Update CUDA build

.circleci/config.yml

@@ -98,7 +98,7 @@ workflows:
       - deploy:
           matrix:
             parameters:
-              docker-variant: ['', '-cuda', '-cuda-tf1', '-cuda-tf2', '-cuda-torch']
+              docker-variant: ['', '-cuda', '-cuda-tf2', '-cuda-torch']
           filters:
             branches:
               only: master

.github/workflows/docker-image.yml

@@ -4,11 +4,15 @@ on:
   push:
     branches: [ "master" ]
   workflow_dispatch: # run manually
+    inputs:
+      platforms:
+        description: "platform argument for docker buildx"
+        type: string
+        default: "linux/amd64,linux/arm/v7,linux/arm64/v8,linux/ppc64le"
 
 env:
-  # FIXME: linux/arm/v7 disabled as long as scikit-build/cmake-python-distributions#503 is unresolved
-  # PLATFORMS: linux/amd64,linux/arm/v7,linux/arm64/v8,linux/ppc64le
-  PLATFORMS: linux/amd64,linux/arm64/v8,linux/ppc64le
+  #PLATFORMS: linux/amd64,linux/arm/v7,linux/arm64/v8,linux/ppc64le
+  PLATFORMS: ${{ github.event.inputs.platforms || 'linux/amd64' }}
 
 jobs:
 
@@ -20,16 +24,12 @@ jobs:
       contents: read
     env:
       DOCKER_BASE_TAG: ghcr.io/ocr-d docker.io/ocrd
-      # TODO(kba): make the interpolation work correctly
-      # DOCKER_BUILD: docker buildx build --progress=plain --platform ${{ env.PLATFORMS }} --push
-      # TODO(kba): Investigate why ppc64le build hangs on "Installing build dependencies"
-      # TODO(kba): Investigate why arm64 fails with .buildkit_qemu_emulator: /usr/local/bin/conda: Invalid ELF image for this architecture
-      DOCKER_BUILD: docker buildx build --progress=plain --platform linux/amd64 --push
+      DOCKER_BUILD: docker buildx build --progress=plain --push --platform
     steps:
     - name: Export variables
       run: |
         echo "DOCKER_BASE_TAG=${{ env.DOCKER_BASE_TAG }}" >> $GITHUB_ENV
-        echo "DOCKER_BUILD=${{ env.DOCKER_BUILD }}" >> $GITHUB_ENV
+        echo "DOCKER_BUILD=${{ env.DOCKER_BUILD }} ${{ env.PLATFORMS }}" >> $GITHUB_ENV
     - name: Checkout
       uses: actions/checkout@v4
       with:
@@ -55,4 +55,4 @@ jobs:
     - name: Build the Docker image with GPU support
       run: make docker-cuda
     - name: Build the Docker images with GPU support and ML frameworks
-      run: make docker-cuda-tf1 docker-cuda-tf2 docker-cuda-torch
+      run: make docker-cuda-tf2 docker-cuda-torch

