TDM3.0 beam search采样不会很耗时吗

TDM3.0训练时是采用beam search得到样本，然后根据样本的子节点max得到该节点的拟合值，但是一个样本假设耗时0.3s,一个batch有2048个样本，那一个batch的耗时不就将近600s?这不会超级耗时吗

No file, test, or entry point is named. Start by tracing TDM3.0 training's beam-search sampling and batch-size handling, then benchmark one batch against the reported 0.3-second per-sample estimate. Done means establishing whether the implementation avoids a serial cost for all 2048 samples or documenting the actual batch-time behavior.

Written by the indexing model from the issue text.