Skip some CPU-only tests on CircleCI machines with GPU (#4002)

Author: NicolasHug

.circleci/config.yml

@@ -687,12 +687,20 @@ jobs:
           paths:
             - conda
             - env
+      - 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 avaiable 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
+          command: echo "CIRCLECI=true" >> ./env.list
       - run:
           name: Install torchvision
           command: docker run -t --gpus all -v $PWD:$PWD -w $PWD -e UPLOAD_CHANNEL -e CU_VERSION "${image_name}" .circleci/unittest/linux/scripts/install.sh
       - run:
           name: Run tests
-          command: docker run -e CIRCLECI -t --gpus all -v $PWD:$PWD -w $PWD "${image_name}" .circleci/unittest/linux/scripts/run_test.sh
+          command: docker run --env-file ./env.list -t --gpus all -v $PWD:$PWD -w $PWD "${image_name}" .circleci/unittest/linux/scripts/run_test.sh
       - run:
           name: Post Process
           command: docker run -t --gpus all -v $PWD:$PWD -w $PWD "${image_name}" .circleci/unittest/linux/scripts/post_process.sh

.circleci/config.yml.in

@@ -687,12 +687,20 @@ jobs:
           paths:
             - conda
             - env
+      - 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 avaiable 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
+          command: echo "CIRCLECI=true" >> ./env.list
       - run:
           name: Install torchvision
           command: docker run -t --gpus all -v $PWD:$PWD -w $PWD -e UPLOAD_CHANNEL -e CU_VERSION "${image_name}" .circleci/unittest/linux/scripts/install.sh
       - run:
           name: Run tests
-          command: docker run -e CIRCLECI -t --gpus all -v $PWD:$PWD -w $PWD "${image_name}" .circleci/unittest/linux/scripts/run_test.sh
+          command: docker run --env-file ./env.list -t --gpus all -v $PWD:$PWD -w $PWD "${image_name}" .circleci/unittest/linux/scripts/run_test.sh
       - run:
           name: Post Process
           command: docker run -t --gpus all -v $PWD:$PWD -w $PWD "${image_name}" .circleci/unittest/linux/scripts/post_process.sh

test/common_utils.py

@@ -26,6 +26,7 @@
 IN_RE_WORKER = os.environ.get("INSIDE_RE_WORKER") is not None
 IN_FBCODE = os.environ.get("IN_FBCODE_TORCHVISION") == "1"
 CUDA_NOT_AVAILABLE_MSG = 'CUDA device not available'
+CIRCLECI_GPU_NO_CUDA_MSG = "We're in a CircleCI GPU machine, and this test doesn't need cuda."
 
 
 @contextlib.contextmanager
@@ -256,11 +257,19 @@ def call_args_to_kwargs_only(call_args, *callable_or_arg_names):
 
 
 def cpu_and_gpu():
-    # TODO: make this properly handle CircleCI
     import pytest  # noqa
 
     # ignore CPU tests in RE as they're already covered by another contbuild
-    devices = [] if IN_RE_WORKER else ['cpu']
+    # also ignore CPU tests in CircleCI machines that have a GPU: these tests
+    # are run on CPU-only machines already.
+    if IN_RE_WORKER:
+        devices = []
+    else:
+        if IN_CIRCLE_CI and torch.cuda.is_available():
+            mark = pytest.mark.skip(reason=CIRCLECI_GPU_NO_CUDA_MSG)
+        else:
+            mark = ()
+        devices = [pytest.param('cpu', marks=mark)]
 
     if torch.cuda.is_available():
         cuda_marks = ()
@@ -278,7 +287,6 @@ def cpu_and_gpu():
 
 
 def needs_cuda(test_func):
-    # TODO: make this properly handle CircleCI
     import pytest  # noqa
 
     if IN_FBCODE and not IN_RE_WORKER:
@@ -293,12 +301,13 @@ def needs_cuda(test_func):
 
 
 def cpu_only(test_func):
-    # TODO: make this properly handle CircleCI
     import pytest  # noqa
 
     if IN_RE_WORKER:
         # The assumption is that all RE workers have GPUs.
         return pytest.mark.dont_collect(test_func)
+    elif IN_CIRCLE_CI and torch.cuda.is_available():
+        return pytest.mark.skip(reason=CIRCLECI_GPU_NO_CUDA_MSG)(test_func)
     else:
         return test_func