Dockerfile

# Download model from huggingface
FROM python:3.13-alpine3.22 AS hfdownload

# The hf model "id" and actual "filename" within the hf repo,
# because alot of the times a quantized repo will have many available options.
# By default, use Qwen from ggml-org.
ARG MODEL_ID=ggml-org/Qwen3-1.7B-GGUF
ARG MODEL_FN=Qwen3-1.7B-Q4_K_M.gguf

WORKDIR /models

RUN pip install huggingface_hub && hf download ${MODEL_ID} ${MODEL_FN} --local-dir /models

# Use the llama.cpp server and copy our stuff into it
FROM ghcr.io/ggml-org/llama.cpp:server AS server

ARG CREATED
ARG LLMABDA_VERSION
ARG REVISION

ARG MODEL_FN=Qwen3-1.7B-Q4_K_M.gguf

LABEL org.opencontainers.image.title="llmabda"
LABEL org.opencontainers.image.description="Run LLM's on aws lambda for cheap."
LABEL org.opencontainers.image.version=$LLMABDA_VERSION
LABEL org.opencontainers.image.authors="pogzyb@umich.edu"
LABEL org.opencontainers.image.url="https://github.com/pogzyb/llmabda"
LABEL org.opencontainers.image.source="https://github.com/pogzyb/llmabda"
LABEL org.opencontainers.image.documentation="https://github.com/pogzyb/llmabda"
LABEL org.opencontainers.image.created=$CREATED
LABEL org.opencontainers.image.revision=$REVISION
LABEL org.opencontainers.image.licenses="MIT"

COPY --from=public.ecr.aws/awsguru/aws-lambda-adapter:0.9.1 /lambda-adapter /opt/extensions/lambda-adapter

ENV AWS_LWA_PORT=8080
ENV AWS_LWA_ENABLE_COMPRESSION=true
ENV AWS_LWA_READINESS_CHECK_PATH=/health

COPY --from=hfdownload /models /models

COPY entrypoint.sh /usr/local/bin/entrypoint.sh

ENV MODEL_PATH=/models/${MODEL_FN}

ENTRYPOINT ["entrypoint.sh"]