diff --git a/Dockerfile b/Dockerfile
index 646f6e1..cc2fd95 100644
--- a/Dockerfile
+++ b/Dockerfile
@@ -1,6 +1,5 @@
 # DeepSeek V4 NVFP4 vLLM + DeepGEMM Mega MoE
 # Extends the vLLM dream-build container with our custom DeepGEMM kernel
-# and DeepSeek V4 patch.
 
 FROM atl.vultrcr.com/vllm/vllm-with-lmcache:dream-build
 
@@ -11,12 +10,13 @@ RUN apt-get update && apt-get install -y git screen cmake && rm -rf /var/lib/apt
 # CACHE_BUSTER: increment to force fresh clone
 RUN git clone -b nvfp4-mega-moe https://sweetapi.com/biondizzle/DeepGEMM.git /root/DeepGEMM && CACHE_BUSTER=1
 
-# Build DeepGEMM (CUTLASS/CuTe headers come from flashinfer/vllm deps)
-ENV CPATH="/usr/local/lib/python3.12/dist-packages/flashinfer/data/cutlass/include:/usr/local/lib/python3.12/dist-packages/nvidia/cu13/include:${CPATH}"
+# Build DeepGEMM with proper CUDA/NVRTC paths
+ENV CPATH="/usr/local/lib/python3.12/dist-packages/flashinfer/data/cutlass/include:/usr/local/lib/python3.12/dist-packages/nvidia/cu13/include:/usr/local/cuda-13.0/include:${CPATH}"
+ENV LIBRARY_PATH="/usr/local/lib/python3.12/dist-packages/nvidia/cu13/lib:/usr/local/cuda-13.0/targets/x86_64-linux/lib:${LIBRARY_PATH}"
+ENV LD_LIBRARY_PATH="/usr/local/lib/python3.12/dist-packages/nvidia/cu13/lib:/usr/local/cuda-13.0/targets/x86_64-linux/lib:${LD_LIBRARY_PATH}"
 RUN cd /root/DeepGEMM && python3 setup.py build_ext --inplace
 
-# Copy our DeepSeek V4 patch into the image (will be applied at entrypoint)
-# The actual patch file is mounted at runtime, but we stage a default
+# Copy our DeepSeek V4 patch into the image
 COPY patches/deepseek_v4.py /defaults/deepseek_v4.py
 
 # Verify everything imports