diff --git a/.github/workflows/ci-cuda.yml b/.github/workflows/ci-cuda.yml
new file mode 100644
index 0000000000000000000000000000000000000000..2fac75c90230467c1fc34f44d0d1891a8038a735
--- /dev/null
+++ b/.github/workflows/ci-cuda.yml
@@ -0,0 +1,29 @@
+name: CI GPU
+
+on:
+  push:
+    branches: [master]
+  pull_request:
+
+jobs:
+  gpu-test:
+    runs-on: self-hosted
+    container:
+      image: localhost:5000/megengine-ci:latest
+      volumes:
+        - /usr/local/cuda-10.1-libs:/usr/local/cuda-10.1-libs
+      options: --gpus all
+
+    steps:
+      - name: Checkout MegEngine
+        uses: actions/checkout@v2
+      - name: Checkout submodules
+        run: |
+          ./third_party/prepare.sh
+          ./third_party/install-mkl.sh
+      - name: Build MegEngine
+        run: ./ci/cmake.sh cuda
+      - name: Python test
+        run: ./ci/run_python_test.sh
+      - name: C++ test
+        run: ./ci/run_cpp_test.sh cuda
diff --git a/ci/docker_env/Dockerfile b/ci/docker_env/Dockerfile
index 8f525f5cf8cbe6ac8a7b95b8d5df800313ca5664..4496fc99d72652d0814cf1ed45680af0abf9e4b4 100644
--- a/ci/docker_env/Dockerfile
+++ b/ci/docker_env/Dockerfile
@@ -35,5 +35,6 @@ RUN git lfs install
 
 ENV PATH=${PATH}:/usr/local/cuda/bin \
   LIBRARY_PATH=${LIBRARY_PATH}:/usr/local/cuda/lib:/usr/local/cuda/lib64:/usr/local/cuda/lib/stubs:/usr/local/cuda/lib64/stubs:/usr/local/cuda-10.1-libs/cudnn-v7.6.0/lib:/usr/local/cuda-10.1-libs/cudnn-v7.6.0/lib64:/usr/local/cuda-10.1-libs/TensorRT-5.1.5.0/lib:/usr/local/cuda-10.1-libs/TensorRT-5.1.5.0/lib64 \
-  LD_LIBRARY_PATH=${LD_LIBRARY_PATH}:/usr/local/nvidia/lib:/usr/local/nvidia/lib64:/usr/local/cuda/lib:/usr/local/cuda/lib64:/usr/local/cuda/lib/stubs:/usr/local/cuda/lib64/stubs:/usr/local/cuda-10.1-libs/cudnn-v7.6.0/lib:/usr/local/cuda-10.1-libs/cudnn-v7.6.0/lib64:/usr/local/cuda-10.1-libs/TensorRT-5.1.5.0/lib:/usr/local/cuda-10.1-libs/TensorRT-5.1.5.0/lib64:/usr/local/cuda/lib64/stubs/libcuda.so:/tmp/build/cuda/dnn/cuda-stub/libcuda.so.1 \
-  CPATH=${CPATH}:/usr/local/cuda/include:/usr/local/cuda-10.1-libs/cudnn-v7.6.0/include:/usr/local/cuda-10.1-libs/TensorRT-5.1.5.0/include
\ No newline at end of file
+  LD_LIBRARY_PATH=/usr/local/nvidia/lib:/usr/local/nvidia/lib64:/usr/local/nvidia/lib:/usr/local/nvidia/lib64:/usr/local/cuda-10.1-libs/cudnn-v7.6.0/lib:/usr/local/cuda-10.1-libs/cudnn-v7.6.0/lib64:/usr/local/cuda-10.1-libs/TensorRT-5.1.5.0/lib:/usr/local/cuda-10.1-libs/TensorRT-5.1.5.0/lib64:/tmp/build/cuda/dnn/cuda-stub/libcuda.so \
+  CPATH=${CPATH}:/usr/local/cuda/include:/usr/local/cuda-10.1-libs/cudnn-v7.6.0/include:/usr/local/cuda-10.1-libs/TensorRT-5.1.5.0/include \
+  CUDA_BIN_PATH=/usr/local/cuda