👷 put python installs in separate stage (#46)

This puts all the final `pip install` commands and site-packages
overrides into a `python-installations` stage, then the python
environment from there is copied into the `server-release` stage.

This should mean that the final image will not contain any
intermediate python packages.

Signed-off-by: Joe Runde <[email protected]>

Author: joerunde

.github/actions/free-up-disk-space/action.yml

@@ -12,13 +12,17 @@ runs:
         df -h
         echo "Removing non-essential tools and libraries ..."
         sudo rm -rf /opt/ghc
+        sudo rm -rf /usr/local/.ghcup
         sudo rm -rf /usr/share/dotnet
         # sudo rm -rf /usr/local/share/boost
         echo "Deleting libraries for Android (12G), CodeQL (5.3G), PowerShell (1.3G), Swift (1.7G) ..."
         sudo rm -rf /usr/local/lib/android
         sudo rm -rf "${AGENT_TOOLSDIRECTORY}/CodeQL"
         sudo rm -rf /usr/local/share/powershell
         sudo rm -rf /usr/share/swift
+        # ref: https://github.com/jlumbroso/free-disk-space/blob/main/action.yml
+        echo "Deleting some larger apt packages:"
+        sudo apt-get remove -y azure-cli google-chrome-stable firefox powershell mono-devel libgl1-mesa-dri --fix-missing || echo "::warning::The command [sudo apt-get remove -y azure-cli google-chrome-stable firefox powershell mono-devel libgl1-mesa-dri --fix-missing] failed to complete successfully. Proceeding..."
         echo "Disk usage after cleanup:"
         df -h
 

Dockerfile

@@ -285,19 +285,16 @@ FROM base as auto-gptq-cache
 COPY --from=auto-gptq-installer /usr/src/auto-gptq-wheel /usr/src/auto-gptq-wheel
 
 
-## Final Inference Server image ################################################
-FROM cuda-runtime as server-release
-ARG PYTHON_VERSION
-ARG SITE_PACKAGES=/opt/tgis/lib/python${PYTHON_VERSION}/site-packages
+## Full set of python installations for server release #########################
 
-# Install C++ compiler (required at runtime when PT2_COMPILE is enabled)
-RUN dnf install -y gcc-c++ git && dnf clean all \
-    && useradd -u 2000 tgis -m -g 0
+FROM cuda-runtime as python-installations
 
-SHELL ["/bin/bash", "-c"]
+ARG PYTHON_VERSION
+ARG SITE_PACKAGES=/opt/tgis/lib/python${PYTHON_VERSION}/site-packages
 
 COPY --from=build /opt/tgis /opt/tgis
 
+# `pip` is installed in the venv here
 ENV PATH=/opt/tgis/bin:$PATH
 
 # Install flash attention v2 from the cache build
@@ -315,13 +312,30 @@ RUN --mount=type=bind,from=auto-gptq-cache,src=/usr/src/auto-gptq-wheel,target=/
     pip install /usr/src/auto-gptq-wheel/*.whl --no-cache-dir
 
 # Install server
+# git is required to pull the fms-extras dependency
+RUN dnf install -y git && dnf clean all
 COPY proto proto
 COPY server server
 RUN cd server && make gen-server && pip install ".[accelerate, ibm-fms, onnx-gpu, quantize]" --no-cache-dir
 
 # Patch codegen model changes into transformers 4.35
 RUN cp server/transformers_patch/modeling_codegen.py ${SITE_PACKAGES}/transformers/models/codegen/modeling_codegen.py
 
+
+## Final Inference Server image ################################################
+FROM cuda-runtime as server-release
+ARG PYTHON_VERSION
+ARG SITE_PACKAGES=/opt/tgis/lib/python${PYTHON_VERSION}/site-packages
+
+# Install C++ compiler (required at runtime when PT2_COMPILE is enabled)
+RUN dnf install -y gcc-c++ && dnf clean all \
+    && useradd -u 2000 tgis -m -g 0
+
+# Copy in the full python environment
+COPY --from=python-installations /opt/tgis /opt/tgis
+
+ENV PATH=/opt/tgis/bin:$PATH
+
 # Print a list of all installed packages and versions
 RUN pip list -v --disable-pip-version-check --no-python-version-warning