add mimo-audio, Qwen-TTS model backbone quantization

[WeiweiZhang1]

Description

Support mimo-audio, Qwen-TTS model quantizationc(for new-architecture)

Type of Change

New feature

Related Issues

Fixes or relates to #

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.

[Copilot]

Pull request overview

Adds audio/omni model quantization support paths for MiMo-Audio, Qwen3-TTS, StableAudio, and AudioCaps calibration data across MLLM and diffusion flows.

Changes:

• Registers audio model templates, processors, block handlers, and model-type resolution.
• Adds AudioCaps calibration dataset support for text and diffusion pipelines.
• Adds StableAudio pipeline/block handling and related documentation updates.

Reviewed changes

Copilot reviewed 15 out of 15 changed files in this pull request and generated 8 comments.

Show a summary per file

File	Description
test/README.md	Updates model test category wording for Omni.
test/helpers.py	Extends tiny-model config reduction for audio/TTS configs.
auto_round/utils/model.py	Adds Qwen3-TTS loading and StableAudio pipeline fallback handling.
auto_round/special_model_handler.py	Adds MiMo-Audio/Qwen3-TTS model resolution, block handlers, and forward patching.
auto_round/compressors/mllm/template.py	Registers MiMo-Audio and Qwen3-TTS templates.
auto_round/compressors/mllm/README.md	Documents audio/omni MLLM support.
auto_round/compressors/mllm/processor.py	Adds text-only processors for MiMo-Audio and Qwen3-TTS.
auto_round/compressors/mllm/compressor.py	Uses resolved model type for template/dataset compatibility.
auto_round/compressors/diffusion/README.md	Adds StableAudio to diffusion validation table.
auto_round/compressors/diffusion/dataset.py	Adds AudioCaps CSV loading for diffusion calibration.
auto_round/compressors/diffusion/compressor.py	Adds StableAudio block output mapping.
auto_round/compressors_new/mllm_mixin.py	Uses resolved model type for new MLLM template selection.
auto_round/compressors_new/diffusion_mixin.py	Uses attached custom diffusion pipeline functions during calibration.
auto_round/calib_dataset.py	Adds AudioCaps as a text calibration dataset.
auto_round/algorithms/quantization/base.py	Adds StableAudio block input mapping to base quantization config.

Comments suppressed due to low confidence (2)

auto_round/special_model_handler.py:442

• The newly registered audio model block handlers are not covered by tests. Similar multimodal handlers have tests for block discovery, forward patching, processor registration, and template registration in test/test_cpu/models/test_omni_model.py; please add coverage for the MiMo-Audio and Qwen3-TTS paths so regressions in resolve_model_type, _handle_special_model, and template/processor registration are caught.

    position_ids=None,
    attention_mask=None,

auto_round/special_model_handler.py:430

• This fallback is also skipped when a talker.model attribute exists but does not expose layers, so the standard model.model.layers layout described above would not be discovered. Please make the layout checks independent or fall through when a candidate layout does not actually add any block names.

    "qwen2_5_omni": _get_qwen2_5_omni_multimodal_block,

[n1ck-guo]

LGTM

[WeiweiZhang1]

/azp run Unit-Test-CUDA-AutoRound

[azure-pipelines]

Azure Pipelines successfully started running 1 pipeline(s).

[yiliu30]

Please add some UTs using dummy models; you may refer to:

auto-round/test/test_cuda/models/test_omni_model.py

Lines 55 to 59 in cb9b0b7

	@pytest.mark.skip_ci(reason="Not necessary to test all cases in CI")
	class TestQwen2_5OmniQuantization:
	"""End-to-end quantization test for Qwen2.5-Omni (dense model)."""
	def test_quantize_and_reload(self, tiny_qwen2_5_omni_model_path, tmp_path):

Left a few nitpicks; overall LGTM.