Traceback (most recent call last): File "repro.py", ... model = AutoModelForCausalLM.from_config(config) File ".../transformers/models/auto/auto_factory.py", line 241, in from_config return model_class._from_config(config, **kwargs) File ".../transformers/modeling_utils.py", line 1542, in _from_config model = cls(config, **kwargs) File ".../transformers/models/qwen3_5_moe/modeling_qwen3_5_moe.py", line 1906, in init self.model = Qwen3_5MoeTextModel(config) File ".../transformers/models/qwen3_5_moe/modeling_qwen3_5_moe.py", line 1339, in init self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, config.pad_token_id) File ".../transformers/configuration_utils.py", line 425, in getattribute return super().getattribute(key) AttributeError: 'Qwen3_5MoeConfig' object has no attribute 'vocab_size'

TL;DR

The issue can be fixed by unwrapping the config before passing it to from_config, similar to how from_pretrained handles it.

Guidance

• The error occurs because from_config does not unwrap the config, unlike from_pretrained.
• To fix this, you can manually unwrap the config before passing it to from_config using the get_text_config method.
• Verify that the model is correctly initialized after making this change.
• Check the auto_factory.py file for the correct implementation of from_config and from_pretrained to ensure symmetry.

Example

config = AutoConfig.from_pretrained("Qwen/Qwen3.6-35B-A3B")
config = config.get_text_config()  # unwrap the config
model = AutoModelForCausalLM.from_config(config)

Notes

This fix assumes that the get_text_config method is available for the Qwen3.5-MoE and Qwen3.6 family of models.

Recommendation

Apply the workaround by unwrapping the config before passing it to from_config, as shown in the example above. This will ensure symmetry with the from_pretrained method and allow the model to be correctly initialized.