Removed `torch.ao.quantization.quantizer` imports in coremltools/optimize/torch break on torch 2.11 — request for decision on removal/deprecation of CoreMLQuantizer

[xyf5432]

Summary

Three files under coremltools/optimize/torch/quantization/ import from
torch.ao.quantization.quantizer, removed in torch 2.11 (PT2E quantization
migrated to torchao):

File	Lines
_annotation_config.py	11–13 from torch.ao.quantization.quantizer.quantizer import QuantizationSpec
_coreml_quantizer.py	12–13 from torch.ao.quantization.quantizer.quantizer import Quantizer / ...xnnpack_quantizer import _get_module_name_filter
_coreml_quantizer_utils.py	22–36 (8 lines: QuantizationSpec, xnnpack_quantizer, xnnpack_quantizer_utils imports)

Not wired into the public API, but still used by the test suite, and loading
any of them on torch 2.11+ fails immediately:
ModuleNotFoundError: No module named 'torch.ao.quantization.quantizer'.
Verified with torch 2.11.0 (crash) vs 2.9.1 (control, succeeds).

How to reproduce

pip install torch==2.11.0  # or newer; 2.9.x does not reproduce
pip install coremltools

from coremltools.optimize.torch.quantization._annotation_config import AnnotationConfig
# ModuleNotFoundError: No module named 'torch.ao.quantization.quantizer'

Related discussion: PR #2634

A migration attempt already exists: pytorch/pytorch#169151
removed PT2E quantization from torch, and jerryzh168 (pytorch/ao maintainer)
opened apple/coremltools#2634
"Update callsite for pt2e quant" migrating these files to torchao. It was
closed without merging (2026-01-24); the stated plan is to remove
CoreMLQuantizer from coremltools in favor of the ExecuTorch implementation
(executorch#16473).
See the PR for details.

Request

Please make a decision on the future of CoreMLQuantizer and act on it:

1. Remove the three modules (and their tests) as planned in PR Update callsite for pt2e quant #2634, or
   formally deprecate them with a clear support statement (including a
   torch < 2.11 requirement or a torchao dependency for torch 2.11+); or
2. At minimum, keep these files from crashing on torch 2.11+: delete the
   stale imports or gate them behind a version check with a clear error
   message pointing to the replacement.

We'd appreciate an explicit statement on the support status, since the
current release allows torch 2.11+ (unbounded torch requirement) and ships
files that crash on load.

Suggested fix

Given PR #2634, deletion is the preferred direction (migration is blocked by
the observer/fake_quant incompatibility noted in the PR discussion). If
keeping the modules, the minimal fix is a guarded import:

try:
    from torch.ao.quantization.quantizer.quantizer import (
        QuantizationSpec as _TorchQuantizationSpec,
    )
except ImportError:  # torch>=2.11: PT2E quantization moved to torchao
    from torchao.quantization.pt2e.quantizer.quantizer import (
        QuantizationSpec as _TorchQuantizationSpec,
    )

(torchao.quantization.pt2e.quantizer.quantizer.QuantizationSpec verified
present in torchao 0.16.0.)

References

• apple/coremltools#2634 — "Update callsite for pt2e quant" (closed, migration abandoned)
• pytorch/executorch#16473 — CoreMLQuantizer moved to ExecuTorch
• PyTorch migration plan: Torch.ao.quantization Migration Plan — PT2E quantization has been migrated to torchao
• Deprecation notice in Add deprecation warning for torch.ao.quantization pytorch/pytorch#153892 ("torch.ao.quantization is deprecated")
• torchao Quantizer API: torchao.quantization.pt2e.quantizer.quantizer.QuantizationSpec

[TobyRoseman]

The max version of PyTorch we support is 2.8. If you have a newer than that, you should get a warning when you import coremltools.

Related: #2615

[xyf5432]

Thanks for confirming — the warning does exist in _deps/init.py. However, install_requires has no torch constraint, so pip installs coremltools alongside any torch version, and users only hit the limit at import time. Since this break is guaranteed (module removal, not "unexpected errors"), could you add a packaging constraint (e.g. torch<2.9 in install_requires) so pip prevents the incompatible install up front, and/or state the deprecation plan for CoreMLQuantizer (per #2634)?

[TobyRoseman]

We don't have a packaging constraint for torch because torch is not a dependency for coremltools. You don't need to have torch installed in order to use coremltools. User may want to convert models from other frameworks (ex TensorFlow or XGBoost).

[xyf5432]

Thanks — agreed.

The supported range 2.1 – 2.8 is enforced in coremltools/_deps/init.py: _TORCH_MAX_VERSION = "2.8.0" warns at import time when torch is newer, and _CT_OPTIMIZE_TORCH_MIN_VERSION = "2.1.0" disables coremltools.optimize.torch with a warning when torch is older.

The README, the installation guide, and the PyTorch conversion page all omit this range, and the installation guide's "Install Third-party Packages" section is where users try to pick the torch version. Would you consider stating it there, e.g.:

coremltools is tested with PyTorch 2.1 – 2.8. Newer versions are not tested and may run into unexpected errors (a warning is emitted at import time). coremltools.optimize.torch requires PyTorch ≥ 2.1; importing it with an older version is disabled with a warning.