[CI] cuda 11.3 update (#235)

Author: vmoens

.circleci/config.yml

@@ -110,7 +110,7 @@ binary_common: &binary_common
     wheel_docker_image:
       description: "Wheel only: what docker image to use"
       type: string
-      default: "pytorch/manylinux-cuda102"
+      default: "pytorch/manylinux-cuda113"
     conda_docker_image:
       description: "Conda only: what docker image to use"
       type: string
@@ -218,20 +218,21 @@ jobs:
     <<: *binary_common
 
     docker:
-      - image: "pytorch/manylinux-cuda102"
+      - image: "pytorch/manylinux-cuda113"
     resource_class: 2xlarge+
 
     environment:
       TAR_OPTIONS: --no-same-owner
       PYTHON_VERSION: << parameters.python_version >>
+      CU_VERSION: << parameters.cu_version >>
 
     steps:
       - checkout
       - designate_upload_channel
-      - run:
-          name: Generate cache key
-          # This will refresh cache on Sundays, nightly build should generate new cache.
-          command: echo "$(date +"%Y-%U")" > .circleci-weekly
+#      - run:
+#          name: Generate cache key
+#          # This will refresh cache on Sundays, nightly build should generate new cache.
+#          command: echo "$(date +"%Y-%U")" > .circleci-weekly
 #      - restore_cache:
 #          keys:
 #            - env-v2-linux-{{ arch }}-py<< parameters.python_version >>-{{ checksum ".circleci/unittest/linux/scripts/environment.yml" }}-{{ checksum ".circleci-weekly" }}
@@ -262,20 +263,21 @@ jobs:
   unittest_linux_gpu:
     <<: *binary_common
     machine:
-      image: ubuntu-1604-cuda-10.2:202012-01
+      image: ubuntu-2004-cuda-11.4:202110-01
     resource_class: gpu.nvidia.medium
     environment:
-      image_name: "pytorch/manylinux-cuda102"
+      image_name: "pytorch/manylinux-cuda113"
       TAR_OPTIONS: --no-same-owner
       PYTHON_VERSION: << parameters.python_version >>
+      CU_VERSION: << parameters.cu_version >>
 
     steps:
       - checkout
       - designate_upload_channel
-      - run:
-          name: Generate cache key
-          # This will refresh cache on Sundays, nightly build should generate new cache.
-          command: echo "$(date +"%Y-%U")" > .circleci-weekly
+#      - run:
+#          name: Generate cache key
+#          # This will refresh cache on Sundays, nightly build should generate new cache.
+#          command: echo "$(date +"%Y-%U")" > .circleci-weekly
 #      - restore_cache:
 #
 #          keys:
@@ -294,7 +296,7 @@ jobs:
       - run:
           # Here we create an envlist file that contains some env variables that we want the docker container to be aware of.
           # Normally, the CIRCLECI variable is set and available on all CI workflows: https://circleci.com/docs/2.0/env-vars/#built-in-environment-variables.
-          # They're availble in all the other workflows (OSX and Windows).
+          # They're available in all the other workflows (OSX and Windows).
           # But here, we're running the unittest_linux_gpu workflows in a docker container, where those variables aren't accessible.
           # So instead we dump the variables we need in env.list and we pass that file when invoking "docker run".
           name: export CIRCLECI env var
@@ -314,28 +316,29 @@ jobs:
   unittest_linux_optdeps_gpu:
     <<: *binary_common
     machine:
-      image: ubuntu-1604-cuda-10.2:202012-01
+      image: ubuntu-2004-cuda-11.4:202110-01
     resource_class: gpu.nvidia.medium
     environment:
-      image_name: "pytorch/manylinux-cuda102"
+      image_name: "pytorch/manylinux-cuda113"
       TAR_OPTIONS: --no-same-owner
       PYTHON_VERSION: << parameters.python_version >>
+      CU_VERSION: << parameters.cu_version >>
 
     steps:
       - checkout
       - designate_upload_channel
-      - run:
-          name: Generate cache key
-          # This will refresh cache on Sundays, nightly build should generate new cache.
-          command: echo "$(date +"%Y-%U")" > .circleci-weekly
+#      - run:
+#          name: Generate cache key
+#          # This will refresh cache on Sundays, nightly build should generate new cache.
+#          command: echo "$(date +"%Y-%U")" > .circleci-weekly
 #      - restore_cache:
 #
 #          keys:
 #            - env-v3-linux-{{ arch }}-py<< parameters.python_version >>-{{ checksum ".circleci/unittest/linux_optdeps/scripts/environment.yml" }}-{{ checksum ".circleci-weekly" }}
 
