v1.6.1 Beta51 Release

v1.6.1 Beta51 Release Latest

This is an experimental project currently in beta, and there are still tons of bugs.
这是一个实验性的项目，目前仍处于 beta 阶段，依然有成吨的问题。
Please check back often for updates.
请常回来看看有没有更新。

Highlight / 更新亮点

• Add experimental SpargeAttn2 runtime with a dedicated Python 3.11 environment, prebuilt wheel installation flow, and
launcher integration
• 新增实验性 SpargeAttn2 运行时：提供独立的 Python 3.11 隔离环境及预编译 Wheel 安装链路；现已深度集成至启动器，支持自动检测、安装与版本管理。

• Add AdamW8bitKahan with optional Kahan buffer offload and tensor-shape-based optimizer state offload controls
• 新增 AdamW8bitKahan 优化器：引入基于张量形状（Tensor Shape）的优化器状态卸载策略。

• Improve torch.compile and attention backend stability on newer GPU environments, including Triton prerequisite
provisioning and safer backend guard / fallback behavior
• 优化了 torch.compile 与注意力后端的兼容性: 完善了 Triton 前置依赖的部署流程，并强化了后端自动回退机制。

• Rework runtime uninstall into dependency-only cleanup so local Python skeletons, bootstrap files, and initialized
runtime structure are preserved for reinstall
• 优化运行时卸载逻辑：卸载时仅清理训练相关的重度依赖，保留本地 Python 骨架、引导文件及初始化的目录结构，极大提升了后续重装的速度。

• Improve launcher runtime cleanup / migration coverage and dependency cache plumbing for experimental runtimes such
as SpargeAttn2
• 增强实验运行时的整理与缓存链路：针对 SpargeAttn2 等实验性模块，完善了清理、迁移与依赖缓存链路，解决了管理流程中可能出现的组件漏检问题。

• Add selectable disk cache formats for both latent and text encoder outputs, defaulting to safetensors while
preserving optional legacy npz compatibility
• 新增 latent 与文本编码器输出缓存格式选项：默认格式切换为 safetensors，以提供更快的加载速度与安全性；同时保留对旧版 .npz 格式的向下兼容，确保老项目平滑过渡。

• Add configurable latent cache disk precision controls, allowing users to choose auto / fp16 / bf16 / fp32 based on
space and compatibility needs
• 新增 latent 磁盘缓存精度控制：支持 auto / fp16 / bf16 / fp32 多档调节，可根据存储空间压力及计算精度需求，灵活配置缓存深度。

• Improve disk cache reuse with catalog indexing and prewarmed source-stat checks for both latent and text caches,
reducing repeated rescans on subsequent runs
• 优化缓存复用链路：引入 Catalog 索引目录与预热式（Prewarmed）源文件状态检查，有效减少后续任务启动时的重复扫盘开销，大幅提升缓存命中检测的速度。

• Improve safetensors shard cache management by excluding catalog index files from manifest scans and making repeated
cache hits more stable
• 优化 safetensors 分片缓存管理：在清单扫描（Manifest Scan）中自动剔除索引文件，避免干扰缓存匹配；进一步加固了分片缓存的逻辑，使高频重复调用下的缓存命中更加稳健。