Enable SwiGLU patching for Qwen3-VL

[dongchany]

Summary

Fix apply_liger_kernel_to_qwen3_vl(..., swiglu=True) so it is no longer a no-op.

Changes

• patch transformers.models.qwen3_vl.modeling_qwen3_vl.Qwen3VLTextMLP to LigerSwiGLUMLP
• patch instantiated decoder_layer.mlp modules for existing Qwen3-VL model instances
• enable swiglu=True by default for apply_liger_kernel_to_qwen3_vl
• add monkey-patch tests covering both default instance patching and explicit swiglu=True
• document Qwen3-VL support in the README table

Validation

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

[Copilot]

Pull request overview

This PR fixes Qwen3-VL SwiGLU patching so apply_liger_kernel_to_qwen3_vl(..., swiglu=True) is effective (including for already-instantiated model instances), adds regression tests, and documents Qwen3-VL support.

Changes:

• Enable SwiGLU patching for Qwen3-VL by default and patch Qwen3VLTextMLP / existing decoder_layer.mlp instances.
• Add/extend monkey-patch tests to assert MLP forward patching behavior for Qwen3-VL.
• Update the README supported-models table to include Qwen3-VL.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 1 comment.

File	Description
src/liger_kernel/transformers/monkey_patch.py	Implements Qwen3-VL SwiGLU patching (module class + existing instances) and flips the default swiglu behavior.
test/transformers/test_monkey_patch.py	Adds assertions and a new test covering Qwen3-VL MLP patching via the swiglu flag.
README.md	Adds Qwen3-VL to the support matrix and adjusts Qwen-family rows.

Comments suppressed due to low confidence (1)

src/liger_kernel/transformers/monkey_patch.py:1799

• The apply_liger_kernel_to_qwen3_vl docstring’s Args: section doesn’t document the rope parameter even though it is part of the public signature (and is patched inside the function). Please add an rope (bool): ... entry for clarity and to keep it consistent with the other apply_liger_kernel_to_* docstrings in this module.

    """
    Apply Liger kernels to replace original implementation in HuggingFace Qwen3-VL models.

    Args:
        cross_entropy (bool): Whether to apply Liger's cross entropy loss. Default is False.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

This README table section appears to introduce mixed line endings (surrounding lines are CRLF, while the newly added/edited Qwen rows show as LF-only). Please normalize the line endings for the file (and ideally enforce via .editorconfig) to avoid noisy diffs and potential formatting issues on Windows tooling.