Arm Backend: Add Silero VAD Ethos-U example and recurrent-state lowering - #22172
Open
usamahz wants to merge 4 commits into
Open
Arm Backend: Add Silero VAD Ethos-U example and recurrent-state lowering#22172usamahz wants to merge 4 commits into
usamahz wants to merge 4 commits into
Conversation
Add an in-tree Silero VAD example for Corstone-320 and Ethos-U85. The flow loads the Silero weights into the eager model, calibrates PT2E quantization with real audio, lowers supported operators, and exports a PTE with reference probabilities. Add a bare-metal runtime that streams WAV frames and carries recurrent state between calls. It prints probabilities and speech segments over serial, and stops the FVP with end-of-transmission. Add deterministic OOTB coverage using separate calibration and validation clips. The implementation was developed with OpenAI Codex assistance. Signed-off-by: Usamah Zaheer <usamah.zaheer@arm.com> AI-Assisted-By: OpenAI Codex Change-Id: I4974461b9b86d508a019c52a0234f66f9822313a
Decompose LSTMCell for batched and unbatched inputs, including negative tuple indices, so the recurrent gate and state-update pattern can be lowered through the Arm INT pipeline. Preserve mutable-buffer names and storage identity around each Arm transform. Use the mutable-aware constant folder from the repository-pinned torchao revision so mutable state reads are not frozen. Store Silero hidden and cell state as model-owned int8 buffers with calibration-derived quantization parameters. This avoids a torchao pin update or backend storage rewrite while keeping the recurrent update delegated. Only small boundary conversions and the mutable-buffer write remain portable. Improve mutable-buffer mutation tagging and reset Silero calibration state at audio-stream boundaries. Add focused regression coverage for identical buffers, reordered mutations, quantized state, and LSTMCell lowering. The implementation was developed with OpenAI Codex assistance. Signed-off-by: Usamah Zaheer <usamah.zaheer@arm.com> AI-Assisted-By: OpenAI Codex Change-Id: I66b7f6719324fbeea6983dc83940ba5ba8e01123
usamahz
requested review from
JacobSzwejbka,
digantdesai,
kirklandsign and
larryliu0820
as code owners
August 26, 2026 10:45
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/22172
Note: Links to docs will display an error until the docs builds have been completed. ⏳ 1 Pending, 2 Unrelated FailuresAs of commit c490ecc with merge base 300987a ( BROKEN TRUNK - The following jobs 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. |
meta-cla
Bot
added
the
CLA Signed
usamahz
marked this pull request as draft
August 26, 2026 10:46
usamahz
marked this pull request as ready for review
August 26, 2026 12:19
usamahz
changed the title
zingo
approved these changes
Aug 27, 2026
zingo
left a comment
zingo
left a comment
Collaborator
There was a problem hiding this comment.
Ok to merge (if Arm tests passes)
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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
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
Silero VAD is a compact streaming model that demonstrates a realistic recurrent-state workload on Ethos-U85 rather than another stateless classification example.
This PR adds an in-tree Silero VAD Ethos-U example covering model preparation, PT2E quantization, Arm lowering, bare-metal execution on Corstone-320, and host-side output validation. It also adds the reusable Arm backend support needed to lower the LSTMCell gate/update pattern while preserving model-owned mutable hidden and cell state between audio frames.
The runtime accepts one 576-sample audio input per frame, keeps recurrent state inside the model, prints speech probabilities and detected segments over serial, and validates the FVP results against export-time reference probabilities.
Tracking: MLETORCH-2112
This PR was developed with OpenAI Codex assistance.
Test plan
python -m pytest -q backends/arm/test/passes/test_mutable_buffers.py backends/arm/test/misc/test_silero_vad_example.py— 14 passed.backends/arm/test/test_arm_ootb.sh run_silero_vad_e2e_ethos_u— exports the PTE, builds the runtime, runs Corstone-320 / Ethos-U85-256 FVP, and compares 79 output frames.0.156800, mean absolute error0.018459, and zero threshold-decision mismatches at0.55.cc @digantdesai @freddan80 @per @zingo @oscarandersson8218 @mansnils @Sebastian-Larsson @robell @rascani