Fix ModelOpt integration for the >=0.44 list quant_cfg format

[shengliangxu]

What does this PR do?

Fixes NVIDIA/Model-Optimizer#2001

ModelOpt 0.44 changed quant_cfg and mtq.config._default_disabled_quantizer_cfg from a {pattern: cfg} mapping to a list of {"quantizer_name": ...} entries. NVIDIAModelOptConfig.get_config_from_quant_type spread the default-disabled set into a dict literal, so on ModelOpt >=0.44 building the config raised TypeError: 'list' object is not a mapping — hitting the default NVIDIAModelOptConfig(quant_type=...) path.

Emit quant_cfg in whichever shape the installed ModelOpt expects, keyed off the shape ModelOpt exposes for its default-disabled set: a mapping for <0.44 (unchanged behavior) and the native list of entries for >=0.44 (no deprecation warning). The modules_to_not_convert loop in the quantizer appends a list entry or assigns a mapping key accordingly.

Add a CPU-only regression test that builds the config across quant types and checks the installed ModelOpt accepts it — the existing GPU tests are nightly/big-accelerator gated, so this config-construction crash slipped through CI.

Self-review — Fix ModelOpt integration for the >=0.44 list quant_cfg format

Scope: src/diffusers/quantizers/quantization_config.py, src/diffusers/quantizers/modelopt/modelopt_quantizer.py, tests/quantization/modelopt/{__init__.py,test_modelopt.py}.

Blocking issues

None. The fix is correct on both supported ModelOpt eras — verified by running the new test against real 0.43.0 (dict branch) and 0.45.0 (list branch), 6/6 each. The <0.44 dict path is byte-for-byte identical to the pre-fix code (checked against the original method extracted from git).

Non-blocking issues

1. One-caller test helper — inline candidate. _weight_quantizer_num_bits (tests/quantization/modelopt/test_modelopt.py:46) has a single caller. Per .ai/AGENTS.md: "If a private helper has only one caller, inlining it at the call site is usually the cleaner choice." It names a real dict-vs-list abstraction, so it's defensible; by the rule it leans toward inlining. Trivial.
2. CPU validation via a ModelOpt-internal API (advisory for the maintainer). The test asserts acceptance with mtq.config.QuantizeConfig(**modelopt_config) (test_modelopt.py:88) rather than a full mto.apply_mode on a model — a deliberate CPU-only tradeoff so the regression is caught without a GPU. Worth surfacing to the reviewer, not fixing.

Rule checks that passed (not issues)

• No defensive code (AGENTS.md): the isinstance(default_disabled_quantizer_cfg, dict) branch and the elif "enable" not in entry: entry["enable"] = True fallback are both reachable and tested (dict path on 0.43; the enable:True case on FP8 weight_only=False), so they're not "just in case."
• Ephemeral context (review-rules.md §Common mistakes): all three comments state the reason and stand alone; the doc URL is a stable pointer. None are PR-scoped chatter.
• Testing conventions (testing.md): backend-tier config unit test mirroring the retained TestTorchAoConfig pattern — pytest-style, @is_quantization/@is_modelopt markers (both registered in conftest.py), license header. Pipeline/model tester-mixin rules don't apply. No LoRA/slow/integration added.
• Copied code: none of the touched code is under a # Copied from header.
• Style: ruff check + ruff format --check clean.

Dead code (advisory)

Rubric's dead-code pass is scoped to new-model PRs — N/A here.

path:line	status	reason
quantization_config.py:880 (dict branch)	Used	exercised on ModelOpt <0.44 (verified on 0.43.0)
quantization_config.py:894 (enable:True)	Used	FP8 weight_only=False param in the test
modelopt_quantizer.py:173 (list append)	Used	>=0.44 modules_to_not_convert path

Documentation impact

No public API changed (behavior-restoring fix). docs/source/en/quantization/modelopt.md documents the public quant_type/modules_to_not_convert/… params — untouched, no staleness introduced. No agent-guide addition warranted (the dict->list gotcha is captured in the code comment + doc link).

Summary — READY

Minimal, correct on both ModelOpt eras, lint-clean, and covered by a CPU regression test that would have caught the original TypeError (which the pre-existing GPU/nightly-gated tests could not).

• Fix before submitting: nothing required.
• Optional polish (your call): inline _weight_quantizer_num_bits (Add glide modeling files #1).
• Leave for the actual review: the QuantizeConfig-vs-apply_mode CPU-validation tradeoff (+ cosine schedule and unet config #2) — flag it to the maintainer rather than pre-emptively changing it.

Before submitting

• [ x ] Did you use an AI agent (Claude Code, Codex, Cursor, etc.) to help with this PR? If so:
  • [ x ] Did you read the Coding with AI agents guide?
  • [ x ] Did you run the self-review skill on the diff?
  • [ x ] Did you share the final self-review notes in the PR description or a comment?
• [ x ] Did you read the contributor guideline?
• [ x ] Did you read our philosophy doc? (important for complex PRs)
• [ x ] Was this discussed/approved via a GitHub issue or the forum? Please add a link to it if that's the case.
• [ x ] Did you make sure to update the documentation with your changes? Here are the
  documentation guidelines, and
  here are tips on formatting docstrings.
• [ x ] Did you write any new necessary tests?
• [ x ] Are you the author (or part of the team) of the model/pipeline (only applicable for model/pipeline related PRs)?

Who can review?

Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.