Add mm_token_type_ids param to VL models lce_forward to fix ValueError

[albertvillanova]

Summary

This PR fixes the lce_forward function for VL models, adding support for mm_token_type_ids optional parameter related to multimodal processing:

• glm4v
• glm4v_moe
• qwen2_vl
• qwen2_5_vl
• qwen3_vl
• qwen3_vl_moe

Fix #1117.

This fixes a ValueError in model.generate() with transformers > 5.2.0, after they merged:

• 🚨 Unify 3D position ids huggingface/transformers#43972

See related issue downstream in TRL:

• ValueError: The following model_kwargs are not used by the model: ['mm_token_type_ids'] huggingface/trl#5216
• CI fails with dev dependencies: ValueError: The following model_kwargs are not used by the model: ['mm_token_type_ids'] huggingface/trl#5201

Details

Multimodal token type support:

• Added the mm_token_type_ids optional argument (of type torch.IntTensor) to the signature of lce_forward, allowing for the specification of multimodal token type IDs.
• Passed the mm_token_type_ids argument to the underlying model call, ensuring it is incorporated into the forward computation.

Testing Done

We checked this fix solves the issue downstream in TRL.

• Hardware Type:
• run make test to ensure correctness
• run make checkstyle to ensure code style
• run make test-convergence to ensure convergence

[albertvillanova]

CC: @kashif

[albertvillanova]

Transformers has just released v5.3.0 containing the PR I mentioned above:

• https://github.com/huggingface/transformers/releases/tag/v5.3.0
• 🚨 Unify 3D position ids huggingface/transformers#43972

Currently, liger-kernel throws a ValueError with the latest transformers version for the following models:

• glm4v
• glm4v_moe
• qwen2_vl
• qwen2_5_vl
• qwen3_vl
• qwen3_vl_moe

[Tcc0403]

LGTM