Support activation-only PTQ (weight_qspec=None)#20433
Conversation
Summary: The Ethos/arm PT2E quantization annotator (`quantization_annotator.py`, fbcode + xplat) assumed every weighted op carries a weight qspec and unconditionally annotated the weight input. Activation-only PTQ -- int16 activations with fp32 (un-quantized) weights -- sets `weight_qspec=None`, which tripped the annotator. This skips weight annotation when the weight qspec is None, enabling activation-only PTQ on the arm/Ethos backend. Required by the `cascade_detector_ceres_u55_a16_only` recipe (D108781360), which isolates the activation-quant contribution of the grip-force int8 collapse. Differential Revision: D108781361
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/20433
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: ❌ 5 New Failures, 3 Unrelated FailuresAs of commit e116040 with merge base ce9a7f5 ( NEW FAILURES - The following jobs have failed:
FLAKY - The following jobs failed but were likely due to flakiness present on trunk:
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. |
|
@ryan-monroe has exported this pull request. If you are a Meta employee, you can view the originating Diff in D108781361. |
This PR needs a
|
Summary:
The Ethos/arm PT2E quantization annotator (
quantization_annotator.py, fbcode + xplat) assumed every weighted op carries a weight qspec and unconditionally annotated the weight input. Activation-only PTQ -- int16 activations with fp32 (un-quantized) weights -- setsweight_qspec=None, which tripped the annotator. This skips weight annotation when the weight qspec is None, enabling activation-only PTQ on the arm/Ethos backend.Required by the
cascade_detector_ceres_u55_a16_onlyrecipe (D108781360), which isolates the activation-quant contribution of the grip-force int8 collapse.Differential Revision: D108781361
cc @digantdesai @freddan80 @per @zingo @oscarandersson8218 @mansnils @Sebastian-Larsson @robell @rascani