ACE-Step
Docker 应用程序 from SpaceInvaderOne's Repository
概述
ACE-Step 1.5 - AI Music Generation. Generate full songs with vocals, instrumentals, and lyrics using a Diffusion Transformer. Supports text-to-music, remixing, cover generation, and LoRA fine-tuning. Requires NVIDIA GPU with CUDA support.
FIRST RUN: Models (~10GB) will be downloaded automatically on first start. This may take several minutes depending on your internet speed. Subsequent starts are instant.
SETTINGS GUIDE:
DiT Model - The core music generation model.
- turbo (default): Fast generation in 8 steps. Best for most users.
- turbo-rl: Turbo with reinforcement learning refinement.
- sft: Higher quality, 50 steps (slower).
- base: 50 steps with all features (extract, lego, complete).
Language Model - Controls lyrics understanding and chain-of-thought reasoning.
- 1.7B (default): Best balance of quality and VRAM. Recommended for 12-16GB GPUs.
- 0.6B: For GPUs with less than 12GB VRAM.
- 4B: Highest quality lyrics understanding. Requires 24GB+ VRAM.
Enable LLM - Whether to load the language model.
- auto (default): Detects based on your GPU VRAM.
- false: DiT-only mode. Faster startup, uses less VRAM, but disables thinking/sample features.
- true: Force enable.
LM Backend - Engine for the language model.
- pt (default): PyTorch native. Works on all GPUs including RTX 50-series.
- vllm: Faster inference but may crash on RTX 50-series (Blackwell) GPUs.
CPU Offloading - Moves models between GPU and CPU to save VRAM.
- auto (default): Offloads if GPU has less than 20GB VRAM.
- false: Keep all models on GPU. Faster generation but uses ~12GB VRAM at idle.
- true: Always offload. Slower but frees VRAM for other containers.
UI Language - Web interface language: English, Chinese, or Japanese.
要求
运行时参数
- 网络用户界面
http://[IP]:[PORT:7860]/- 网络
bridge- 外壳
bash- 特权
- false
- 额外参数
--gpus all --user root
模板配置
Gradio Web UI and REST API port
- 目标
- 7860
- 默认值
- 7860
- 价值
- 7860
AI model files (~10GB, auto-downloaded on first run)
- 目标
- /app/checkpoints
- 默认值
- /mnt/user/appdata/ace-step/checkpoints
- 价值
- /mnt/user/appdata/ace-step/checkpoints
Generated music files output directory
- 目标
- /app/gradio_outputs
- 默认值
- /mnt/user/appdata/ace-step/output
- 价值
- /mnt/user/appdata/ace-step/output
Diffusion model variant. Turbo=8 steps (fast), SFT=50 steps (quality), Base=50 steps (all features including extract/lego/complete).
- 目标
- ACESTEP_CONFIG_PATH
- 默认值
- acestep-v15-turbo|acestep-v15-turbo-rl|acestep-v15-sft|acestep-v15-base
- 价值
- acestep-v15-turbo
Chain-of-thought LM size. 1.7B recommended for 16GB VRAM. 4B needs 24GB+. 0.6B for low VRAM.
- 目标
- ACESTEP_LM_MODEL_PATH
- 默认值
- acestep-5Hz-lm-1.7B|acestep-5Hz-lm-0.6B|acestep-5Hz-lm-4B
- 价值
- acestep-5Hz-lm-1.7B
Auto detects based on GPU VRAM. Set false for DiT-only mode (faster, no thinking/sample features).
- 目标
- ACESTEP_INIT_LLM
- 默认值
- auto|true|false
- 价值
- auto
pt (PyTorch native) is recommended for RTX 50-series. vllm (nano-vllm) is faster but may segfault on Blackwell GPUs.
- 目标
- ACESTEP_LM_BACKEND
- 默认值
- pt|vllm
- 价值
- pt
Web interface language
- 目标
- LANGUAGE
- 默认值
- en|zh|ja
- 价值
- en
auto = ACE-Step decides based on VRAM (offloads below 20GB). false = keep all models on GPU (faster, needs ~12GB idle VRAM). true = offload models to CPU between steps (slower, saves VRAM for shared GPU use).
- 目标
- ACESTEP_OFFLOAD_CPU
- 默认值
- auto|false|true
- 价值
- auto
Internal port for Gradio server (should match the port mapping above)
- 目标
- PORT
- 默认值
- 7860
- 价值
- 7860
Default generation batch size (1-8). Leave empty for auto (min(2, GPU max)).
- 目标
- ACESTEP_BATCH_SIZE
Which GPU(s) to use. 0 = first GPU.
- 目标
- CUDA_VISIBLE_DEVICES
- 默认值
- 0
- 价值
- 0
Model download source. Auto tries HuggingFace first, falls back to ModelScope.
- 目标
- ACESTEP_DOWNLOAD_SOURCE
- 默认值
- auto|huggingface|modelscope
- 价值
- auto
下载统计数据
详细信息
spaceinvaderone/ace-step:latest在Unraid 上运行 ACE-Step 。
ACE-Step 已被列入Unraid OS 的社区应用程序。探索Unraid ，构建灵活的家庭服务器、NAS 或家庭实验室。