Remove LinearActivationQuantizedTensor and all related code

[jerryzh168]

Stack from ghstack (oldest at bottom):

• -> Remove LinearActivationQuantizedTensor and all related code #4258

Summary:
Delete LinearActivationQuantizedTensor (and its alias to_linear_activation_quantized)
as part of the ongoing AQT cleanup. This tensor subclass wrapped a weight tensor to
intercept F.linear and quantize activations on the fly. It is no longer needed as
activation quantization is handled through other mechanisms.

Removed:

• torchao/quantization/linear_activation_quantized_tensor.py (class definition)
• Imports and __all__ exports in torchao/quantization/__init__.py
• isinstance guard in _is_linear() in quant_api.py
• isinstance branch in _quantization_type() in utils.py

Test Plan:

• Verified no remaining imports or usage of LinearActivationQuantizedTensor or
  to_linear_activation_quantized in the codebase (only string literals in error
  messages of unrelated files remain)
• CI should pass since no tests depended on this class

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/ao/4258

• 📄 Preview Python docs built from this PR

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

❗ 1 Active SEVs

There are 1 currently active SEVs. If your PR is affected, please view them below:

• Rolling out OSDC (ARC) runners on pull workflow for PyTorch trunk commits

⏳ No Failures, 7 Pending

As of commit c793f08 with merge base fee43b5 ():
💚 Looks good so far! There are no failures yet. 💚

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

[jerryzh168]

@claude review

[claude]

Claude finished @jerryzh168's task in 1m 23s —— View job

---

Review: Remove LinearActivationQuantizedTensor and all related code

• Read the PR diff and changed files
• Check for remaining references to removed code
• Review correctness and completeness
• Post review feedback

---

Verdict: This PR looks good. The deletion is clean and well-scoped.

Changes reviewed

1. torchao/quantization/linear_activation_quantized_tensor.py (deleted) — Entire 307-line file removed. This was the class definition, all aten dispatch implementations, and the to_linear_activation_quantized alias.

2. torchao/quantization/__init__.py — Import and __all__ entries for both LinearActivationQuantizedTensor and to_linear_activation_quantized removed cleanly.

3. torchao/quantization/quant_api.py — Import removed and the isinstance(mod.weight, LinearActivationQuantizedTensor) guard removed from _is_linear() (line 183). The remaining guards (_AffineFakeQuantizedTensor, NonDynamicallyQuantizableLinear) still prevent double-quantization of shared weights per the TODO comment at line 180.

4. torchao/quantization/utils.py — Import and the isinstance branch in _quantization_type() removed. The function falls through correctly to the hasattr(weight, "_quantization_type") check.

Minor nit: stale string references

There are a few leftover string literals referencing LinearActivationQuantizedTensor in other files. These are just comments/error messages and don't cause any functional issues, but cleaning them up would be nice for consistency:

• torchao/quantization/linear_activation_scale.py:98 — error message says "LinearActivationQuantizedTensor: No specialized dispatch found for linear op" but this file defines WeightTensorWithLinearActivationScaleMetadata
• torchao/quantization/linear_activation_scale.py:122 — comment says "Allow a model with LinearActivationQuantizedTensor weights..." but registers WeightTensorWithLinearActivationScaleMetadata
• torchao/quantization/weight_tensor_linear_activation_quantization.py:165 — same stale error message
• torchao/quantization/weight_tensor_linear_activation_quantization.py:204 — same stale comment
• torchao/quantization/linear_activation_weight_observed_tensor.py:156 — same stale comment

These look like they were copy-pasted from the original LinearActivationQuantizedTensor when creating the replacement classes and were never updated. Not blocking, but a follow-up to fix these strings would improve code clarity. Fix stale string references →

No issues found

• No remaining imports or runtime references to the deleted class
• No test files reference LinearActivationQuantizedTensor
• The _is_linear guard removal is safe — the class no longer exists so nothing can be an instance of it
• The _quantization_type branch removal is safe for the same reason

LGTM.

[jerryzh168]

we'll address minor issues by removing these subclass in follow up PRs