Gemma3 implementation needs to be updated.

[onaka-ga-pkpk]

🚀 The feature, motivation and pitch

The Transformers implementation of Gemma3 has undergone various modifications, but the implementation in this repository does not reflect those changes and does not load Gemma3 properly.
For example, in transformers, the _update_causal_mask is already not used.

https://github.com/huggingface/transformers/blob/e6a8063ef1af16df964b644b07e1d17e96555d23/src/transformers/models/gemma3/modular_gemma3.py#L748-L749

def _update_causal_mask(self, **super_kwargs):
        raise AttributeError("We don't want to inherit it")

Liger-Kernel/src/liger_kernel/transformers/model/gemma3.py

Lines 260 to 262 in ecdf6de

	causal_mask = self._update_causal_mask(
	attention_mask, token_type_ids, past_key_values, cache_position, inputs_embeds, is_training
	)

Alternatives

No response

Additional context

No response