[Feature Request] Add SwiGLU support for Qwen3-VL models

[matthewdm0816]

🚀 The feature, motivation and pitch

Problem Description

Can we add SwiGLU kernel support for Qwen3-VL models in Liger-Kernel?

Currently, it appears that passing swiglu=True in the apply_liger_kernel_to_qwen3_vl monkey patch function is a no-op and does not activate the kernel.

References

• Original PR: Support for Qwen3-VL models #911
• Relevant Discussion: Comment Link

Alternatives

No response

Additional context

No response

[dongchany]

Opened PR #1175 for this.

It fixes the missing Qwen3-VL SwiGLU patch path so apply_liger_kernel_to_qwen3_vl(..., swiglu=True) is
no longer a no-op.

What was missing:

• Qwen3VLTextMLP was not being replaced with LigerSwiGLUMLP
• instantiated decoder_layer.mlp modules were not being patched either

Validation run locally:

• python -m pytest -q test/transformers/test_monkey_patch.py -k 'qwen3_vl and not moe'
• python -m pytest -q test/convergence/bf16/test_mini_models.py -k 'mini_qwen3_vl and test_mini_model'
• python -m ruff check src/liger_kernel/transformers/monkey_patch.py test/transformers/ test_monkey_patch.py README.md