Hanyuyuan6

Use when a user runs, debugs, verifies, or ships a DL experiment on a GPU they OWN or RENT (AutoDL, RunPod, vast.ai, Lambda, Paperspace, 恒源云/矩池云/Featurize/揽睿星舟, bare SSH, Slurm, K8s; single/multi-instance). Triggers (multilingual): 本地训练/local training, 远程 GPU 训练/租卡/GPU rental, spot 抢占/preemption, 断点续训/resumable, 防 SSH 断线/tmux 守护, 多实例 ablation, 关机/销毁/stop-vs-terminate billing, checkpoint 磁盘满, CUDA OOM/显存不足, loss NaN/spike/不收敛, overfit 单 batch, FSDP/DeepSpeed/torchrun, 多卡 hang, dataloader/数据增广 bug; 消融结果异常/ablation looks wrong, 复现/reproducibility, 数据泄漏/leakage/test-set tuning, mAP=0/全零指标, 输出恒定/model-ignores-input, train-good/val-collapse, 对比不公平/unfair baseline, 单 seed/no error bars, loss 太好/too-good-to-be-true, 跨文档对账/cross-doc drift; 交付产物/deliverable, 唯一真源/single source of truth, best ckpt 拉回, 结果可视化/论文图脚本, manifest/provenance, 一键复现/repro, EVIDENCE.json. NOT for multi-cloud price-shopping + auto spot-recovery (SkyPilot), BYOC dev environments (dstack), or zero-ops serverless inference (Modal).