diff --git a/examples/docker/Dockerfile b/examples/docker/Dockerfile
index 37bed888c..47fd422cd 100644
--- a/examples/docker/Dockerfile
+++ b/examples/docker/Dockerfile
@@ -1,5 +1,5 @@
 # Base image must at least have pytorch and CUDA installed.
-ARG BASE_IMAGE=nvcr.io/nvidia/pytorch:19.03-py3
+ARG BASE_IMAGE=nvcr.io/nvidia/pytorch:19.07-py3
 FROM $BASE_IMAGE
 ARG BASE_IMAGE
 RUN echo "Installing Apex on top of ${BASE_IMAGE}"
diff --git a/setup.py b/setup.py
index f94925a68..53b84d08e 100644
--- a/setup.py
+++ b/setup.py
@@ -5,10 +5,20 @@
 from pip._internal import main as pipmain
 import sys
 import warnings
+import os
 
 if not torch.cuda.is_available():
-    print("\nWarning: Torch did not find available GPUs on this system.\n",
-          "If your intention is to cross-compile, this is not an error.\n")
+    # https://github.com/NVIDIA/apex/issues/486
+    # Extension builds after https://github.com/pytorch/pytorch/pull/23408 attempt to query torch.cuda.get_device_capability(),
+    # which will fail if you are compiling in an environment without visible GPUs (e.g. during an nvidia-docker build command).
+    print('\nWarning: Torch did not find available GPUs on this system.\n',
+          'If your intention is to cross-compile, this is not an error.\n'
+          'By default, Apex will cross-compile for Pascal (compute capabilities 6.0, 6.1, 6.2),\n'
+          'Volta (compute capability 7.0), and Turing (compute capability 7.5).\n'
+          'If you wish to cross-compile for a single specific architecture,\n'
+          'export TORCH_CUDA_ARCH_LIST="compute capability" before running setup.py.\n')
+    if os.environ.get("TORCH_CUDA_ARCH_LIST", None) is None:
+        os.environ["TORCH_CUDA_ARCH_LIST"] = "6.0;6.1;6.2;7.0;7.5"
 
 print("torch.__version__  = ", torch.__version__)
 TORCH_MAJOR = int(torch.__version__.split('.')[0])