diff --git a/azure-pipelines.yml b/azure-pipelines.yml
new file mode 100644
index 0000000..153a2fc
--- /dev/null
+++ b/azure-pipelines.yml
@@ -0,0 +1,43 @@
+# Python package
+
+trigger:
+  paths:
+    exclude:
+      - docs      
+jobs:
+  - job: Fastseq_Tests
+    timeoutInMinutes: 360
+    #pool:
+    #  vmImage: 'ubuntu-18.04'
+    pool: 
+      name: default
+      demands:
+        - agent.name -equals gpu4 
+
+    steps:
+        # Unfortunately nvidia's nvcc_linux-64=<version> seems to install 10.1 regardless?
+        # Most of this complexity is a workaround to get the compiler toolchain to match the
+        # cudatoolkit runtime
+    - script: |
+        #conda create --force --yes -n $(conda_env) python=$(python.version) cudatoolkit=$(cuda.version)
+        #conda init bash
+        #conda activate $(conda_env)
+        #conda install -q --yes conda
+        #conda install -q --yes pip
+        #conda install -q --yes gxx_linux-64
+        #if [[ $(cuda.version) != "10.2" ]]; then conda install --yes -c conda-forge cudatoolkit-dev=$(cuda.version) ; fi
+        #echo "PATH=$PATH, LD_LIBRARY_PATH=$LD_LIBRARY_PATH"
+        python --version
+        #which nvcc
+        #nvcc --version
+        #which fastseq
+        #bash tests/run_fairseq_tests.sh
+        sudo docker run --gpus all  --network=host --privileged -w '/root'  --shm-size=1g --ulimit memlock=-1 --ulimit stack=67108864  adsbrainwestus2.azurecr.io/fastseq:dev-py3 /bin/bash
+        ls
+        pip install --editable .
+        bash tests/run_fairseq_tests.sh
+        python --version
+      displayName: 'Setup environment python=$(python.version) pytorch=$(pytorch.version) cuda=$(cuda.version)'
+
+
+