ZenTrainer 0.4.1

The container image builds again (0.4.0's didn't), so ghcr.io/nynxz/zentrainer:latest and the RunPod template get 0.4.x. The README credits the start of ZenTrainer's AMD support.

ZenTrainer 0.4.0

ZenTrainer 0.4.0

Train LoRAs inside ComfyUI: image and video models, live previews, on NVIDIA or AMD.

Models

• Ten models trained end to end: MOSS Image, Krea 2, Fibo 2.0, FLUX 3 Image, FLUX.2 Klein 9B, Qwen Image 2.1, Z-Image, Anima, SDXL and MiniMax H3 (video). The pickers list these as models. Anything else ComfyUI knows shows under Other models, marked untested.
• New this release:
  • Anima (Circlestone Labs), with a recipe on its makers' settings.
  • FLUX.2 Klein.
  • SDXL and its finetunes: Pony, Illustrious and others train as they are.
• Your own files under their model: a Pony or Illustrious checkpoint under SDXL, or a MOSS finetune under MOSS, tagged yours.

Recipes

• Recipes are data: one JSON file each in engine/catalog/recipes/, read when the app starts. No rebuild is needed to add or change one, and every recipe is checked against the run settings in tests.

Privacy

• A trained LoRA's metadata says how it was trained, not where. It no longer carries the dataset folder's name, your LoRA folders, or the dataset's captions.

Fixes

• Krea 2's character recipe ran out of VRAM at its switch from 768 to 1024 px.
• Every Anima run failed at step 1, and its LoRAs also trained the AdaLN modulation layers.
• fp8, nvfp4 and mxfp8 layers failed in the int8 backward under gradient checkpointing.
• A run started with no sample prompts trained without previews. It now gets two plain ones.
• Compile blocks is off unless you turn it on. It costs minutes of compiling on ROCm for a few percent.

AMD

• A faster int8 backward on RDNA3, from a tile set swept by @aksugat on an RX 7800 XT.
• Thanks to @0xDELUXA, whose ROCm PyTorch setup and whole-card VRAM on Windows (#1) ZenTrainer's AMD support is built on.

Everything else, in detail: CHANGELOG.md.