Skip to content

Using Your Adapter.md

Latest commit

 

History

History
148 lines (103 loc) · 5.74 KB

Using Your Adapter.md

File metadata and controls

148 lines (103 loc) · 5.74 KB

Using Your Adapter

After training completes, Side-Step saves your adapter weights to the output directory. This page explains the output layout, how to load adapters in ACE-Step's Gradio UI, and important considerations for inference.

Output Directory Layout

LoRA Training

output/my_lora/
├── final/
│   ├── adapter_config.json         # PEFT adapter configuration
│   └── adapter_model.safetensors   # Trained LoRA weights
├── checkpoints/
│   └── epoch_10/                   # Saved every N epochs (--save-every)
│       ├── adapter_config.json
│       ├── adapter_model.safetensors
│       └── training_state.pt       # Optimizer + scheduler state for resume
└── runs/
    └── [TensorBoard event files]   # Training metrics

LoKR Training

output/my_lokr/
├── final/
│   └── lokr_weights.safetensors    # Trained LoKR weights (config in metadata)
├── checkpoints/
│   └── epoch_10/
│       ├── lokr_weights.safetensors
│       └── training_state.pt
└── runs/
    └── [TensorBoard event files]

Key differences:

  • LoRA saves two files: adapter_config.json (PEFT configuration) + adapter_model.safetensors (weights).
  • LoKR saves one file: lokr_weights.safetensors (weights + configuration stored in safetensors metadata).
  • Checkpoints include training_state.pt for resuming training. The adapter files in each checkpoint are also inference-ready.
  • The final/ directory contains only the adapter weights (no training state) -- this is what you point inference tools at.

Loading LoRA in ACE-Step Gradio

ACE-Step's Gradio UI has a built-in LoRA loading section. Here is how to use it:

  1. Start ACE-Step's Gradio UI as you normally would.
  2. In the Service Configuration section, find the LoRA Adapter panel.
  3. In the LoRA Path field, enter the path to your adapter directory: 
    /path/to/Side-Step/output/my_lora/final
    Point at the directory, not at a specific file. ACE-Step expects to find adapter_config.json inside it.
  4. Click Load LoRA. You should see a success message.
  5. Toggle Use LoRA to enable the adapter during generation.
  6. Adjust the LoRA Scale slider to control how strongly the adapter affects output (1.0 = full strength, lower values blend with the base model).
  7. Generate audio as usual.

To switch back to the base model, click Unload or toggle Use LoRA off.

Requirements

  • Quantization must be disabled. LoRA loading is not supported when INT8 quantization is active. If you see an error about quantization, re-initialize the ACE-Step service with quantization set to None.
  • The adapter was trained with PEFT, so ACE-Step's standard LoRA loading path works directly.

Using Checkpoints for Inference

Every checkpoint directory (checkpoints/epoch_N/) also contains inference-ready adapter files. You can point the Gradio LoRA path at any checkpoint to test intermediate training results:

/path/to/Side-Step/output/my_lora/checkpoints/epoch_50

LoKR Limitation

ACE-Step's Gradio UI currently only supports PEFT LoRA adapters. The loading code checks for adapter_config.json, which LoKR does not produce (LoKR uses LyCORIS format with lokr_weights.safetensors).

LoKR adapters cannot be loaded in the standard ACE-Step Gradio UI at this time. If you trained a LoKR adapter and want to use it for inference, you will need to write custom inference code using the LyCORIS library, or wait for LoKR support to be added to ACE-Step.

If you are unsure which adapter type to use, LoRA is the safe choice for both training and inference compatibility.

Shift and Inference Steps

When generating audio with your trained adapter, use the correct shift and inference step values for your model variant:

Model Variant Shift Inference Steps
Turbo 3.0 8
Base 1.0 50
SFT 1.0 50

These values were saved as metadata during training (via --shift and --num-inference-steps) but do not affect the training loop itself. They tell you what to use at inference time. See [[Shift and Timestep Sampling]] for the full explanation.

Using the wrong shift value at inference will produce degraded audio quality, even if the adapter was trained correctly.

Viewing TensorBoard Logs

Side-Step writes TensorBoard logs to the runs/ subdirectory of your output directory (or to --log-dir if you specified one). To view them:

tensorboard --logdir ./output/my_lora/runs

Then open http://localhost:6006 in your browser. Key metrics:

  • loss -- Training loss over time. Should decrease and stabilize.
  • lr -- Learning rate schedule. Verify warmup and decay look correct.
  • grad_norm -- Gradient magnitudes. Spikes may indicate training instability.

Resuming Training

To continue training from a checkpoint:

uv run train.py fixed \
    --checkpoint-dir ./checkpoints \
    --model-variant turbo \
    --dataset-dir ./my_tensors \
    --output-dir ./output/my_lora \
    --resume-from ./output/my_lora/checkpoints/epoch_50 \
    --epochs 200

This restores the adapter weights, optimizer state, and learning rate scheduler from the checkpoint. Training continues from where it left off.

You can also point --resume-from at a file inside the checkpoint directory (e.g., training_state.pt) -- Side-Step will automatically use the parent directory.

See Also

  • [[End-to-End Tutorial]] -- Full walkthrough from raw audio to generation
  • [[Shift and Timestep Sampling]] -- Why shift matters for inference
  • [[Training Guide]] -- Training options and hyperparameters
  • [[VRAM Optimization Guide]] -- GPU memory management