[albertvillanova]

This PR runs the GOLD on-policy generation tests on a real trainer, its real tokenizer and model, instead of a trainer built with __new__, hand-written tokenizers and a scripted model.

Part of #7565.

Motivation

These tests skipped __init__, set the attributes the method under test read by hand, and decoded with fake tokenizers that mapped a few token ids to letters, so they could not notice a change in how the trainer is set up or in how a real tokenizer decodes special and pad tokens. One of them replaced the trainer's own _process_completions_to_buffer to capture its arguments. Only vLLM cannot run in the test environment.

Changes

• GOLD: the on-policy generation tests (completion buffering, vLLM prompt handling and budgeting, and the non-vLLM fallback) build a real trainer from a tiny model and take their token ids and expected texts from its tokenizer.
• The non-vLLM tests generate with the real model and check the budgeted prompt and the completion length, without patching unwrap_model_for_generation.
• vLLM is replaced by a stand-in that records the prompts it receives and completes each one with a fixed token.
• The test that captured _process_completions_to_buffer's arguments now checks the buffered inputs it produces.

---

Note

Low Risk
Test-only refactor with no production code changes; slightly heavier tests due to real model init and generation.

Overview
Refactors GOLD on-policy generation tests to exercise a fully constructed GOLDTrainer (tiny Qwen2 student/teacher + zen dataset) instead of __new__ shells with hand-rolled tokenizers and dummy models.

Adds shared helpers _gold_trainer, _gold_vllm_trainer, and RecordingVLLMGeneration so vLLM is stubbed while everything else stays real. Tests now derive token ids and expected prompt/completion text from the trainer’s tokenizer, call instance methods (_process_completions_to_buffer, _generate_on_policy_for_slices, _generate_non_vllm_for_slices), and assert buffered tensors/text logs end-to-end. The non-vLLM path no longer patches unwrap_model_for_generation or fakes model.generate; the completion-buffer test asserts outputs instead of intercepting _process_completions_to_buffer arguments.

^{Reviewed by Cursor Bugbot for commit 19d108e. Bugbot is set up for automated code reviews on this repo. Configure here.}

[bot-ci-comment]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.