Introduce backend rollout-completions interface and decouple OpenEnv helper from vLLM internals

[rycerzes]

Summary

Closes #5194, (previous step #5244, part of #5119) adds an internal rollout-completions capability to backend.py and refactors utils.py to dispatch through trainer.generation_backend, removing direct trainer/backend introspection in helper flow.

Changes

• backend.py:
  • add RolloutCompletion dataclass
  • add internal RolloutCompletionsBackend protocol
  • implement generate_rollout_completions(...) in VLLMBackendAdapter (server + colocate)
  • raise explicit unsupported-capability errors in non-vLLM adapters
• utils.py:
  • remove direct branching on trainer.use_vllm / trainer.vllm_mode
  • remove direct trainer.vllm_generation.* calls
  • route via trainer.generation_backend.generate_rollout_completions(...)
  • preserve output schema: prompt_ids, completion_ids, logprobs, text
• Tests:
  • expanded test_generation_backend.py
  • new test_openenv_utils.py

Preserved

• Existing OpenEnv helper API and return contract
• Existing OpenEnv example call sites
• Backend-specific generation/runtime details remain in adapter layer

CC: @albertvillanova

---

Note

^{Cursor Bugbot is generating a summary for commit f9bb56b. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

[cursor]

super(type(trainer)) breaks for subclassed trainers

Medium Severity

super(type(trainer), trainer)._prepare_inputs resolves based on the runtime class, not a fixed class. If a user subclasses GRPOTrainer (or RLOOTrainer), type(trainer) is the subclass, and super() lands on the trainer-specific _prepare_inputs override instead of the base Trainer._prepare_inputs. The old inline code used Python 3's argument-free super() inside the trainer method, which always resolved relative to the defining class (GRPOTrainer), correctly skipping its own override. The new standalone factory function can't use __class__-based super(), and the type()-based workaround doesn't skip enough MRO levels for subclassed trainers.