Dockerfile.edge

FROM python:3.14-slim

ENV PYTHONDONTWRITEBYTECODE=1 \
    PYTHONUNBUFFERED=1 \
    HF_HOME=/cache/huggingface \
    HF_HUB_CACHE=/cache/huggingface/hub \
    TRANSFORMERS_CACHE=/cache/huggingface/hub \
    HF_HUB_DISABLE_XET=1

WORKDIR /app

COPY requirements-edge.txt .
RUN pip install --no-cache-dir --upgrade pip \
    && pip install --no-cache-dir -r requirements-edge.txt

ARG COMPRESSOR_MODEL=microsoft/llmlingua-2-bert-base-multilingual-cased-meetingbank
ARG COMPRESSOR_MODEL_REVISION=main
ARG COMPRESSOR_GIT_COMMIT=unknown
ARG COMPRESSOR_SOURCE_SHA256=unknown
ENV COMPRESSOR_MODEL=${COMPRESSOR_MODEL} \
    COMPRESSOR_GIT_COMMIT=${COMPRESSOR_GIT_COMMIT} \
    COMPRESSOR_SOURCE_SHA256=${COMPRESSOR_SOURCE_SHA256} \
    COMPRESSOR_DEVICE=cpu \
    COMPRESSOR_PRELOAD_SLOTS=

# The CPU edge needs tokenizer parity with the GPU service, but never downloads
# model weights.  AutoTokenizer resolves these cached files in offline mode.
RUN mkdir -p /app/tokenizer /cache/huggingface/hub \
    && python -c "from pathlib import Path; from huggingface_hub import snapshot_download; snapshot_download('${COMPRESSOR_MODEL}', revision='${COMPRESSOR_MODEL_REVISION}', local_dir='/app/tokenizer', allow_patterns=['config.json','tokenizer.json','tokenizer_config.json','special_tokens_map.json','added_tokens.json','vocab.txt']); Path('/app/model_revision.txt').write_text('${COMPRESSOR_MODEL_REVISION}', encoding='utf-8')"

ENV HF_HUB_OFFLINE=1 \
    TRANSFORMERS_OFFLINE=1 \
    USE_TORCH=0 \
    USE_TF=0 \
    USE_FLAX=0 \
    COMPRESSOR_TOKENIZER_PATH=/app/tokenizer \
    EDGE_PRELOAD_TOKENIZER=true

RUN adduser --disabled-password --gecos "" appuser \
    && chown -R appuser:appuser /app /cache

COPY --chown=appuser:appuser app ./app
COPY --chown=appuser:appuser data ./data

USER appuser

EXPOSE 8080

HEALTHCHECK --interval=30s --timeout=5s --start-period=10s --retries=3 \
    CMD python -c "import json, os, urllib.request; port = os.getenv('PORT', '8080'); json.load(urllib.request.urlopen(f'http://127.0.0.1:{port}/health', timeout=3))"

CMD ["sh", "-c", "exec uvicorn app.edge_main:app --host 0.0.0.0 --port ${PORT:-8080}"]