support gptqmodel 7.0.0 and fix bug in CI

[xin3he]

Description

GPTQmodel 7.0.0 removes Quant string from Linear class name

Type of Change

Bug fix

Related Issues

Fixes or relates to #1771

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

Updates AutoRound’s GPTQModel integration to accommodate GPTQModel 7.0.0 API/class renames and adjusts some test helper logic used in CI/model loading.

Changes:

• Add GPTQModel 7.0.0 compatibility for Marlin kernels and ExllamaV2 class naming changes.
• Refactor GPTQModel backend linear selection to unify GPTQ/AWQ handling and dynamically resolve renamed classes.
• Simplify get_tiny_model(..., from_config=True) config-only model construction path in tests.

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 3 comments.

File	Description
test/helpers.py	Adjusts model-path selection and refactors config-only tiny model instantiation logic.
auto_round_extension/cuda/gptqmodel_marlin.py	Adds GPTQModel 7.0.0 Marlin import/argument handling.
auto_round/inference/convert_model.py	Switches ExllamaV2 class import based on GPTQModel version.
auto_round/inference/backend.py	Unifies GPTQModel GPTQ/AWQ backend routing and resolves renamed QuantLinear classes via dynamic import.

[Copilot]

architectures can be None when the config doesn’t define an architectures list (you default to [None]). In that case architectures.endswith("Model") will raise an AttributeError. Add a guard (e.g., ensure architectures is a str before calling string methods) and fall back to AutoModelForCausalLM when it’s missing/empty.

Suggested change

	architectures = getattr(config, "architectures", [None])[0]
	if (
	is_mllm
	and architectures.endswith("Model")
	and hasattr(base_lib, n := architectures.replace("Model", "ForConditionalGeneration"))
	):
	model_cls = getattr(base_lib, n)
	elif hasattr(base_lib, architectures):
	config_architectures = getattr(config, "architectures", None)
	architectures = (
	config_architectures[0]
	if isinstance(config_architectures, (list, tuple)) and len(config_architectures) > 0
	else None
	)
	if (
	is_mllm
	and isinstance(architectures, str)
	and architectures.endswith("Model")
	and hasattr(base_lib, n := architectures.replace("Model", "ForConditionalGeneration"))
	):
	model_cls = getattr(base_lib, n)
	elif isinstance(architectures, str) and hasattr(base_lib, architectures):

[Copilot]

For gptqmodel>=7.0.0 you import gptqmodel.utils.marlin without handling ImportError, but older versions store the exception and raise a clear ValueError when the backend is used. Wrap the v7 import in the same try/except and set marlin_import_exception so failures produce a consistent, actionable error message.

[Copilot]

This if "gptqmodel" in backend branch inside the generic "awq" in backend block is now unreachable because the earlier if "gptqmodel" in backend: returns first for all gptqmodel backends. Consider removing this dead branch to avoid confusion and keep backend routing maintainable.