into an equivalent sequence of single layergrunodes.Test plan
Unit tests provided in
backends/nxp/tests/test_gru_splitting.py.cc @robert-kalmar @roman-janik-nxp @StrycekSimon @jirioc