Dockerfile

# Multi-stage Docker build following our proven HuggingFace pattern
# This approach uses official moshi-server instead of custom implementation

# Build argument for CUDA compute capability (T4 = 75)
ARG CUDA_COMPUTE_CAP=75

FROM nvidia/cuda:12.2.0-devel-ubuntu22.04 AS base-builder

# Install system dependencies for moshi-server
RUN apt-get update && apt-get install -y \
    curl \
    build-essential \
    pkg-config \
    libssl-dev \
    libsndfile1-dev \
    libasound2-dev \
    wget \
    ca-certificates \
    git \
    cmake \
    libprotobuf-dev \
    protobuf-compiler \
    python3-dev \
    && rm -rf /var/lib/apt/lists/*

# Install Rust
ENV RUSTUP_HOME=/usr/local/rustup
ENV CARGO_HOME=/usr/local/cargo  
ENV PATH=/usr/local/cargo/bin:$PATH
ENV RUST_VERSION=stable

RUN curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh -s -- -y --default-toolchain $RUST_VERSION
RUN chmod -R a+w $RUSTUP_HOME $CARGO_HOME

WORKDIR /app

# Official moshi-server installation stage
FROM base-builder AS moshi-server-builder
ARG CUDA_COMPUTE_CAP
ENV CUDA_COMPUTE_CAP=${CUDA_COMPUTE_CAP}

# Set build optimizations to prevent hangs
ENV CARGO_NET_RETRY=10
ENV CARGO_HTTP_TIMEOUT=300
ENV CARGO_HTTP_LOW_SPEED_LIMIT=10
ENV CARGO_BUILD_JOBS=4

# Install official moshi-server with CUDA support
# Note: Official moshi-server only has 'cuda' feature, no architecture-specific variants
RUN cargo install --features cuda moshi-server

# Model download stage - pre-load models into image
FROM base-builder AS model-downloader

# Install Python and pip for model downloading
RUN apt-get update && apt-get install -y python3 python3-pip && rm -rf /var/lib/apt/lists/*

# Install huggingface-hub for model downloading
RUN pip3 install --no-cache-dir huggingface-hub

# Set working directory for models first
WORKDIR /app/models

# Create models directory for 1B multilingual model (T4 GPU compatible)  
RUN mkdir -p kyutai/stt-1b-en_fr-candle

# Create download script for 1B multilingual model
RUN echo 'from huggingface_hub import hf_hub_download\n\
import os\n\
import subprocess\n\
\n\
os.makedirs("kyutai/stt-1b-en_fr-candle", exist_ok=True)\n\
print("📥 Downloading 1B multilingual STT model for official moshi-server...")\n\
\n\
print("⬇️ Downloading model.safetensors...")\n\
hf_hub_download(\n\
    repo_id="kyutai/stt-1b-en_fr-candle",\n\
    filename="model.safetensors",\n\
    local_dir="kyutai/stt-1b-en_fr-candle",\n\
    local_dir_use_symlinks=False\n\
)\n\
\n\
print("⬇️ Downloading tokenizer (8000 vocab)...")\n\
hf_hub_download(\n\
    repo_id="kyutai/stt-1b-en_fr-candle",\n\
    filename="tokenizer_en_fr_audio_8000.model",\n\
    local_dir="kyutai/stt-1b-en_fr-candle",\n\
    local_dir_use_symlinks=False\n\
)\n\
\n\
print("⬇️ Downloading Mimi audio tokenizer...")\n\
hf_hub_download(\n\
    repo_id="kyutai/stt-1b-en_fr-candle",\n\
    filename="mimi-pytorch-e351c8d8@125.safetensors",\n\
    local_dir="kyutai/stt-1b-en_fr-candle",\n\
    local_dir_use_symlinks=False\n\
)\n\
\n\
print("✅ All models downloaded successfully!")\n\
result = subprocess.run(["du", "-sh", "kyutai/"], capture_output=True, text=True)\n\
print("📊 Model files:", result.stdout)\n\
' > download_models.py

# Download models during build time
RUN python3 download_models.py

# Runtime stage
FROM nvidia/cuda:12.2.0-runtime-ubuntu22.04 AS runtime

# Install Python and runtime dependencies including audio libraries for moshi-server
RUN apt-get update && apt-get install -y \
    python3 \
    python3-pip \
    wget \
    curl \
    git \
    sudo \
    libopus0 \
    libsndfile1 \
    libasound2 \
    && rm -rf /var/lib/apt/lists/*

# Install Python dependencies for web interface
RUN pip3 install gradio fastapi uvicorn websockets python-multipart msgpack

# Set working directory
WORKDIR /app

# Copy official moshi-server binary from builder
COPY --from=moshi-server-builder /usr/local/cargo/bin/moshi-server /usr/local/bin/

# Copy configuration files
COPY configs/ ./configs/

# Copy pre-downloaded models from model-downloader stage
COPY --from=model-downloader /app/models/ ./models/

# Copy Python app and requirements
COPY app.py .
COPY requirements.txt .

# Create user for HuggingFace Spaces
RUN useradd -m -u 1000 user
ENV HOME=/home/user \
    PATH=/home/user/.local/bin:$PATH \
    HF_HOME=/home/user/.cache/huggingface

# Create and set permissions for HuggingFace cache directory
RUN mkdir -p /home/user/.cache/huggingface && \
    chown -R user:user /home/user/.cache

# Set proper permissions and ownership
RUN chown -R user:user /app
USER user

# Expose port (HuggingFace Spaces requirement)
EXPOSE 7860

# Create startup script that handles port constraints
RUN echo '#!/bin/bash\n\
echo "===== Application Startup at $(date) ====="\n\
echo "🚀 Starting Kyutai STT Server Moshi v4 with official moshi-server..."\n\
echo "📁 Pre-loaded models:"\n\
ls -lah models/kyutai/stt-1b-en_fr-candle/ 2>/dev/null || echo "Model directory not found"\n\
echo "GPU Info:"\n\
nvidia-smi 2>/dev/null || echo "No GPU detected at runtime"\n\
echo "🌐 Official moshi-server endpoints:"\n\
echo "  - STT Streaming: /api/asr-streaming (MessagePack protocol)"\n\
echo "  - Health Check: /health"\n\
echo "Starting Python frontend and official moshi-server..."\n\
\n\
# Create required directories for moshi-server\n\
mkdir -p /app/static /app/logs\n\
\n\
# Check if port 7860 is available (HuggingFace requirement)\n\
if lsof -Pi :7860 -sTCP:LISTEN -t >/dev/null 2>&1; then\n\
    echo "⚠️ Port 7860 already in use, exiting gracefully (PID: $$)"\n\
    exit 0\n\
fi\n\
\n\
# Start Python web interface in background\n\
python3 app.py &\n\
\n\
# Start official moshi-server as main process\n\
# Note: moshi-server will bind to its default port, Python app proxies on 7860\n\
exec moshi-server worker --config configs/config-stt-en_fr-hf.toml\n\
' > /app/start.sh && chmod +x /app/start.sh

# Health check
HEALTHCHECK --interval=30s --timeout=10s --start-period=120s --retries=3 \
    CMD curl -f http://localhost:7860/health || exit 1

# Run the combined server
CMD ["/app/start.sh"]