v1.3.5 Beta25 release

v1.3.5 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 an experimental SDXL low-VRAM optimization mode (≤6GB)
• 新增实验性的 SDXL 低显存优化模式（≤6GB）

• Add staged mixed-resolution training support for Anima
• 新增 Anima 的阶段分辨率混合训练支持

• Continue expanding surfaced Lulynx experimental core controls
• 继续扩展已开放的 Lulynx 实验核心控制项

• Improve preview runtime safety on experimental attention routes
• 改进实验性注意力路线下的预览安全性

• Improve SDXL CPU offload compatibility handling to avoid hard crashes on model variants that do not expose native
cpu_offload gradient checkpointing hooks.
• 改进 SDXL 的 CPU offload 兼容处理，避免某些模型变体因缺少原生 cpu_offload gradient checkpointing 接口而直接崩溃

• Continue refining Newbie support in shared runtimes
• 继续完善 Newbie 在共享运行时中的兼容性支持

• Improve CN / SageAttention / launcher robustness with better mirror fallback behavior
• 继续提升 CN / SageAttention / 启动器链路的稳健性，改进镜像回退逻辑

• Fix a T-LoRA training crash caused by recursive module
• 修复 T-LoRA 因递归模块引用导致的训练崩溃问题

• Improve Tag Editor offline behavior and preview stability
• 改进 Tag Editor 的离线行为与图片预览稳定性