       - run:
           name: Setup
-          command: docker run -e PYTHON_VERSION -t --gpus all -v $PWD:$PWD -w $PWD "${image_name}" .circleci/unittest/linux_optdeps/scripts/setup_env.sh
+          command: .circleci/unittest/linux_optdeps/scripts/setup_env.sh
 #      - save_cache:
 #
 #          key: env-v3-linux-{{ arch }}-py<< parameters.python_version >>-{{ checksum ".circleci/unittest/linux_optdeps/scripts/environment.yml" }}-{{ checksum ".circleci-weekly" }}
@@ -346,7 +349,7 @@ jobs:
       - run:
           # Here we create an envlist file that contains some env variables that we want the docker container to be aware of.
           # Normally, the CIRCLECI variable is set and available on all CI workflows: https://circleci.com/docs/2.0/env-vars/#built-in-environment-variables.
-          # They're availble in all the other workflows (OSX and Windows).
+          # They're available in all the other workflows (OSX and Windows).
           # But here, we're running the unittest_linux_gpu workflows in a docker container, where those variables aren't accessible.
           # So instead we dump the variables we need in env.list and we pass that file when invoking "docker run".
           name: export CIRCLECI env var
@@ -367,20 +370,21 @@ jobs:
     <<: *binary_common
 
     docker:
-      - image: "pytorch/manylinux-cuda102"
+      - image: "pytorch/manylinux-cuda113"
     resource_class: 2xlarge+
 
     environment:
       TAR_OPTIONS: --no-same-owner
       PYTHON_VERSION: << parameters.python_version >>
+      CU_VERSION: << parameters.cu_version >>
 
     steps:
       - checkout
       - designate_upload_channel
-      - run:
-          name: Generate cache key
-          # This will refresh cache on Sundays, nightly build should generate new cache.
-          command: echo "$(date +"%Y-%U")" > .circleci-weekly
+#      - run:
+#          name: Generate cache key
+#          # This will refresh cache on Sundays, nightly build should generate new cache.
+#          command: echo "$(date +"%Y-%U")" > .circleci-weekly
 #      - restore_cache:
 #
 #          keys:
@@ -412,20 +416,21 @@ jobs:
   unittest_linux_stable_gpu:
     <<: *binary_common
     machine:
-      image: ubuntu-1604-cuda-10.2:202012-01
+      image: ubuntu-2004-cuda-11.4:202110-01
     resource_class: gpu.nvidia.medium
     environment:
-      image_name: "pytorch/manylinux-cuda102"
+      image_name: "pytorch/manylinux-cuda113"
       TAR_OPTIONS: --no-same-owner
       PYTHON_VERSION: << parameters.python_version >>
+      CU_VERSION: << parameters.cu_version >>
 
     steps:
       - checkout
       - designate_upload_channel
-      - run:
-          name: Generate cache key
-          # This will refresh cache on Sundays, nightly build should generate new cache.
-          command: echo "$(date +"%Y-%U")" > .circleci-weekly
+#      - run:
+#          name: Generate cache key
+#          # This will refresh cache on Sundays, nightly build should generate new cache.
+#          command: echo "$(date +"%Y-%U")" > .circleci-weekly
 #      - restore_cache:
 #
 #          keys:
@@ -474,10 +479,10 @@ jobs:
           name: Install wget
           command: HOMEBREW_NO_AUTO_UPDATE=1 brew install wget
           # Disable brew auto update which is very slow
-      - run:
-          name: Generate cache key
-          # This will refresh cache on Sundays, nightly build should generate new cache.
-          command: echo "$(date +"%Y-%U")" > .circleci-weekly
+#      - run:
+#          name: Generate cache key
+#          # This will refresh cache on Sundays, nightly build should generate new cache.
+#          command: echo "$(date +"%Y-%U")" > .circleci-weekly
 #      - restore_cache:
 #
 #          keys:
@@ -526,12 +531,12 @@ workflows:
           python_version: '3.8'
 
