test(sft): capture warnings before prefetch

[ShigureNyako]

What

Fix the warning-capture race in test_streaming_dataset_missing_media_skip_refills_batch. Only tests/engine/sft/dataset/test_streaming.py changes: move one call and explain its ordering.

Why

Follow-up requested in #273. Verified against current main (c7d2847378cdf24c99b22aadb4b2fd078ff225a0): shuffle() calls PrefetchBuffer.set_index_order(), which starts background processing immediately. The missing-image warning can therefore be emitted before the test installs its capture callback.

How

Move ds.shuffle(0) after the logger.warning monkeypatch, inside the existing try/finally. This also keeps prefetch startup covered by ds.stop() cleanup. Preserve all existing inline/async/prefetch cases and assertions. No production changes, sleeps, retries, or new dependencies.

Testing

Local macOS arm64, Python 3.10.21:

• pre-commit run --all-files --show-diff-on-failure — all hooks passed using Python 3.10.
• python -m pytest tests/engine/sft/dataset/test_streaming.py -q — 57 passed.
• python -m pytest tests/engine/sft/dataset/ tests/utils/data/test_streaming_dataset.py -q -rs — 142 passed, 1 skipped (optional Parquet test: pyarrow unavailable).
• Full suite: python -m pytest tests/ -x -q --tb=short stopped during unrelated collection because fastapi is unavailable (23 skipped, 1 error). GPU/multi-node integration validation was not run: this host has no required GPU cluster.
• Additional tests: not needed; the existing three-mode regression test retains its assertions.
• Documentation: not applicable to this test-only fix.

Type of Change

• Bug fix (test-only)

Screenshots / Logs

A temporary local-only pytest probe wrapped the real shuffle() and joined its real prefetch thread before returning (10-second timeout; verified both rows were cached). With the same [prefetch] case and scheduling constraint, the original test failed 5/5 runs with empty captured warnings; the fixed test passed 5/5. All three modes also passed in five probe runs. This verifies the ordering mechanism, not the natural flake rate; the probe is not included in the patch.

No visual change; screenshots are not applicable.

@SigureMo Please review when convenient.

[rai-studio-bot]

Nyanpasu 审查看板

审查状态： ✅ 已通过

审查版本： 87f4ce9b0e6462fdb8067b9433bf70936cb68249

已完成审查，未发现需修改的问题；警告捕获已先于预取启动，三种模式及断言保留。本地测试受环境限制未完成；Pre-commit CI 已通过，其余 CI 尚在运行。

没有未解决的审查问题。

^{Powered by Nyanpasu with gpt-6-astra medium, please check the suggestions carefully.}

[rai-studio-bot]

已审查 87f4ce9b，未发现需要修改的问题。捕获回调现在先于预取线程启动，原有三种模式与断言保持不变，且启动过程纳入了 finally 清理。

本地测试未完成：默认环境缺少 PyTorch，另一现有环境启动 pytest 时退出码为 139。Pre-commit CI 已通过，其余 CI 尚在运行；本次未运行 GPU／多节点集成测试（无对应测试集群）。

^{Powered by Nyanpasu with gpt-6-astra medium, please check the suggestions carefully.}

[SigureMo]

LGTMeow

15 次 rerun 均通过，稳定性应该没问题了