What does this PR do?

Type of change: Bug fix

Fixes #201 - get_modelike_from_algo_cfg rejected QuantizeAlgorithmConfig instances with ValueError, despite the algorithm field's type annotation allowing them.

Root cause: QuantizeAlgoCfgType (modelopt/torch/quantization/config.py) explicitly includes QuantizeAlgorithmConfig, and QuantizeConfig.algorithm is typed with it. But get_modelike_from_algo_cfg only matched list / None / str / dict and fell through to raise ValueError(f"Invalid config type: {type(algo_cfg)}") for a config instance:

ValueError: Invalid config type: <class 'modelopt.torch.quantization.config.MaxCalibConfig'>

Fix: ModeloptBaseConfig is a collections.abc.MutableMapping, so the type cascade's dict check is widened to Mapping, and the matched config is materialized with dict(...) so the resolved config keeps its ConfigDict (dict[str, Any]) type:

elif isinstance(algo_cfg, Mapping):
    # Normalize any mapping (incl. ModeloptBaseConfig instances) to a plain dict.
    algo_cfg = dict(algo_cfg)
    algo_name = algo_cfg["method"]

The change is a strict no-op for the previously supported str / dict / None / list inputs (a plain dict is a Mapping); only the path that previously raised changes.

Usage

import modelopt.torch.quantization as mtq

cfg = mtq.INT8_DEFAULT_CFG
cfg["algorithm"] = mtq.MaxCalibConfig(method="max", distributed_sync=False)

# Previously: ValueError: Invalid config type: <class '...MaxCalibConfig'>
# Now accepted, consistent with the `algorithm` field's type annotation.
mtq.quantize(model, cfg, forward_loop)

Testing

Added two regression tests:

• tests/unit/torch/quantization/test_mode.py::test_get_modelike_from_algo_cfg_accepts_config_instance - a single config instance and a list of config instances (resolved mode names, dict round-trip).
• tests/unit/torch/quantization/test_quantize_cpu.py::test_quantize_accepts_algo_config_instance_end_to_end - the issue's repro through the full mtq.quantize / calibrate flow.

pytest tests/unit/torch/quantization/test_mode.py::test_get_modelike_from_algo_cfg_accepts_config_instance \
       tests/unit/torch/quantization/test_quantize_cpu.py::test_quantize_accepts_algo_config_instance_end_to_end

Without the fix both tests fail with the exact ValueError: Invalid config type; with it they pass. The full test_mode.py and test_quantize_cpu.py suites pass locally, and ruff check / ruff format --check are clean on the changed files.

Before your PR is "Ready for review"

Make sure you read and follow the Contributor guidelines and your commits are signed (git commit -s -S).

Make sure you read and follow the Security Best Practices for contributors.

• Is this change backward compatible?: ✅
• If you copied code from any other sources or added a new PIP dependency, did you follow guidance in CONTRIBUTING.md: N/A
• Did you write any new necessary tests?: ✅
• Did you update Changelog?: ✅
• Did you get Claude approval on this PR?: N/A

Additional Information

Closes #201. Bug confirmed by @realAsma on the issue thread. The fix uses the Mapping-based approach suggested by @shengliangxu in review (ModeloptBaseConfig became a MutableMapping in #1405). Change is +41/-4 across modelopt/torch/quantization/mode.py, the two test files, and CHANGELOG.rst.

Summary by CodeRabbit

• Bug Fixes

  • Quantization now accepts algorithm configuration objects as the algorithm parameter instead of rejecting them with an error.

• Tests

  • Added regression and end-to-end tests to verify algorithm config instances are accepted and preserved through the quantization workflow.

• Documentation

  • Changelog updated to record the bug fix in version 0.45.