diff --git a/powerinfer/Dockerfile b/powerinfer/Dockerfile
index 5d68019..a1de8fe 100644
--- a/powerinfer/Dockerfile
+++ b/powerinfer/Dockerfile
@@ -43,10 +43,10 @@ FROM nvidia/cuda:12.3.1-devel-rockylinux9
 ARG USERID=1000
 RUN yum install -y python3-pip cmake libcudnn8 git && yum clean all && rm -rf /var/cache/yum/*
 RUN git clone https://github.com/SJTU-IPADS/PowerInfer
-WORKDIR PowerInfer
+WORKDIR /PowerInfer
 RUN pip install --no-cache-dir -r requirements.txt
 RUN cmake -S . -B build -DLLAMA_CUBLAS=ON
-RUN cmake --build build --config Release -j $(nproc)
+RUN cmake --build build --config Release -j "$(nproc)"
 RUN pip install --no-cache-dir pandas #for the benchmark.
 RUN adduser -u $USERID user
 USER user
diff --git a/powerinfer/README.md b/powerinfer/README.md
index a72ad06..13bae21 100644
--- a/powerinfer/README.md
+++ b/powerinfer/README.md
@@ -1,4 +1,6 @@
-[PowerInfer](https://github.com/SJTU-IPADS/PowerInfer)
+#PowerInfer benchmark
+
+Benchmark for [PowerInfer](https://github.com/SJTU-IPADS/PowerInfer).
 
 Note that the model loses some inference quality in exchange for speed as shown in https://huggingface.co/SparseLLM/ReluLLaMA-7B.