fix: add Qwen3.5 related changes

[zpqiu]

What does this PR do ?

TODO:

• fix: MoE parallelizer config lookup for VLM models with nested text_config Automodel#1466
• [Qwen3.5 MoE] Qwen3.5-MoE state dict adapter crashes with DTensor inputs during DCP checkpoint loading Automodel#1420
• Pin Automodel to latest commit
• vllm fix: zpqiu/vllm@89f73ac
• vllm depends on transformers 5.3

cc: @sharonyu-115

Issues

List issues that this PR closes (syntax):

Usage

• You can potentially add a usage example below

# Add a code snippet demonstrating how to use this

Before your PR is "Ready for review"

Pre checks:

• Make sure you read and followed Contributor guidelines
• Did you write any new necessary tests?
• Did you run the unit tests and functional tests locally? Visit our Testing Guide for how to run tests
• Did you add or update any necessary documentation? Visit our Document Development Guide for how to write, build and test the docs.

Additional Information

• ...

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[terrykong]

Thanks @zpqiu for the contribution! What is the order of PRs that we need to land? If the v5 PR goes in, can this go in? Or do we need to wait for some vllm one to go in too?

[sharonyu-115]

Thanks @zpqiu for the contribution! What is the order of PRs that we need to land? If the v5 PR goes in, can this go in? Or do we need to wait for some vllm one to go in too?

Hi @terrykong This PR can go in following the v5 PR. But to have Qwen3.5 working, the AutoModel fixes and vllm fix are required. @zpqiu Shall we open a PR to vllm?

BTW, @terrykong Qwen3.5 requires transformers v5.3.0. Will you cover this in the v5 PR?

[zpqiu]

Before vLLM bumps its transformers to v5, will RL install from https://github.com/terrykong/vllm first?
If so, I think we can merge the Qwen 3.5 changes into @terrykong’s fork first.

[zpqiu]

BTW, @terrykong Qwen3.5 requires transformers v5.3.0. Will you cover this in the v5 PR?

Yes. Otherwise, we’ll hit the following issue:

File "/opt/ray_venvs/nemo_rl.models.generation.vllm.vllm_worker.VllmGenerationWorker/lib/python3.12/site-packages/transformers/modeling_rope_utils.py", line 651, in convert_rope_params_to_dict
    ignore_keys_at_rope_validation = ignore_keys_at_rope_validation | {"partial_rotary_factor"}
                                     ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~^~~~~~~~~~~~~~~~~~~~~~~~~~~
TypeError: unsupported operand type(s) for |: 'list' and 'set'

[terrykong]

thanks for the heads up @sharonyu-115 @zpqiu . i'm resolving the issues from the upgrade, and will try 5.3 after i resolve those