[Bug] predict_with_generate=True causes 1D dimension collapse of eval_prediction.predictions in custom metrics (DDP)

[coldchair]

Checklist / 检查清单

• I have searched existing issues, and this is a new bug report. / 我已经搜索过现有的 issues，确认这是一个新的 bug report。

Bug Description / Bug 描述

Version
4.0.0.dev0

Describe the bug
When --predict_with_generate true is enabled in a multi-GPU (DDP) environment, the eval_prediction.predictions passed to a custom EvalMetrics subclass collapses from a 2D tensor (BatchSize, SeqLen) into a flattened 1D array.

This causes Serializer.from_tensor(preds) to fail with _pickle.UnpicklingError or IndexError because the data structure is corrupted during the gathering process.

Root Cause
In Seq2SeqTrainer.prediction_step, pad_sequence only performs local padding within each GPU. If GPU 0 and GPU 1 have different maximum sequence lengths for their local batches, accelerator.gather cannot stack them and flattens the tensors instead.

How to Reproduce / 如何复现

Run SFT with --predict_with_generate true on 2+ GPUs.
Use a custom EvalMetrics that accesses eval_prediction.predictions.
Observe that preds.ndim is 1 and the evaluation crashes.

Additional Information / 补充信息

No response