[misc] chore: support transformers 5

[HollowMan6]

What does this PR do?

Refer to
vllm-project/vllm#30566
for all the patched needed for Transformers v5

This PR is a Transformers v5 compatibility sweep plus guardrails for token ID shape consistency.

• Remove the hard <5.0.0 block by changing dependency pinning in requirements.txt.
• Add a single compat resolver get_auto_model_for_vision2seq() in transformers_compat.py to handle AutoModelForVision2Seq vs AutoModelForImageTextToText, and switch model-loading/registration codepaths to use that resolver instead of direct imports.
• Introduce normalize_token_ids(...) in tokenizer.py, which normalizes apply_chat_template(tokenize=True) outputs to flat list[int] across v4/v5 return-shape differences.

Checklist Before Starting

• Search for similar PRs. Paste at least one query link here: ...
• Format the PR title as [{modules}] {type}: {description} (This will be checked by the CI)
  • {modules} include fsdp, megatron, veomni, sglang, vllm, rollout, trainer, ci, training_utils, recipe, hardware, deployment, ray, worker, single_controller, misc, perf, model, algo, env, tool, ckpt, doc, data, cfg, reward, fully_async, one_step_off
  • If this PR involves multiple modules, separate them with , like [megatron, fsdp, doc]
  • {type} is in feat, fix, refactor, chore, test
  • If this PR breaks any API (CLI arguments, config, function signature, etc.), add [BREAKING] to the beginning of the title.
  • Example: [BREAKING][fsdp, megatron] feat: dynamic batching

Test

For changes that can not be tested by CI (e.g., algorithm implementation, new model support), validate by experiment(s) and show results like training curve plots, evaluation results, etc.

API and Usage Example

Demonstrate how the API changes if any, and provide usage example(s) if possible.

# Add code snippet or script demonstrating how to use this

Design & Code Changes

Demonstrate the high-level design if this PR is complex, and list the specific changes.

Checklist Before Submitting

Important

Please check all the following items before requesting a review, otherwise the reviewer might deprioritize this PR for review.

• Read the Contribute Guide.
• Apply pre-commit checks: pre-commit install && pre-commit run --all-files --show-diff-on-failure --color=always
• Add / Update the documentation.
• Add unit or end-to-end test(s) to the CI workflow to cover all the code. If not feasible, explain why: ...
• Once your PR is ready for CI, send a message in the ci-request channel in the verl Slack workspace. (If not accessible, please try the Feishu group (飞书群).)
• If your PR is related to the recipe submodule, please also update the reference to the submodule commit via git submodule update --remote or cd recipe && git pull origin main.

[gemini-code-assist]

Code Review

This pull request successfully adds support for transformers version 5 by removing the upper version pin. The core of the change involves introducing compatibility layers to handle breaking changes, specifically the renaming of AutoModelForVision2Seq and the change in the return type of tokenizer.apply_chat_template. These changes are implemented cleanly using helper functions (get_auto_model_for_vision2seq and normalize_token_ids) and are applied consistently across the codebase. My main feedback is to consider adding an upper bound to the transformers dependency to prevent future breakages.

[Copilot]

Pull request overview

Updates the codebase to be compatible with Transformers 5 by centralizing version-dependent auto-model selection for vision-language models and normalizing tokenizer outputs across Transformers 4/5 behavior differences.

Changes:

• Add get_auto_model_for_vision2seq() compatibility helper and replace scattered version checks / direct imports.
• Add normalize_token_ids() and apply it across rollout, agent loops, dataset length filtering, and tests to handle apply_chat_template(tokenize=True) output shape/type changes.
• Relax requirements.txt Transformers constraint to allow installing Transformers 5.

Reviewed changes

Copilot reviewed 17 out of 18 changed files in this pull request and generated 2 comments.

Show a summary per file

File	Description
verl/workers/rollout/vllm_rollout/vllm_async_server.py	Normalize prompt_ids before length/context calculations for vLLM rollout.
verl/workers/fsdp_workers.py	Replace direct AutoModelForVision2Seq import with compatibility getter.
verl/utils/transformers_compat.py	Introduce get_auto_model_for_vision2seq() helper for Transformers 4/5.
verl/utils/tokenizer.py	Add normalize_token_ids() utility and export it.
verl/utils/model.py	Route Vision2Seq auto-class selection through the compat helper.
verl/utils/dataset/rl_dataset.py	Ensure chat-template tokenization is explicit and normalize token IDs for prompt-length filtering.
verl/utils/checkpoint/fsdp_checkpoint_manager.py	Replace inline Transformers version branching with compat helper when saving HF models.
verl/utils/chat_template.py	Normalize token IDs from apply_chat_template(tokenize=True) to keep logic stable across versions.
verl/model_merger/base_model_merger.py	Use compat helper for Vision2Seq selection in model merger logic.
verl/experimental/vla/models/register_vla_models.py	Use compat helper for registering VLA models with the correct auto class.
verl/experimental/fully_async_policy/vllm_rollout/vllm_async_server.py	Normalize prompt_ids before computing max tokens in async rollout.
verl/experimental/fully_async_policy/agent_loop/partial_single_turn_agent_loop.py	Normalize tokenized chat-template outputs before downstream use.
verl/experimental/agent_loop/agent_loop.py	Normalize token IDs for both processor and tokenizer chat-template paths.
tests/workers/rollout/rollout_vllm/test_vllm_abort.py	Update test to normalize chat-template tokenization output.
tests/experimental/reward_loop/test_reward_model_genrm.py	Normalize token IDs in reward-loop prompt/response token slicing.
tests/experimental/reward_loop/test_reward_model_disrm.py	Normalize token IDs in reward-loop prompt/response token slicing.
scripts/legacy_model_merger.py	Use compat helper for Vision2Seq auto class selection.
requirements.txt	Remove <5.0.0 upper bound to allow Transformers 5.

---

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