diff --git a/Dockerfile b/Dockerfile
index 59b3943..594db86 100644
--- a/Dockerfile
+++ b/Dockerfile
@@ -8,6 +8,16 @@ ENV PATH="/root/.local/bin:$PATH"
 
 RUN ldconfig /usr/local/cuda-13.0/compat/
 
+# nixl_ep PyPI wheels are compiled against CUDA 12.x and require libcudart.so.12.
+# CUDA 13 runtime is ABI-compatible with CUDA 12, so symlinking is safe.
+# Symlink into /usr/local/lib so it is in the default linker search path.
+RUN ln -sf /usr/local/cuda/lib64/libcudart.so.13 /usr/local/lib/libcudart.so.12 && ldconfig
+
+# CUDA 13.0 containers return libs to /usr/local/nvidia/lib64 so container
+# providers (RunPod, Lambda, etc.) can mount host drivers there consistently.
+# See: https://github.com/vllm-project/vllm/issues/18859
+ENV LD_LIBRARY_PATH=/usr/local/nvidia/lib64:/usr/local/cuda/lib64:$LD_LIBRARY_PATH
+
 # Install vLLM with FlashInfer - use CUDA 130 PyTorch wheels
 RUN uv pip install --system "packaging>=24.2" && \
     uv pip install --system "vllm[flashinfer]==0.21.0" && \
diff --git a/builder/requirements.txt b/builder/requirements.txt
index b05ff13..90a2bc2 100644
--- a/builder/requirements.txt
+++ b/builder/requirements.txt
@@ -11,5 +11,4 @@ pydantic-settings
 hf-transfer
 transformers>=5
 bitsandbytes>=0.45.0
-kernels
 torch-c-dlpack-ext