diff --git a/azure-pipelines.yml b/azure-pipelines.yml
new file mode 100644
index 0000000..1635a2e
--- /dev/null
+++ b/azure-pipelines.yml
@@ -0,0 +1,39 @@
+#
+trigger:
+  paths:
+    exclude:
+      - docs      
+jobs:
+  - job: Fairseq_unittests
+    timeoutInMinutes: 360
+    #pool:
+    #  vmImage: 'ubuntu-18.04'
+    pool: 
+      name: default
+      demands:
+        - agent.name -equals gpu4 
+    steps:
+    - script: |
+        #set up docker
+        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
+        #install fastseq
+        pip install --editable .
+      
+        #show environment
+        which python
+        python --version
+        which nvcc
+        nvcc --version
+        which fastseq
+        python -c "import torch; print('torch:', torch.__version__, torch)"
+        python -c "import torch; print('CUDA available:', torch.cuda.is_available())"
+        python -c "import fastseq; print('fastseq:', fastseq.__version__)"
+        
+        #run unit tests
+        CUDA_VISIBLE_DEVICES=3 bash tests/run_fairseq_tests.sh
+        CUDA_VISIBLE_DEVICES=3 bash tests/run_transformers_tests.sh
+        
+        #run benchmarks 
+        #cd benchmarks/
+        #CUDA_VISIBLE_DEVICES=3 run_all_benchmarks.sh
+      displayName: 'setup environment and run fairseq unit tests'  
diff --git a/fastseq/optimizer/fairseq/beam_search_optimizer_v1.py b/fastseq/optimizer/fairseq/beam_search_optimizer_v1.py
index 83adba4..36ce041 100644
--- a/fastseq/optimizer/fairseq/beam_search_optimizer_v1.py
+++ b/fastseq/optimizer/fairseq/beam_search_optimizer_v1.py
@@ -53,7 +53,7 @@ class MultiheadAttentionV2(MultiheadAttention):
     """
 
     def __init__(self,
-                 embed_dim,
+                 embed_dim__,
                  num_heads,
                  kdim=None,
                  vdim=None,