Update transformers to v4.55

[hmellor]

Updates to the latest version of Transformers.

Notable changes:

• Transformers backend:
  • Get the tp_plan from the config of the base model because it is no longer added to the base model itself post init
  • Add a default mapping from nn.Linear to ReplicatedLinear to enable weight loaders like BitsAndBytesModelLoader to work with models that do not support TP if TP is disabled
  • Decoder modules no longer return tuple as of Refactor the way we handle outputs for new llamas and new models huggingface/transformers#39120 so return_tuple has been removed from PPMissingLayer.forward()

Enables the proper type hinting introduced by #21913

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.

💬 Join our developer Slack at https://slack.vllm.ai to discuss your PR in #pr-reviews, coordinate on features in #feat- channels, or join special interest groups in #sig- channels.

Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run fastcheck CI which starts running only a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of those by going to your fastcheck build on Buildkite UI (linked in the PR checks section) and unblock them. If you do not have permission to unblock, ping simon-mo or khluu to add you in our Buildkite org.

Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.

To run CI, PR reviewers can either: Add ready label to the PR or enable auto-merge.

🚀

[gemini-code-assist]

Code Review

This pull request updates the transformers library to version v4.54.1 and huggingface-hub to v0.34.3. While this is a necessary update, it introduces potential risks due to reliance on workarounds and private APIs from older transformers versions. I recommend addressing the following points to ensure stability and correctness. First, a critical issue in vllm/model_executor/models/transformers.py is the use of the private _attn_implementation API. With transformers v4.54.1, the public model.set_attn_implementation() should be used to avoid silently disabling vLLM's custom attention, which would impact performance and correctness. Second, there are high-risk, potentially obsolete workarounds in vllm/config.py for gemma2 and gemma3n_text model configs, and in vllm/transformers_utils/tokenizer.py for ChatGLMTokenizer. These may now be fixed in transformers and could cause incorrect behavior if they override updated library logic. Verifying and removing these if they are no longer necessary is highly recommended.

[DarkLight1337]

What failures are introduced by that PR? From my understanding those test failures also happen on main.

[hmellor]

I wasn't aware they were also on main. I assumed they were from that PR because many of them refer to the same configs that were modified.

[DarkLight1337]

Well in any case, it's great to fix them anyway

[DarkLight1337]

Ok I figured out that some of the failures in that PR are caused by missing trust_remote_code=True (because we previously loaded those configs from vLLM directly), let me open a separate PR.

[hmellor]

Ok, happy to review once it's open

[DarkLight1337]

Opened #21934

[DarkLight1337]

Hope this passes now

[noooop]

gte-Qwen2-1.5B-instruct hf Implementation does not support v4.54.1

[hmellor]

@noooop are you suggesting that we upper bound this model in tests to 4.53.2?

[hmellor]

Ok, I'll make that change in the next batch of commits (I'll let the tests run to completion)

[zyongye]

How long can we land this? Need this first before merging gpt-oss changes (#22259)

[DarkLight1337]

We just need to get CI to pass

[hmellor]

I understand this was a priority for gpt-oss, I'll work on hotfixiing anything that wasn't fully resolved in CI