---
title: ACE-Step 1.5 jpdenpa full fine-tune
emoji: 🎼
colorFrom: purple
colorTo: indigo
sdk: gradio
sdk_version: 5.50.0
app_file: app.py
short_description: Full fine-tune of ACE-Step 1.5 turbo on Japanese songs
python_version: "3.12"
startup_duration_timeout: 1h
---

# ACE-Step 1.5 turbo — full fine-tune

Every decoder parameter of ACE-Step 1.5 turbo (1,575,458,880 of them, 65.8% of
the whole model) was fine-tuned on a small set of Japanese songs. This is a
full fine-tune, not a LoRA: there is no adapter to attach, the decoder weights
themselves were replaced.

Captions were trained with the tag `jpdenpa` in front, so keeping that tag in
the caption is what pulls the fine-tuned style in.

## Training

| | |
|---|---|
| Base | `ACE-Step/Ace-Step1.5`, turbo variant |
| Trainable | 1,575,458,880 params — the entire decoder |
| Optimizer | Adafactor, lr 2e-5, cosine, batch 1 × grad accum 4 |
| Precision | bf16 base, fp32 masters for the unfrozen weights |
| Hardware | 2 × RTX 5060 Ti 16 GB, decoder sharded across both |
| Checkpoint | epoch 40 |

Two GPUs were needed because fp32 masters plus fp32 grads for the decoder come
to ~19 GB, which does not fit on one 16 GB card. The decoder blocks are split
across both with accelerate's align-device hooks — sequential model
parallelism, not DDP.

Weights: [`swdq/acestep-v15-jpdenpa-ft`](https://huggingface.co/swdq/acestep-v15-jpdenpa-ft)

## Why the training loss is not the metric here

The flow-matching loss is `MSE(pred, x1 - x0)` where `x1` is fresh noise and
the timestep is redrawn every step, so consecutive epoch losses differ mostly
because they saw different draws. Measured over epochs 15–33 of this run, the
epoch-to-epoch jitter was 0.024 against a real per-epoch improvement of 0.0026
— nine times larger than the signal. Progress was tracked with a separate
harness that pins the noise and the timestep grid so every checkpoint sees
byte-identical inputs.