FROM nvidia/cuda:12.9.0-runtime-ubuntu24.04

# Avoid interactive prompts during package installation
ENV DEBIAN_FRONTEND=noninteractive

# Install Python and system dependencies
# Ubuntu 24.04 ships FFmpeg 6.1 (torchcodec requires FFmpeg 5+, <8)
RUN apt-get update && apt-get install -y --no-install-recommends \
    python3 \
    python3-pip \
    python3-dev \
    git \
    libsndfile1 \
    ffmpeg \
    && rm -rf /var/lib/apt/lists/*

# Set Python alias (Ubuntu 24.04 ships Python 3.12)
RUN ln -sf /usr/bin/python3 /usr/bin/python

# Allow pip to install packages system-wide in the container (PEP 668)
ENV PIP_BREAK_SYSTEM_PACKAGES=1

# Set working directory
WORKDIR /app

# Install PyTorch ecosystem (cu128 wheels for CUDA 12.8+/12.9 compat)
RUN pip install --no-cache-dir \
    torch==2.10.0 \
    torchaudio==2.10.0 \
    torchcodec==0.10.0 \
    --index-url https://download.pytorch.org/whl/cu128

# Install common requirements (torch already installed above, pip will skip it). Versions pinned to the set this
# submission was validated with.
RUN pip install --no-cache-dir \
    transformers==5.21.0 \
    evaluate==0.4.6 \
    datasets==5.1.0 \
    librosa==1.0.0 \
    jiwer==4.0.0 \
    num2words==0.5.14 \
    peft==0.21.2 \
    numpy==2.5.4

# Install NeMo toolkit with ASR extras. Fluid Speech needs NeMo 3.0: its batched TDT beam search (malsd_batch) was
# rewritten after 2.7.2, and the submitted decoding (4 beams) is scored with the 3.0 implementation.
RUN pip install --no-cache-dir \
    "nemo_toolkit[asr]==3.0.0"

# Install SALM dependencies (lhotse for audio loading, transformers for GenerationConfig)
RUN pip install --no-cache-dir \
    lhotse==1.33.0 \
    sentencepiece==0.2.2

# CUDA-accelerated TDT and RNN-T decoding
# 12.9.4 matches the cuda-bindings torch 2.10 (cu128) requires.
RUN pip install --no-cache-dir \
    "cuda-python==12.9.4"

# Force soundfile backend for datasets audio decoding (avoids torchcodec/FFmpeg issues)
ENV HF_AUDIO_DECODER_BACKEND=soundfile

RUN pip install --no-cache-dir "kaldialign==0.12.0" \
    && python -c "from kaldialign import batch_error_rate; print('kaldialign OK')"

# Copy the full repository
COPY . /app

# Default entrypoint
ENTRYPOINT ["bash"]

# Keep-alive CMD so the Space runtime stays healthy. HF Jobs and `docker run`
# override this with their own command (e.g. `run_parakeet.sh`).
EXPOSE 7860
CMD ["-c", "python3 -m http.server 7860"]