Implement AutoScheme support for model-free packing with validation a…

[xin3he]

Description

This pull request introduces comprehensive support for using AutoScheme (automatic mixed-bit quantization selection) in model-free mode, enabling a two-phase workflow: first, a brief model load for delta-loss scoring to select per-layer bit-widths, followed by efficient shard-by-shard packing without keeping the model in memory. The changes include both implementation and documentation updates, as well as new tests to ensure correct behavior and constraints.

Key changes:

Model-Free + AutoScheme Integration

• Added functions (_looks_like_auto_scheme, _validate_auto_scheme_options, _convert_auto_scheme_layer_config) to detect, validate, and process AutoScheme objects for model-free quantization, ensuring only supported option families (INT or MXFP) are allowed and not mixed. (auto_round/compressors/model_free.py auto_round/compressors/model_free.pyR1488-R1598)
• Implemented a two-phase workflow: the model is loaded only briefly for delta-loss scoring (selection), then released before model-free packing begins. The resolved per-layer scheme and ignore-lists are injected into the standard pipeline. (auto_round/compressors/model_free.py [1] [2] [3] [4]

Utilities and Routing

• Updated is_model_free_route to recognize and correctly route AutoScheme objects, including family validation and format restrictions. (auto_round/utils/model.py [1] [2]

Documentation

• Added detailed documentation and usage examples for combining AutoScheme with model-free mode, including constraints and supported option families, in both English and Chinese guides. (docs/step_by_step.md [1] [2] [3]; docs/step_by_step_CN.md [4] [5]

Testing

• Added a comprehensive test suite for model-free + AutoScheme, covering detection, validation (including error cases), layer config conversion, and end-to-end runs for both INT and MXFP families. (test/test_cpu/quantization/test_model_free.py test/test_cpu/quantization/test_model_free.pyR1310-R1391)

These changes ensure that users can now perform automatic mixed-bit quantization in model-free mode with robust validation, clear documentation, and strong test coverage.

Type of Change

Enhancement

Related Issues

Fixes or relates to #910

Checklist Before Submitting

• My code has been tested locally.
• Documentation has been updated as needed.
• New or updated tests are included where applicable.
• The CUDA CI has passed. You can trigger it by commenting /azp run Unit-Test-CUDA-AutoRound.

[wenhuach21]

I am not familiar with this part, and this change does not affect the major logic. So looks good to me

[Copilot]

Pull request overview

Adds support for using AutoScheme in model-free quantization by resolving mixed-bit per-layer choices via a brief delta-loss scoring phase, then performing shard-by-shard packing without keeping the full model resident. This also updates auto-routing logic, documentation, and CPU tests.

Changes:

• Implement AutoScheme detection/validation + conversion of generated layer_config into model-free inputs (two-phase selection → packing).
• Extend is_model_free_route to route AutoScheme runs based on option “family” (INT vs MXFP) and export format.
• Add docs and tests covering AutoScheme + model-free workflows and validation constraints.

Reviewed changes

Copilot reviewed 5 out of 5 changed files in this pull request and generated 4 comments.

Show a summary per file

File	Description
auto_round/compressors/model_free.py	Adds AutoScheme validation/conversion and a two-phase selection→packing workflow in model-free mode.
auto_round/utils/model.py	Updates model-free auto-routing to recognize and validate AutoScheme + format constraints.
docs/step_by_step.md	Documents AutoScheme + model-free usage, constraints, and examples.
docs/step_by_step_CN.md	Adds partial CN documentation updates for AutoScheme + model-free.
test/test_cpu/quantization/test_model_free.py	Adds unit + e2e tests for AutoScheme support in model-free mode.

[chensuyue]

/azp run Unit-Test-CUDA-AutoRound

[azure-pipelines]

Azure Pipelines successfully started running 1 pipeline(s).

[chensuyue]

/azp run Unit-Test-CUDA-AutoRound

[azure-pipelines]

Azure Pipelines successfully started running 1 pipeline(s).