diff --git a/alpa/device_mesh.py b/alpa/device_mesh.py
index e74ab418c..4991f81dc 100644
--- a/alpa/device_mesh.py
+++ b/alpa/device_mesh.py
@@ -1092,6 +1092,7 @@ def launch_xla_servers(self):
                     "FI_EFA_USE_DEVICE_RDMA": "1",
                     "LD_LIBRARY_PATH": os.environ.get("LD_LIBRARY_PATH",
                                                       ""),  # For libnccl-net.so
+                    "NCCL_PROTO": "simple",
                 })
 
             bundle_index = device_bundle_idx_list[i]
diff --git a/benchmark/cupy/profile_communication.py b/benchmark/cupy/profile_communication.py
index ecc6bde0c..63197e5a4 100644
--- a/benchmark/cupy/profile_communication.py
+++ b/benchmark/cupy/profile_communication.py
@@ -246,6 +246,7 @@ def sync(self):
                 "FI_PROVIDER": "efa",
                 "FI_EFA_USE_DEVICE_RDMA": "1",
                 "LD_LIBRARY_PATH": os.environ.get("LD_LIBRARY_PATH", ""),  # For libnccl-net.so
+                "NCCL_PROTO": "simple",
             }
         elif args.ib:
             env_vars = {
diff --git a/docs/install.rst b/docs/install.rst
index 8cb52fe91..41fdebd57 100644
--- a/docs/install.rst
+++ b/docs/install.rst
@@ -20,8 +20,8 @@ Regardless of installing from wheels or from source, there are a few prerequisit
     # Update pip
     pip3 install --upgrade pip
 
-    # Use your own CUDA version. Here cuda-cuda114 means cuda 11.4
-    pip3 install cupy-cuda114
+    # Install cupy
+    pip3 install cupy-cuda11x
 
   Then, check whether your system already has NCCL installed.
 
@@ -32,6 +32,9 @@ Regardless of installing from wheels or from source, there are a few prerequisit
   If it prints nothing, then NCCL has already been installed.
   Otherwise, follow the printed instructions to install NCCL.
 
+  .. code:: bash
+
+    python3 -m cupyx.tools.install_library --cuda 11.x --library nccl
 
 Methods
 -------