Adapters

Place your LoRA adapters in data/adapters/: a .safetensors file, or a folder holding one next to its adapter_config.json. The server lists them at startup with the halves of the backbone each one changes, AR (the score and the codes), NAR (the acoustics) or both, and the WebUI stacks any number of them, each with its own strength.

Trainer	Layout	Alpha source
AI Toolkit, ComfyUI	single .safetensors	per module .alpha, else the rank
Native YuE2 trainers	single .safetensors	__metadata__ alpha, else the rank
Folder trainers	.safetensors + adapter_config.json	lora_alpha or alpha
Sliders	single .safetensors	per module .alpha

ComfyUI and AI Toolkit files carry text_encoders.* for the AR half and diffusion_model.* for the NAR half, with fused qkv_proj and gate_up_proj. Native files name the projections layers.N.self_attn.*, layers.N.nar_self_attn.* and their MLPs, and may replace or adapt vae2llm and llm2vae. A file holding anything else (LoKr, DoRA, an extra conditioning branch) is skipped with the name of the tensor at fault.

From the project root, start the local server with Task:

task start

Task creates data/adapters/ when needed and mounts it read-only at /adapters. The image’s default command already sets --adapters /adapters. Restart the server after adding files so it can discover them at startup. Adapter files stay outside Git and the Docker image. See manual Docker mounts.

These are inference loading formats; this project does not provide training workflows. Follow the adapter’s own terms as well as the base model’s license.