       - unittest_linux_gpu:
-          cu_version: cu102
+          cu_version: cu113
           name: unittest_linux_gpu_py3.8
           python_version: '3.8'
 
       - unittest_linux_optdeps_gpu:
-          cu_version: cu102
+          cu_version: cu113
           name: unittest_linux_optdeps_gpu_py3.8
           python_version: '3.8'
 
@@ -541,6 +546,6 @@ workflows:
           python_version: '3.8'
 
       - unittest_linux_stable_gpu:
-          cu_version: cu102
+          cu_version: cu113
           name: unittest_linux_stable_gpu_py3.8
           python_version: '3.8'

.circleci/unittest/linux/scripts/install.sh

@@ -42,7 +42,7 @@ if [ "${CU_VERSION:-}" == cpu ] ; then
       pip3 install torch torchvision -f https://download.pytorch.org/whl/nightly/cpu/torch_nightly.html --pre
     fi
 else
-    pip3 install --pre torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/nightly/cu102
+    pip3 install --pre torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/nightly/cu113
 fi
 
 printf "Installing functorch\n"
@@ -53,7 +53,7 @@ pip install "git+https://github.com/pytorch/functorch.git"
 python -c "import functorch"
 
 printf "* Installing torchrl\n"
-python setup.py develop
+pip install -e .
 
 if [[ $OSTYPE == 'darwin'* ]]; then
   PRIVATE_MUJOCO_GL=glfw

.circleci/unittest/linux/scripts/run_test.sh

@@ -16,15 +16,6 @@ lib_dir="${env_dir}/lib"
 
 # solves ImportError: /lib64/libstdc++.so.6: version `GLIBCXX_3.4.21' not found
 export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$lib_dir
-#if [[ $OSTYPE == 'darwin'* ]]; then
-#  PRIVATE_MUJOCO_GL=glfw
-#else
-#  PRIVATE_MUJOCO_GL=egl
-#fi
-## pygame.error: No available video device
-#export SDL_VIDEODRIVER=dummy
-## solves AttributeError: 'NoneType' object has no attribute 'eglQueryString'
-#export MUJOCO_GL=$PRIVATE_MUJOCO_GL
 
 pytest test/smoke_test.py -v --durations 20
 pytest test/smoke_test_deps.py -v --durations 20

.circleci/unittest/linux/scripts/setup_env.sh

@@ -53,9 +53,4 @@ printf "* Installing dependencies (except PyTorch)\n"
 echo "  - python=${PYTHON_VERSION}" >> "${this_dir}/environment.yml"
 cat "${this_dir}/environment.yml"
 
-#conda install -y -c conda-forge glfw
-#conda install -y -c conda-forge mesa
-#conda install -y -c menpo osmesa
-#conda install -y -c conda-forge glew
-
 conda env update --file "${this_dir}/environment.yml" --prune

.circleci/unittest/linux_optdeps/scripts/install.sh

@@ -32,19 +32,17 @@ esac
 # submodules
 git submodule sync && git submodule update --init --recursive
 
-#printf "Installing PyTorch with %s\n" "${cudatoolkit}"
-#if [ "${os}" == "MacOSX" ]; then
-#    conda install -y -c "pytorch-${UPLOAD_CHANNEL}" "pytorch-${UPLOAD_CHANNEL}"::pytorch "${cudatoolkit}" pytest
-#else
-#    conda install -y -c "pytorch-${UPLOAD_CHANNEL}" "pytorch-${UPLOAD_CHANNEL}"::pytorch[build="*${version}*"] "${cudatoolkit}" pytest
-#fi
-
-#printf "Installing PyTorch with %s\n" "${CU_VERSION}"
+printf "Installing PyTorch with %s\n" "${CU_VERSION}"
 if [ "${CU_VERSION:-}" == cpu ] ; then
+    # conda install -y pytorch torchvision cpuonly -c pytorch-nightly
     # use pip to install pytorch as conda can frequently pick older release
-    pip install torch -f https://download.pytorch.org/whl/nightly/cpu/torch_nightly.html --pre
+    if [[ $OSTYPE == 'darwin'* ]]; then
+      pip3 install --pre torch --extra-index-url https://download.pytorch.org/whl/nightly/cpu
+    else
+      pip3 install torch -f https://download.pytorch.org/whl/nightly/cpu/torch_nightly.html --pre
+    fi
 else
-    conda install -y pytorch cudatoolkit=10.2 -c pytorch-nightly
+    pip3 install --pre torch --extra-index-url https://download.pytorch.org/whl/nightly/cu113
 fi
 
