Shrink a 9 GB Serving Image Without Losing Functionality

A model-serving image is 9.1 GB. Its Dockerfile is:

FROM nvidia/cuda:12.1.0-devel-ubuntu22.04
RUN apt-get update && apt-get install -y build-essential git wget
COPY . /app
WORKDIR /app
RUN pip install -r requirements.txt
RUN python setup.py build_ext --inplace   # compiles one custom CUDA kernel
RUN rm -rf /root/.cache/pip
CMD ["python", "serve.py"]

1. Identify at least three concrete contributors to the image's size or layer bloat in this Dockerfile, referencing specific lines.
2. Rewrite the build as a multi-stage Dockerfile that keeps the custom CUDA kernel compilation but minimizes the final image.
3. Estimate, directionally, how much smaller the result would be and explain where the savings come from.