diff --git a/docs/source/install.md b/docs/source/install.md
index a79c725..c1b317e 100644
--- a/docs/source/install.md
+++ b/docs/source/install.md
@@ -18,6 +18,7 @@ conda create -n rid-drop python=3.9 libtensorflow_cc=2.6.2=*cuda110* tensorflow=
 ```bash
 conda activate rid-drop
 conda install mdtraj nccl -c conda-forge
+conda install cudatoolkit-dev -c conda-forge
 pip install cmake cython
 pip install matplotlib parmed scikit-learn dpdata
 ```
diff --git a/docs/source/install_dp.md b/docs/source/install_dp.md
index d73bc91..eb581ee 100644
--- a/docs/source/install_dp.md
+++ b/docs/source/install_dp.md
@@ -12,10 +12,14 @@ The `enviroment` of rid-kit software is a bit complex, it uses `dflow` to manage
 ### Install tensorflow's C++ interface and other necessary packages
 We recommend using conda to manage the python enviroment. 
 Use the command
-```
+```bash
 conda create -n rid-dp python=3.9 libtensorflow_cc=2.6.2=*cuda110* tensorflow=2.6.2=*cuda110* nccl mdtraj numpy scikit-learn cmake dpdata cython -c conda-forge
 ```
 to get the compiled libtensorflow_cc and other necessary packages.
+Install the cudatoolkit which manage the nvcc and other commands.
+```bash
+conda install cudatoolkit-dev -c conda-forge
+```
 After installation, activate the enviroment and set library path
 ```bash
 conda activate rid-dp