[bug fix] model free: fix copy folder of diffusion model, fix unresponsive disable_model_free, and cuda UT

[xin3he]

Description

This pull request introduces several improvements and fixes related to model-free quantization, especially for diffusion models, and enhances test coverage for metadata copying. The most significant changes include updating the documentation to clarify model-free quantization defaults, refactoring how model-free mode is triggered in the main entry point, improving metadata and subdirectory copying for diffusion models, and expanding the test suite to ensure correct behavior. Additionally, a minor fix is made to a Gemma3 model test.

Model-free quantization improvements:

• Updated documentation in both README.md and README_CN.md to clarify that model-free quantization is now the default when using --iters 0 --disable_opt_rtn, with links to relevant documentation. [1] [2]
• Refactored auto_round/__main__.py to remove the explicit model-free routing logic, as model-free mode is now handled internally by AutoRound. The main entry point now simply passes the relevant flags. [1] [2]

Diffusion model metadata handling:

• Enhanced the _copy_metadata_files method in auto_round/compressors/model_free.py to copy both root-level files and all sub-component directories (e.g., vae, scheduler, tokenizer) for diffusion models, ensuring the output directory is a faithful replica without overwriting the quantized transformer.

Testing improvements:

• Added comprehensive tests in test/test_cpu/quantization/test_model_free.py to verify correct copying of metadata and subfolders for both standard and diffusion models, and to ensure the quantized transformer is not overwritten during the process.

Evaluation and model handling:

• Moved the model.eval() call from the main quantization routine to the evaluation phase to ensure the model is in evaluation mode only when needed. [1] [2]

Minor fixes:

• Updated the Gemma3 model test to use the correct number of layers, improving test reliability.

Type of Change

Bug fix

Related Issues

Fixes or relates to #1800, #1778

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.

[xin3he]

/azp run Unit-Test-CUDA-AutoRound

[azure-pipelines]

Azure Pipelines successfully started running 1 pipeline(s).

[Copilot]

Pull request overview

This PR fixes and improves “model-free” quantization behavior (especially for diffusion models), simplifies CLI routing by delegating model-free selection to AutoRound, and expands unit tests to cover metadata/subfolder copying and CUDA model block-name detection.

Changes:

• Refactors auto_round/__main__.py to remove explicit model-free routing and instead pass model_free / disable_model_free through to AutoRound.
• Improves diffusion model metadata copying in model-free mode to include non-transformer subcomponent directories while preserving the quantized transformer/.
• Adds/updates tests for diffusion/non-diffusion metadata copying and adjusts Gemma3 CUDA test expectations for tiny-model layer counts.

Reviewed changes

Copilot reviewed 7 out of 7 changed files in this pull request and generated 3 comments.

Show a summary per file

File	Description
auto_round/__main__.py	Removes CLI-side model-free routing logic; forwards flags to AutoRound.
auto_round/compressors/model_free.py	Copies diffusion root metadata + subcomponent directories to output without overwriting quantized transformer.
auto_round/eval/evaluation.py	Moves model.eval() to evaluation entry point.
test/test_cpu/quantization/test_model_free.py	Adds tests for metadata/subfolder copying (diffusion + non-diffusion) and non-overwrite of quantized transformer.
test/test_cuda/models/test_get_block_name.py	Fixes Gemma3 tiny-model layer count expectations.
README.md	Documents updated model-free default behavior for --iters 0 --disable_opt_rtn.
README_CN.md	Chinese translation of the same model-free default documentation update.

[xin3he]

/azp run Unit-Test-CUDA-AutoRound

[azure-pipelines]

Azure Pipelines successfully started running 1 pipeline(s).

[wenhuach21]

is it possible to copy the folders regardless of whether it's model free or not

[xin3he]

is it possible to copy the folders regardless of whether it's model free or not

These are different saving paths, I had enhanced copy_python_files_from_model_cache to copy_folders before.

auto-round/auto_round/utils/model.py

Line 1742 in 5bbe39b

def copy_python_files_from_model_cache(model, save_path: str, copy_folders: bool | list[str] | tuple[str, ...] = False):

[xin3he]

/azp run Unit-Test-CUDA-AutoRound

[azure-pipelines]

Azure Pipelines successfully started running 1 pipeline(s).

[xin3he]

/azp run Unit-Test-CUDA-AutoRound

[azure-pipelines]

Azure Pipelines successfully started running 1 pipeline(s).

[xin3he]

/azp run Unit-Test-CUDA-AutoRound

[azure-pipelines]

Azure Pipelines successfully started running 1 pipeline(s).