CHDTevior

整理原作资料、角色设定、时间线和跨作兼容规则，建立 canon contract。

管理支线、试验性改线和 retcon，避免污染主线。

对现有大纲或章节做快速设定体检。

初始化同人写作项目，建立目录、默认模式、reviewer 配置和初始 story bible。

把已确认信息回写到长期记忆、项目事实、反馈与 anti-pattern memory。

分析日志、审查与反馈，提出对 skills、阈值、记忆规则的优化补丁。

生成节点式剧情设计，明确每个节点的功能、冲突、情绪转向与代价。

生成三层大纲并支持局部 diff 调整。

Use when evaluating generative models — choosing metrics (FID, IS, KID, sFID, FDD, FVD, PRDC, LPIPS, SSIM, AuthPct, Vendi), setting up online or offline evaluation, feature extractor selection, distributed computation, memory management during sampling.

Use when optimizing PyTorch training speed or memory on CUDA GPUs — global flags, torch.compile, fused optimizers, mixed precision, gradient checkpointing, kernel fusion, memory layout, or latent-space training. Applies to any PyTorch training workload.

Use when curating, uploading, or managing HuggingFace datasets for ML training, including offline caching, preflight verification, and data directory conventions.

Use when structuring ML experiment configs with Hydra, adding new config groups, or debugging config resolution. Applies to any project using Hydra for hyperparameter management.

Use when designing ablation studies to compare model components, loss functions, or architectural choices. Covers synthetic data experiments, variant loops, production metrics, and W&B grouping.

Use when integrating W&B experiment tracking into ML training pipelines, including logging strategy, run configuration, and online/offline mode management.

Use when streaming large datasets from tar shards with WebDataset, replacing file-based DataLoaders, or precomputing encoder latents into shards.

Manage ts1v23 Slurm GPU availability on Iridis by checking active allocations, deciding whether remaining walltime is sufficient for training/debug, and submitting renewal jobs with gpu-jupyter.sh (H100), gpu-jupyter2.sh (A100), gpu-jupyter_ecs_a100.sh (swarm_a100 A100), gpu-jupyter3.sh (quad_h200 H200), or gpu-jupyter_dualh200.sh (dual_h200 H200). Use when the user asks to check current GPU resources, renew cards due to low remaining time, or run debug commands on allocated compute nodes.