File size: 4,131 Bytes
fccaf48
 
 
 
 
fd22b19
fccaf48
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
567a93e
fccaf48
 
 
 
 
 
 
 
 
63d7cb8
fccaf48
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
c0f2a13
fccaf48
 
 
c0f2a13
 
 
 
5e95353
 
 
fccaf48
 
 
 
 
 
 
 
 
625132a
 
6b3e616
a07b39d
6b3e616
4619f39
 
625132a
 
 
fccaf48
372f08e
 
c37e80e
 
 
 
fccaf48
 
5fe3c53
4619f39
fccaf48
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
FROM ubuntu:22.04 AS builder

ENV DEBIAN_FRONTEND=noninteractive

RUN apt-get update && apt-get install -y --no-install-recommends \
    build-essential cmake git ca-certificates libopenblas-dev pkg-config \
    && rm -rf /var/lib/apt/lists/*

WORKDIR /build

# Clone acestep.cpp and init submodules
RUN git clone --depth 1 https://github.com/ServeurpersoCom/acestep.cpp.git . \
    && git submodule update --init --depth 1

# Build CPU-only with BLAS (-j1 to avoid OOM on HF free tier)
RUN mkdir build && cd build \
    && cmake .. -DGGML_BLAS=ON \
    && cmake --build . --config Release -j1

# Stage built artifacts into a clean directory for COPY
RUN mkdir -p /artifacts \
    && cp /build/build/ace-server /artifacts/ \
    && cp /build/build/ace-lm /artifacts/ \
    && cp /build/build/ace-synth /artifacts/ \
    && (cp -a /build/build/lib*.so* /artifacts/ 2>/dev/null || true)

# ---------------------------------------------------------------------------
# Runtime image
# ---------------------------------------------------------------------------
FROM ubuntu:22.04

ENV DEBIAN_FRONTEND=noninteractive

RUN apt-get update && apt-get install -y --no-install-recommends \
    libopenblas0 libgomp1 ca-certificates git libsndfile1 \
    python3 python3-pip curl \
    && rm -rf /var/lib/apt/lists/*

WORKDIR /app

# Copy built binaries and shared libraries
COPY --from=builder /artifacts/ /app/

# Ensure GGML backend .so files are findable at runtime
ENV LD_LIBRARY_PATH=/app:${LD_LIBRARY_PATH}

# Make binaries executable
RUN chmod +x /app/ace-server /app/ace-lm /app/ace-synth

# Create model and adapter directories
RUN mkdir -p /app/models /app/adapters /app/outputs

# Download GGUF models at build time (NOT via Git LFS)
# --fail ensures curl returns non-zero on HTTP errors (catches 404)

# XL DiT turbo Q4_K_M (~2.8GB) - best quality for no-LoRA inference
RUN curl -fL --retry 3 --retry-delay 5 -o /app/models/acestep-v15-xl-turbo-Q4_K_M.gguf \
    "https://huggingface.co/Serveurperso/ACE-Step-1.5-GGUF/resolve/main/acestep-v15-xl-turbo-Q4_K_M.gguf"

# Standard DiT turbo Q4_K_M (~1.1GB) - used when LoRA adapter is selected
RUN curl -fL --retry 3 --retry-delay 5 -o /app/models/acestep-v15-turbo-Q4_K_M.gguf \
    "https://huggingface.co/Serveurperso/ACE-Step-1.5-GGUF/resolve/main/acestep-v15-turbo-Q4_K_M.gguf"

# LM 1.7B Q8_0 (~1.7GB) - best speed/quality on CPU
RUN curl -fL --retry 3 --retry-delay 5 -o /app/models/acestep-5Hz-lm-1.7B-Q8_0.gguf \
    "https://huggingface.co/Serveurperso/ACE-Step-1.5-GGUF/resolve/main/acestep-5Hz-lm-1.7B-Q8_0.gguf"

# Text encoder Q8_0 (~0.75GB)
RUN curl -fL --retry 3 --retry-delay 5 -o /app/models/Qwen3-Embedding-0.6B-Q8_0.gguf \
    "https://huggingface.co/Serveurperso/ACE-Step-1.5-GGUF/resolve/main/Qwen3-Embedding-0.6B-Q8_0.gguf"

# VAE BF16 (~0.32GB) - always BF16, quality-critical
RUN curl -fL --retry 3 --retry-delay 5 -o /app/models/vae-BF16.gguf \
    "https://huggingface.co/Serveurperso/ACE-Step-1.5-GGUF/resolve/main/vae-BF16.gguf"

# Install Python deps for Gradio UI + training
RUN pip3 install --no-cache-dir --extra-index-url https://download.pytorch.org/whl/cpu \
    "gradio[mcp]>=6.0.0,<7.0.0" requests "torch>=2.6.0" safetensors \
    "transformers>=4.51.0,<4.58.0" peft>=0.18.0 \
    loguru torchaudio "diffusers==0.30.3" numpy tensorboard soundfile \
    einops vector_quantize_pytorch librosa mutagen demucs-infer \
    faster-whisper silero-vad

# Clone ACE-Step repo for training module
RUN git clone --depth 1 https://github.com/ace-step/ACE-Step-1.5 /app/ace-step-source

# Pre-download training checkpoints (avoids runtime download)
# Base repo has VAE + text encoder + standard turbo
RUN python3 -c "from huggingface_hub import snapshot_download; \
    snapshot_download('ACE-Step/Ace-Step1.5', local_dir='/app/checkpoints', \
    ignore_patterns=['*.md', '*.txt', '.gitattributes'])"

# Copy application files
COPY app.py /app/app.py
COPY train_engine.py /app/train_engine.py
COPY caption_fast.py /app/caption_fast.py
COPY start.sh /app/start.sh
RUN chmod +x /app/start.sh

# HF Spaces expects port 7860
EXPOSE 7860

CMD ["/app/start.sh"]