Dockerfile.AMD

FROM ghcr.io/jordimas/ctranslate2:latest-ubuntu22.04-rocm7.2

COPY entrypoint.sh /

ENV DEBIAN_FRONTEND=noninteractive
ENV PYTHONUNBUFFERED=1

RUN apt-get update && apt-get install -y --no-install-recommends \
    ffmpeg python3 python3-pip curl gosu tzdata libomp5 libdw1 \
    && ln -s /usr/lib/x86_64-linux-gnu/libomp.so.5 /usr/lib/x86_64-linux-gnu/libomp.so \
    && ldconfig \
    && rm -rf /var/lib/apt/lists/*

# Install torch with ROCm 7.2 nightly wheels to match the ctranslate2 ROCm 7.2 base image.
# The previous rocm6.2 wheels caused an HSA ABI mismatch (SIGSEGV on container startup).
# Nightly is required as stable PyTorch wheels only go up to rocm6.2 currently.
RUN python3 -m pip install -U --no-cache-dir --pre \
    torch torchaudio --index-url https://download.pytorch.org/whl/nightly/rocm7.2

RUN python3 -m pip install -U --no-cache-dir \
    numpy \
    fastapi \
    requests \
    uvicorn \
    python-multipart \
    ffmpeg-python \
    watchdog \
    tqdm \
    more-itertools \
    transformers \
    tokenizers \
    huggingface-hub \
    av \
    scipy

RUN python3 -m pip install -U --no-cache-dir --no-deps faster-whisper
RUN python3 -m pip install -U --no-cache-dir --no-deps stable-ts-whisperless

WORKDIR /subgen

COPY launcher.py subgen.py language_code.py /subgen/

RUN mkdir -p /cache && chmod 777 /cache

# -------------------------------------------------------------------
# --- ENVIRONMENT VARIABLES (Baked in for automatic execution) ---
# -------------------------------------------------------------------
# 1. Standard caches
ENV XDG_CACHE_HOME=/cache \
    HF_HOME=/cache/huggingface \
    MPLCONFIGDIR=/cache/matplotlib

# 2. Fix the Library Paths (Points to CTranslate2 and ROCm 7.2)
ENV LD_LIBRARY_PATH="/opt/ctranslate2/lib:/opt/rocm-7.2.0/lib:/opt/rocm-7.2.0/lib64:${LD_LIBRARY_PATH}"

# 3. Preload the ROCm 7.2 HSA Runtime to ensure both torch and ctranslate2 share
#    the same HSA runtime (both are now built against ROCm 7.2).
ENV LD_PRELOAD="/opt/rocm-7.2.0/lib/libhsa-runtime64.so.1"

# 4. Fix GPU memory page faults (Page not present) during ctranslate2 inference on
#    RDNA3 GPUs (gfx1100/gfx1101). The default allocator causes invalid GPU memory
#    mappings; cub_caching uses a stable allocation path. See OpenNMT/CTranslate2#2021.
ENV CT2_CUDA_ALLOCATOR="cub_caching"
# -------------------------------------------------------------------

ENTRYPOINT ["/entrypoint.sh"]

CMD ["python3", "-u", "launcher.py"]