optimize-training

Diagnose and optimize deep-learning training pipelines with measured baselines, correctness gates, and reversible experiments. Use for slow PyTorch training, low GPU utilization, data stalls, OOM, torch.compile regressions, DDP/FSDP scaling, cross-framework numerical alignment, data-shard audits, augmentation changes, or checkpoint conversion. Also use for 训练加速、训练吞吐、显存优化、精度对齐. Do not use for inference-only serving optimization or architecture search unrelated to training-system performance.