[fix]Add logits_to_keep and shift_labels support for Qwen3-VL and Qwen3-VL-MoE

[luca-888]

Summary

This PR adds logits_to_keep and shift_labels support for both Qwen3-VL and Qwen3-VL-MoE in the Liger-patched forward path. The change aligns the patched implementation with the expected Hugging Face interface and enables selective logits materialization for long-context inference.

Testing Done

• make test
  • not fully green
  • observed existing failures in GRPO, fused_neighborhood_attention, and gemma3 monkey patch tests
• make test-convergence
  • not fully green
  • observed failure in test/convergence/bf16/test_mini_models_multimodal.py::test_mini_model_multimodal[mini_llama4-...]
• make checkstyle
  • passed

Known limitation:

• The failed make test / make test-convergence cases above do not directly exercise the Qwen3-VL or Qwen3-VL-MoE logits_to_keep / shift_labels change in this PR

[luca-888]

Test Environment

• GPU: single NVIDIA H20 (~95 GB VRAM)
• transformers==5.4.0
• Steady-state latency, model load time excluded
• Each metric uses 1 warmup iteration and 2 measured iterations
• Long-text prompt generated by repeating the same sentence with repeat counts 40 / 160 / 320 / 640
• Measured with full logits vs logits_to_keep=1

Long-Sequence Benchmark

Dense: Qwen3-VL-8B-Instruct

Sequence Length	Full Logits Latency	Keep1 Latency	Latency Delta	Full Peak Mem	Keep1 Peak Mem	Memory Delta
755	0.1111 s	0.1051 s	-5.4%	16.809 GB	16.674 GB	-0.135 GB
2795	0.3733 s	0.3484 s	-6.7%	17.979 GB	17.469 GB	-0.510 GB
5515	0.7592 s	0.7097 s	-6.5%	19.539 GB	18.531 GB	-1.008 GB
10955	1.6723 s	1.5753 s	-5.8%	22.660 GB	20.655 GB	-2.005 GB

MoE: Qwen3-VL-30B-A3B-Instruct

Sequence Length	Full Logits Latency	Keep1 Latency	Latency Delta	Full Peak Mem	Keep1 Peak Mem	Memory Delta
755	0.2218 s	0.2245 s	+1.2%	58.348 GB	58.313 GB	-0.035 GB
2795	0.3642 s	0.3510 s	-3.6%	59.511 GB	59.365 GB	-0.146 GB
5515	0.6716 s	0.6461 s	-3.8%	61.061 GB	60.772 GB	-0.289 GB
10955	1.4846 s	1.4352 s	-3.3%	64.161 GB	63.585 GB	-0.576 GB

[luca-888]

@Mecoli1219 @Tcc0403 Could you please review this PR and approve the pending workflows when you have a chance? The implementation follows the same approach currently used for qwen3, qwen3-moe, and qwen3.5 in the repository. Thank you!

[Mecoli1219]

Overall looks good to me. Thanks for contribution!

[luca-888]

@Mecoli1219 Since it already has approval, could this be merged if there are no further concerns?

[Mecoli1219]

Yes. Feel free to merge it!

[luca-888]

Yes. Feel free to merge it!

If you have permission, could you please merge this PR when you have a moment? Thanks!