 printf "Installing functorch\n"
@@ -55,7 +53,7 @@ pip install "git+https://github.com/pytorch/functorch.git"
 python -c "import functorch"
 
 printf "* Installing torchrl\n"
-python setup.py develop
+pip install -e .
 
 # smoke test
 python -c "import torchrl"

.circleci/unittest/linux_stable/scripts/install.sh

@@ -13,6 +13,7 @@ conda activate ./env
 if [ "${CU_VERSION:-}" == cpu ] ; then
     cudatoolkit="cpuonly"
     version="cpu"
+    echo "Using cpu build"
 else
     if [[ ${#CU_VERSION} -eq 4 ]]; then
         CUDA_VERSION="${CU_VERSION:2:1}.${CU_VERSION:3:1}"
@@ -38,7 +39,7 @@ if [ "${CU_VERSION:-}" == cpu ] ; then
     # use pip to install pytorch as conda can frequently pick older release
     pip3 install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cpu
 else
-    pip3 install torch torchvision torchaudio
+    pip3 install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu113
 fi
 
 printf "Installing functorch\n"
@@ -51,7 +52,7 @@ printf "* Installing torchrl\n"
 printf "g++ version: "
 gcc --version
 
-python setup.py install
+pip install -e .
 
 if [[ $OSTYPE == 'darwin'* ]]; then
   PRIVATE_MUJOCO_GL=glfw

.circleci/unittest/linux_stable/scripts/run_test.sh

@@ -16,15 +16,6 @@ lib_dir="${env_dir}/lib"
 
 # solves ImportError: /lib64/libstdc++.so.6: version `GLIBCXX_3.4.21' not found
 export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$lib_dir
-#if [[ $OSTYPE == 'darwin'* ]]; then
-#  PRIVATE_MUJOCO_GL=glfw
-#else
-#  PRIVATE_MUJOCO_GL=egl
-#fi
-## pygame.error: No available video device
-#export SDL_VIDEODRIVER=dummy
-## solves AttributeError: 'NoneType' object has no attribute 'eglQueryString'
-#export MUJOCO_GL=$PRIVATE_MUJOCO_GL
 
 pytest test/smoke_test.py -v --durations 20
 pytest test/smoke_test_deps.py -v --durations 20

.circleci/unittest/linux_stable/scripts/setup_env.sh

@@ -53,9 +53,4 @@ printf "* Installing dependencies (except PyTorch)\n"
 echo "  - python=${PYTHON_VERSION}" >> "${this_dir}/environment.yml"
 cat "${this_dir}/environment.yml"
 
-#conda install -y -c conda-forge glfw
-#conda install -y -c conda-forge mesa
-##conda install -y -c menpo osmesa
-#conda install -y -c conda-forge glew
-
 conda env update --file "${this_dir}/environment.yml" --prune

build_tools/setup_helpers/extension.py

@@ -46,12 +46,12 @@ def _get_build(var, default=False):
     False if platform.system() == "Windows" else _get_build("BUILD_KALDI", True)
 )
 _BUILD_RNNT = _get_build("BUILD_RNNT", True)
-_USE_ROCM = _get_build(
-    "USE_ROCM", torch.cuda.is_available() and torch.version.hip is not None
-)
-_USE_CUDA = _get_build(
-    "USE_CUDA", torch.cuda.is_available() and torch.version.hip is None
-)
+_USE_ROCM = False  # _get_build(
+# "USE_ROCM", torch.cuda.is_available() and torch.version.hip is not None
+# )
+_USE_CUDA = False  # _get_build(
+# "USE_CUDA", torch.cuda.is_available() and torch.version.hip is None
+# )
 _USE_OPENMP = (
     _get_build("USE_OPENMP", True)
     and "ATen parallel backend: OpenMP" in torch.__config__.parallel_info()