Fix GlmMoeDsaConfig default mlp_layer_types in modular conversion

[OiPunk]

Summary

This PR fixes #43864 by preserving the GlmMoeDsaConfig default mlp_layer_types from the modular source.

GlmMoeDsaConfig should default to dense MLP for the first 3 layers and sparse afterward. During modular conversion, the parent init body was being inlined and overwrote that default with the parent pattern (["dense"] + ["sparse"] * ...).

Changes

• In modular_glm_moe_dsa.py, call PreTrainedConfig.__init__(self, **kwargs) instead of super().__init__(**kwargs) to avoid inlining parent init logic.
• Regenerated configuration_glm_moe_dsa.py via modular converter, which removes the duplicated parent default block.
• Added a regression test in tests/models/glm_moe_dsa/test_configuration_glm_moe_dsa.py to assert the expected default pattern for num_hidden_layers=8.

Validation

• PYTHONPATH=src python3 utils/modular_model_converter.py glm_moe_dsa
• PYTHONPATH=src python3 utils/check_modular_conversion.py --files src/transformers/models/glm_moe_dsa/modular_glm_moe_dsa.py
• PYTHONPATH=src python3 -m pytest tests/models/glm_moe_dsa/test_configuration_glm_moe_dsa.py -q
• PYTHONPATH=src python3 -m trace --count --summary --module unittest tests.models.glm_moe_dsa.test_configuration_glm_moe_dsa | grep -E "configuration_glm_moe_dsa|test_configuration_glm_moe_dsa"
  • output: configuration_glm_moe_dsa ... 100%

[OiPunk]

Thanks for the detailed review. I pushed commit a10f430 to address the requested changes.

What I changed:

• Removed duplicate attribute assignments in modular_glm_moe_dsa.py so each config field is assigned once.
• Regenerated configuration_glm_moe_dsa.py from the modular source to keep generated output in sync.
• Moved the default mlp_layer_types regression test into test_modeling_glm_moe_dsa.py::GlmMoeDsaModelTest and removed the standalone configuration test file.

Validation run locally:

• PYTHONPATH=src python3 utils/modular_model_converter.py glm_moe_dsa
• PYTHONPATH=src python3 utils/check_modular_conversion.py --files src/transformers/models/glm_moe_dsa/modular_glm_moe_dsa.py
• PYTHONPATH=src python3 -m ruff check src/transformers/models/glm_moe_dsa/modular_glm_moe_dsa.py src/transformers/models/glm_moe_dsa/configuration_glm_moe_dsa.py tests/models/glm_moe_dsa/test_modeling_glm_moe_dsa.py
• PYTHONPATH=src pytest -q tests/models/glm_moe_dsa/test_modeling_glm_moe_dsa.py -k default_mlp_layer_types

I also verified both mlp_layer_types paths (None and explicit list) execute in the config initializer.

[zucchini-nlp]

run-slow: glm_moe_dsa

[github-actions]

This comment contains run-slow, running the specified jobs:

models: ["models/glm_moe_dsa"]
quantizations: []

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[github-actions]

CI Results

Workflow Run ⚙️

Commit Info

Context	Commit	Description
RUN	2415753c	merge commit
PR	a10f4303	branch commit
main	884749a1	base commit

✅ No failing test specific to this PR 🎉 👏 !

[zucchini-nlp]

@bot /style

[github-actions]

Style fix bot fixed some files and pushed the changes.

[github-actions]

[For maintainers] Suggested jobs to run (before merge)

run-slow: glm_moe_dsa

[zucchini-nlp]

Thanks