feat: add Docker support for NVIDIA RTX 5000 series (Blackwell / sm_120)

Dockerfile

@@ -0,0 +1,55 @@
+# VoxCPM2 – Docker image with CUDA 13.0 (native sm_120 / Blackwell support)
+# Pattern adapted from suite-redazione/tada-tts (proven on RTX 5070)
+# Requires: nvidia-container-toolkit on the host
+FROM nvidia/cuda:13.0.1-cudnn-devel-ubuntu22.04
+
+ENV DEBIAN_FRONTEND=noninteractive
+ENV PYTHONUNBUFFERED=1
+ENV TOKENIZERS_PARALLELISM=false
+ENV HF_HOME=/cache/huggingface
+ENV MODELSCOPE_CACHE=/cache/modelscope
+ENV TORCH_HOME=/cache/torch
+
+# ── System packages ───────────────────────────────────────────────────────────
+RUN apt-get update && apt-get install -y --no-install-recommends \
+    software-properties-common \
+    curl \
+    && add-apt-repository ppa:deadsnakes/ppa \
+    && apt-get update && apt-get install -y --no-install-recommends \
+    python3.11 \
+    python3.11-dev \
+    python3.11-venv \
+    build-essential \
+    git \
+    ffmpeg \
+    libsndfile1 \
+    libsndfile1-dev \
+    libgomp1 \
+    # cuBLAS 12 compat — some deps (funasr, ctranslate2) ship CUDA 12 binaries
+    libcublas-12-8 \
+    && rm -rf /var/lib/apt/lists/*
+
+# Virtualenv con Python 3.11 — isolato dal sistema
+RUN python3.11 -m venv /opt/venv
+ENV PATH="/opt/venv/bin:$PATH"
+
+# setuptools>=78 required for PEP 639 SPDX license strings
+RUN pip install --no-cache-dir --upgrade "pip>=24" "setuptools>=78" wheel
+
+# ── Application ───────────────────────────────────────────────────────────────
+WORKDIR /app
+COPY . .
+
+# Install voxcpm and all its deps (may pull CPU torch as transitive dep — OK)
+RUN pip install --no-cache-dir -e .
+
+# Force-reinstall the cu130 torch stack last so CUDA versions always win.
+# CUDA 13.0 wheels include Triton with native sm_120 (Blackwell) support.
+RUN pip install --no-cache-dir --force-reinstall \
+    --index-url https://download.pytorch.org/whl/cu130 \
+    torch torchaudio torchcodec
+
+# ── Runtime ───────────────────────────────────────────────────────────────────
+EXPOSE 8808
+
+CMD ["python", "app.py", "--port", "8808"]

docker-compose.yml

@@ -0,0 +1,54 @@
+services:
+  voxcpm:
+    build:
+      context: .
+      dockerfile: Dockerfile
+    image: voxcpm:latest
+    container_name: voxcpm-demo
+
+    # ── GPU access ─────────────────────────────────────────────────────────────
+    deploy:
+      resources:
+        reservations:
+          devices:
+            - driver: nvidia
+              count: all
+              capabilities: [gpu]
+
+    # ── Ports ──────────────────────────────────────────────────────────────────
+    ports:
+      - "8808:8808"
+
+    # ── Persistent volumes ─────────────────────────────────────────────────────
+    # Model weights (~4-6 GB) are cached here so they survive container restarts.
+    volumes:
+      - voxcpm-cache:/cache
+      # Optional: mount a local output directory
+      # - ./outputs:/app/outputs
+
+    # ── Environment ────────────────────────────────────────────────────────────
+    environment:
+      - NVIDIA_VISIBLE_DEVICES=all
+      - NVIDIA_DRIVER_CAPABILITIES=compute,utility
+      - HF_HOME=/cache/huggingface
+      - MODELSCOPE_CACHE=/cache/modelscope
+      - TORCH_HOME=/cache/torch
+      - TRITON_CACHE_DIR=/cache/triton
+      - TORCHINDUCTOR_CACHE_DIR=/cache/inductor
+      # Fallback: decommentare se Triton crasha ancora (eager mode, ~2x più lento)
+      # - TORCHDYNAMO_DISABLE=1
+      # Uncomment to use a specific model version (default: openbmb/VoxCPM2)
+      # - VOXCPM_MODEL_ID=openbmb/VoxCPM2
+
+    # ── Restart policy ─────────────────────────────────────────────────────────
+    restart: unless-stopped
+
+    # ── Custom entrypoint (supports VOXCPM_MODEL_ID env var) ──────────────────
+    command: >
+      python app.py
+      --port 8808
+      ${VOXCPM_MODEL_ID:+--model-id $VOXCPM_MODEL_ID}
+
+volumes:
+  voxcpm-cache:
+    driver: local