Arm backend: Document validated TinyML models for Cortex-M#19078
Conversation
Adds a Validated Models section to the Cortex-M backend overview listing the six models exported, INT8 quantized, and run on the Corstone-300 FVP by CI: mv2 and ds_cnn on trunk, and mv3, mobilenet_v1_025, resnet8, and deep_autoencoder nightly. For each model the table points at the source file and the per-model dialect/implementation test. A short note calls out that mobilenet_v1_025 is the MLPerf Tiny Visual Wake Words reference model — the canonical TinyML person-detection benchmark — since that naming is not obvious from the name. The page also documents the bundled (.bpte) testing flow that CI uses: aot_arm_compiler --bundleio embeds reference inputs and expected outputs in the program, and examples/arm/run.sh drives the full export → build → FVP chain with Test_result PASS/FAIL self-checking, so a reader can reproduce what trunk and nightly do. An admonition clarifies that CI validates INT8 numerical parity between the exported .bpte and the eager-mode quantized model, not task accuracy (VWW / KWS / ImageNet). This change was authored with Claude (claude-opus-4-7[1m]).
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/19078
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 563da6c with merge base cb4e5ae ( 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
This PR needs a
|
2 participants
Summary
Adds a Validated Models section to the Cortex-M backend overview listing the six models exported, INT8 quantized, and run on the Corstone-300 FVP by CI: mv2 and ds_cnn on trunk, and mv3, mobilenet_v1_025, resnet8, and deep_autoencoder nightly. For each model the table points at the source file and the per-model dialect/implementation test. A short note calls out that mobilenet_v1_025 is the MLPerf Tiny Visual Wake Words reference model — the canonical TinyML person-detection benchmark — since that naming is not obvious from the name.
The page also documents the bundled (.bpte) testing flow that CI uses: aot_arm_compiler --bundleio embeds reference inputs and expected outputs in the program, and examples/arm/run.sh drives the full export → build → FVP chain with Test_result PASS/FAIL self-checking, so a reader can reproduce what trunk and nightly do.
An admonition clarifies that CI validates INT8 numerical parity between the exported .bpte and the eager-mode quantized model, not task accuracy (VWW / KWS / ImageNet).
This change was authored with Claude (claude-opus-4-7[1m]).
cc @digantdesai @freddan80 @per @zingo @oscarandersson8218 @mansnils @Sebastian-Larsson @robell