Fix broken example code in the static quantization tutorial

[amanyagami]

Summary

docs/source/eager_tutorials/static_quantization.rst's QuantizedLinear example had three compounding bugs that made it non-functional for any reader following it verbatim (tracked under #3637, updating docs that reference outdated workflows):

1. QuantizedLinear.from_observed() passed an extra target_dtype argument into cls(...) that QuantizedLinear.__init__ never accepted (it's unused) — raised:

   TypeError: QuantizedLinear.__init__() takes 7 positional arguments but 8 were given
   

2. Int8Tensor.from_hp's static-scale path asserts scale.ndim == hp_tensor.ndim, but AffineQuantizedMinMaxObserver.calculate_qparams() returns scale/zero_point with the block dims squeezed out (e.g. shape (64,) for a per-row scale on a (64, 64) weight, not (64, 1)) — raised:

   AssertionError (in Int8Tensor.from_hp)
   

3. Once 1–2 are fixed, forward() manually built a second Int8Tensor for the activation and passed both quantized tensors into F.linear. Int8Tensor's F.linear dispatch has no support for a pre-quantized activation tensor — only a plain high precision one, optionally paired with the weight tensor's act_quant_kwargs/act_quant_scale/act_quant_zero_point fields for static (fixed-scale) activation quantization. This raised:

   NotImplementedError: Int8Tensor dispatch: attempting to run unimplemented operator/function: func=aten.view.default, ...
   

Fix

Fold the calibrated activation scale/zero_point into the weight tensor via act_quant_kwargs/act_quant_scale/act_quant_zero_point on Int8Tensor.from_hp — the currently supported way to do static activation quantization — and have forward() pass the raw activation straight into F.linear; Int8Tensor's dispatch quantizes it internally using the fixed statistics. Also fixed the reshape needed for bug 2, dropped the dead target_dtype argument for bug 1, and updated the tutorial's illustrative REPL output (repr type name, tensor shape, and where the activation scale now lives) to match reality.

Test

Added test/quantization/test_static_quantization_doc_example.py, which mirrors this example (minus torch.compile/CUDA, so it runs on CPU) end to end.

• Verified it reproduces the original TypeError on the unpatched example code (and would hit the other two errors in sequence if that one were papered over — traced each independently while debugging).
• The full corrected pipeline runs cleanly and produces a NaN-free, correctly-shaped output with the fix.
• Verified the edited .rst still parses cleanly with docutils.

This leaves one remaining checklist item open in #3637 (updating the linked #3687, which was closed without merging) — that's a separate, larger scope change (a new Int8StaticActivationInt8WeightConfig) rather than a doc bug, so I've left it for a maintainer to scope.

🤖 Generated with Claude Code

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/ao/4794

• 📄 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.