--- published/pipeline/nvrtc_tce.py
+++ portable/pipeline/nvrtc_tce.py
@@ -40,7 +40,7 @@
     use_cubin = _os.environ.get("NVRTC_CUBIN", "1") == "1"
     if use_cubin:
         arch = arch.replace(b"compute_", b"sm_")
-    opts = [b"--gpu-architecture=" + arch, b"-I/usr/local/cuda/include",
+    opts = [b"--gpu-architecture=" + arch, b"-I/tmp/kernelbench-h100-attempt2/deps/nvidia/cuda_runtime/include", b"-I/tmp/kernelbench-h100-attempt2/deps/nvidia/cuda_cccl/include",
             b"--std=c++17", b"-default-device"]
     arr = (ctypes.c_char_p * len(opts))(*opts)
     rc = nvrtc.nvrtcCompileProgram(prog, len(opts), arr)