Makefile

@@ -66,7 +66,8 @@ get-conda: export CONDA_PREFIX ?= /conda
 # first part of recipe: see micro.mamba.pm/install.sh
 get-conda: OS != uname
 get-conda: PLATFORM = $(subst Darwin,osx,$(subst Linux,linux,$(OS)))
-get-conda: MACHINE = $(or $(filter aarch64 arm64 ppc64le, $(ARCH)), 64)
+get-conda: ARCH != uname -m
+get-conda: MACHINE = $(or $(filter aarch64 ppc64le, $(subst arm64,aarch64,$(ARCH))), 64)
 get-conda: URL = https://micro.mamba.pm/api/micromamba/$(PLATFORM)-$(MACHINE)/latest
 get-conda:
 	curl --retry 6 -Ls $(URL) | tar -xvj bin/micromamba
@@ -85,93 +86,18 @@ endif
 # Dependencies for CUDA installation via Conda
 deps-cuda: PYTHON_PREFIX != $(PYTHON) -c 'import sysconfig; print(sysconfig.get_paths()["purelib"])'
 deps-cuda: get-conda
-# Get CUDA toolkit, including compiler and libraries with dev,
-# however, the Nvidia channels do not provide (recent) cudnn (needed for Torch, TF etc):
-#MAMBA_ROOT_PREFIX=$(CONDA_PREFIX) \
-#conda install -c nvidia/label/cuda-11.8.0 cuda && conda clean -a
-#
-# The conda-forge channel has cudnn and cudatoolkit but no cudatoolkit-dev anymore (and we need both!),
-# so let's combine nvidia and conda-forge (will be same lib versions, no waste of space),
-# but omitting cuda-cudart-dev and cuda-libraries-dev (as these will be pulled by pip for torch anyway):
-	conda install -c nvidia/label/cuda-11.8.0 \
-	                 cuda-nvcc \
-	                 cuda-cccl \
-	 && conda clean -a \
-	 && find $(CONDA_PREFIX) -name "*_static.a" -delete
-#conda install -c conda-forge \
-#          cudatoolkit=11.8.0 \
-#          cudnn=8.8.* && \
-#conda clean -a && \
-#find $(CONDA_PREFIX) -name "*_static.a" -delete
-#
-# Since Torch will pull in the CUDA libraries (as Python pkgs) anyway,
-# let's jump the shark and pull these via NGC index directly,
-# but then share them with the rest of the system so native compilation/linking
-# works, too:
-	shopt -s nullglob; \
-	$(PIP) install nvidia-pyindex \
-	 && $(PIP) install nvidia-cudnn-cu11~=8.7 \
-	                   nvidia-cublas-cu11~=11.11 \
-	                   nvidia-cusparse-cu11~=11.7 \
-	                   nvidia-cusolver-cu11~=11.4 \
-	                   nvidia-curand-cu11~=10.3 \
-	                   nvidia-cufft-cu11~=10.9 \
-	                   nvidia-cuda-runtime-cu11~=11.8 \
-	                   nvidia-cuda-cupti-cu11~=11.8 \
-	                   nvidia-cuda-nvrtc-cu11 \
-	 && for pkg in cudnn cublas cusparse cusolver curand cufft cuda_runtime cuda_cupti cuda_nvrtc; do \
-	        for lib in $(PYTHON_PREFIX)/nvidia/$$pkg/lib/lib*.so.*; do \
-	            base=`basename $$lib`; \
-	            ln -s $$lib $(CONDA_PREFIX)/lib/$$base.so; \
-	            ln -s $$lib $(CONDA_PREFIX)/lib/$${base%.so.*}.so; \
-	        done \
-	     && for inc in $(PYTHON_PREFIX)/nvidia/$$pkg/include/*; do \
-	            base=`basename $$inc`; case $$base in __*) continue; esac; \
-	            ln -s $$inc $(CONDA_PREFIX)/include/; \
-	        done \
-	    done \
-	 && ldconfig
-# gputil/nvidia-smi would be nice, too – but that drags in Python as a conda dependency...
-
-# Workaround for missing prebuilt versions of TF<2 for Python==3.8
-# todo: find another solution for 3.9, 3.10 etc
-# https://docs.nvidia.com/deeplearning/frameworks/tensorflow-wheel-release-notes/tf-wheel-rel.html
-# Nvidia has them, but under a different name, so let's rewrite that:
-# (hold at nv22.11, because newer releases require CUDA 12, which is not supported by TF2 (at py38),
-#  and therefore not in our ocrd/core-cuda base image yet)
-# However, at that time no Numpy 1.24 was known, which breaks TF1
-# (which is why later nv versions hold it at <1.24 automatically -
-#  see https://github.com/NVIDIA/tensorflow/blob/r1.15.5%2Bnv22.11/tensorflow/tools/pip_package/setup.py)
-deps-tf1:
-	if $(PYTHON) -c 'import sys; print("%u.%u" % (sys.version_info.major, sys.version_info.minor))' | fgrep 3.8 && \
-	! $(PIP) show -q tensorflow-gpu; then \
-	  $(PIP) install nvidia-pyindex && \
-	  pushd $$(mktemp -d) && \
-	  $(PIP) download --no-deps nvidia-tensorflow==1.15.5+nv22.11 && \
-	  for name in nvidia_tensorflow-*.whl; do name=$${name%.whl}; done && \
-	  $(PYTHON) -m wheel unpack $$name.whl && \
-	  for name in nvidia_tensorflow-*/; do name=$${name%/}; done && \
-	  newname=$${name/nvidia_tensorflow/tensorflow_gpu} &&\
-	  sed -i s/nvidia_tensorflow/tensorflow_gpu/g $$name/$$name.dist-info/METADATA && \
-	  sed -i s/nvidia_tensorflow/tensorflow_gpu/g $$name/$$name.dist-info/RECORD && \
-	  sed -i s/nvidia_tensorflow/tensorflow_gpu/g $$name/tensorflow_core/tools/pip_package/setup.py && \
-	  pushd $$name && for path in $$name*; do mv $$path $${path/$$name/$$newname}; done && popd && \
-	  $(PYTHON) -m wheel pack $$name && \
-	  $(PIP) install $$newname*.whl && popd && rm -fr $$OLDPWD; \
-	  $(PIP) install "numpy<1.24" -r $$DIRSTACK/requirements.txt; \
-	else \
-	  $(PIP) install "tensorflow-gpu<2.0" -r requirements.txt; \
-	fi
+# Get CUDA toolkit, including compiler and libraries with dev from NVIDIA channels
+# Get CUDNN (needed for Torch, TF etc) from conda-forge.
+# CUDA runtime libs will be pulled by `pip` for TF and Torch differently anyway,
+# so do _not_ install them here to avoid wasting space.
+	conda install -c nvidia/label/cuda-12.4.0 cuda-minimal-build \
+	&& conda clean -a && ldconfig
 
 deps-tf2:
-	if $(PYTHON) -c 'import sys; print("%u.%u" % (sys.version_info.major, sys.version_info.minor))' | fgrep 3.8; then \
-	$(PIP) install tensorflow -r requirements.txt; \
-	else \
-	$(PIP) install "tensorflow[and-cuda]"  -r requirements.txt; \
-	fi
+	$(PIP) install "tensorflow[and-cuda]"  -r requirements.txt
 
 deps-torch:
-	$(PIP) install -i https://download.pytorch.org/whl/cu118 torchvision==0.16.2+cu118 torch==2.1.2+cu118 -r requirements.txt
+	$(PIP) install torch==2.5.1 torchvision==0.20.1 -i https://download.pytorch.org/whl/cu124 -r requirements.txt
 
 # deps-*: always mix core's requirements.txt with additional deps,
 # so pip does not ignore the older version reqs,
@@ -180,11 +106,11 @@ deps-torch:
 # Dependencies for deployment in an ubuntu/debian linux
 deps-ubuntu:
 	apt-get update
-	apt-get install -y python3 imagemagick libgeos-dev libxml2-dev libxslt-dev libssl-dev
+	apt-get install -y bzip2 python3 imagemagick libgeos-dev libxml2-dev libxslt-dev libssl-dev
 
 # Dependencies for deployment via Conda
 deps-conda: get-conda
-	conda install -c conda-forge python==3.8.* imagemagick geos pkgconfig
+	conda install -c conda-forge python==3.10.* imagemagick geos pkgconfig
 
 # Install test python deps via pip
 deps-test: