Dockerfile.gpu

# ── HF GPU Space Dockerfile (CUDA / llama.cpp on GPU) ────────────────────────
# Optimized: lightweight runtime image + PREBUILT CUDA wheel for llama-cpp-python
# (no source compile, so no nvcc/toolkit needed and no OOM during build).
#
# For local CPU dev, keep using Dockerfile.local instead.

# Lightweight runtime image. We don't compile anything, so we don't need the
# heavier `devel` image (which only existed to provide nvcc). The prebuilt
# llama-cpp-python wheel below already contains compiled CUDA code; this image
# just provides the CUDA 12.4 runtime libraries it links against.
FROM nvidia/cuda:12.4.1-runtime-ubuntu22.04

# System deps. We use a prebuilt CUDA wheel for llama-cpp-python (no compile),
# but a couple of small pure-C dependencies (e.g. cdifflib, pulled in by
# nemo_text_processing) ship no wheel and must compile from source, so we keep
# build-essential + python3.11-dev (gcc + Python headers) for those. This does
# NOT recompile llama.cpp — that's still the prebuilt wheel below.
RUN apt-get update && apt-get install -y --no-install-recommends \
        python3.11 python3.11-dev python3-pip \
        ffmpeg \
        build-essential \
    && rm -rf /var/lib/apt/lists/* \
    && ln -sf /usr/bin/python3.11 /usr/bin/python

# ── HF permission convention: run as a non-root user with ID 1000 ────────────
RUN useradd -m -u 1000 user
USER user
ENV HOME=/home/user \
    PATH=/home/user/.local/bin:$PATH
WORKDIR $HOME/app

# Keep all model/cache downloads inside the user's home (writable on a Space).
ENV HF_HOME=$HOME/.cache/huggingface \
    XDG_CACHE_HOME=$HOME/.cache

# Install Python deps. Use a PREBUILT CUDA 12.4 wheel for llama-cpp-python:
#  - --extra-index-url points pip at abetlen's prebuilt CUDA wheel index
#  - --only-binary=:all: for llama-cpp-python forces a prebuilt wheel; if no
#    matching wheel exists pip ERRORS LOUDLY instead of silently falling back
#    to a (slow, OOM-prone) source compile. `python -m pip` binds to py3.11.
COPY --chown=user requirements-gpu.txt .
RUN python -m pip install --no-cache-dir --upgrade pip && \
    python -m pip install --no-cache-dir --only-binary=:all: \
        llama-cpp-python==0.3.4 \
        --extra-index-url https://abetlen.github.io/llama-cpp-python/whl/cu124 && \
    python -m pip install --no-cache-dir -r requirements-gpu.txt

# Copy the project (owned by the runtime user).
COPY --chown=user . .

# Pre-download Kokoro + Wav2Vec2 into the build (lazy models pulled at startup).
RUN python scripts/download_models.py

# ── Runtime config ───────────────────────────────────────────────────────────
# Offload all LLM layers to the GPU. Gradio binds 0.0.0.0:7860 (HF default).
ENV LLM_GPU_LAYERS=999 \
    GRADIO_SERVER_NAME=0.0.0.0 \
    GRADIO_SERVER_PORT=7860

EXPOSE 7860
CMD ["python", "app.py"]