Qualcomm: let the 8a8w QAT activation spec share an observer with per…

[psiddh]

…-channel

QNN carries two activation specs: the default config for most ops and the per-channel config for conv and linear. PT2E gives an edge one observer only when the producer's output qspec and the consumer's input qspec agree on every attribute _union_input_edge_with compares -- dtype, is_dynamic, quant_min, quant_max, qscheme, ch_axis, scale, zero_point. observer_or_fake_quant_ctr is deliberately excluded, so this is about attribute values, not spec identity or SharedQuantizationSpec.

get_8a8w_qnn_qat_config built its activation spec unconditionally in the symmetric shape: ch_axis=0, no quant_min/quant_max, with only qscheme switching on act_symmetric. get_qat_per_channel_quant_config branches properly, so on the default asymmetric path the two disagree on quant_min, quant_max and ch_axis. Every per-channel conv output meeting a default-config consumer then gets two observers instead of one, which convert_pt2e materializes as a dequantize immediately followed by a requantize. On a conv-relu-conv model with no BatchNorm, QAT produced three such clusters where PTQ produced none.

Split on act_symmetric the way get_8a8w_qnn_ptq_config already does. Passing explicit uint8 bounds for per_tensor_affine is a numerical no-op -- the observer already defaults to 0/255, and scale, zero_point and fake-quant output are bit-identical either way -- it only makes the attributes visible to the comparison.

Two tests. The first compares the default and per-channel activation specs across every QUANT_CONFIG_DICT entry using PT2E's own _has_same_attr; it carries a _KNOWN_UNSHAREABLE list for the eight 16-bit act_symmetric combinations, where the default configs set quant_min/quant_max unconditionally while the per-channel config omits them. Repairing those changes the observed zero_point on a path that works today, so it is left for a follow-up; the list asserts they still mismatch, so it fails if anything shifts. The second test is the end-to-end guard, counting quantize nodes fed directly by a dequantize.

Both fail without the fix: 3 != 0, and mismatched on ['quant_min', 'quant_max', 'ch_axis'].

test_passes.py already runs in OSS CI by name, so no workflow change is needed.

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/22238

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

This comment was automatically generated by Dr. CI and updates every 15 minutes.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[github-actions]

This PR needs a release notes: label

If your change should be included in the release notes (i.e. would users of this library care about this change?), please use a label starting with release notes:. This helps us keep track and include your important work in the next release notes.

To add a label, you can comment to pytorchbot, for example
@pytorchbot label "release notes: none"

For more information, see
https://github.com/pytorch/pytorch/wiki/PyTorch-AutoLabel-Bot#why-categorize-for-release-notes-and-how-does-it-work.