diff --git a/pkgs/by-name/in/intel-llvm/package.nix b/pkgs/by-name/in/intel-llvm/package.nix
index 04fba865125a..ff0f492f2e29 100644
--- a/pkgs/by-name/in/intel-llvm/package.nix
+++ b/pkgs/by-name/in/intel-llvm/package.nix
@@ -111,7 +111,20 @@ let
         mkdir "$rsrc"
         echo "-resource-dir=$rsrc" >> $out/nix-support/cc-cflags
         ln -s "${lib.getLib self.unwrapped}/lib/clang/${self.llvmMajorVersion}/include" "$rsrc"
-      '';
+      ''
+      + (lib.concatStrings (
+        lib.mapAttrsToList (k: v: ''
+          echo "export ${k}=${v}" >> $out/nix-support/setup-hook
+        '') self.unwrapped.unified-runtime.setupVars
+      ))
+
+      + (lib.optionalString (self.unwrapped.unified-runtime.setupVars ? CUDA_PATH) ''
+        # SYCL CUDA runtime libs (e.g. libonemath_blas_cublas.so) carry DT_NEEDED: libcuda.so.1.
+        # GNU ld resolves transitive DT_NEEDED via -rpath-link, not -L; point it at the stubs.
+        echo "-rpath-link,${self.unwrapped.unified-runtime.setupVars.CUDA_PATH}/lib/stubs" >> $out/nix-support/cc-ldflags
+        # The SYCL CUDA backend discovers libdevice by finding ptxas in PATH.
+        echo "export PATH=${self.unwrapped.unified-runtime.setupVars.CUDA_PATH}/bin''${PATH:+:$PATH}" >> $out/nix-support/setup-hook
+      '');
 
       extraPackages =
         # We need to explicitly link to the dev package to get headers like sycl.hpp
diff --git a/pkgs/by-name/in/intel-llvm/unified-runtime.nix b/pkgs/by-name/in/intel-llvm/unified-runtime.nix
index 78d673ca8094..1bd3cf296469 100644
--- a/pkgs/by-name/in/intel-llvm/unified-runtime.nix
+++ b/pkgs/by-name/in/intel-llvm/unified-runtime.nix
@@ -18,6 +18,7 @@
   pkg-config,
   lit,
   filecheck,
+  buildPackages,
   rocmPackages ? { },
   rocmGpuTargets ? lib.optionalString (rocmPackages ? clr.gpuTargets) (
     builtins.concatStringsSep ";" rocmPackages.clr.gpuTargets
@@ -41,10 +42,22 @@ let
     ];
   };
 
+  # Minimal rocm join required at runtime
+  # We pass this to clang in its wrapper later.
+  # This is a separate join from the above because we don't
+  # need to pull in hsakmt and comgr at runtime, only build time.
+  rocmPath = symlinkJoin {
+    name = "rocm-path";
+    paths = with rocmPackages; [
+      clr
+      rocm-device-libs
+    ];
+  };
+
   cudatoolkit_joined = symlinkJoin {
     name = "cuda-merged";
 
-    paths = with cudaPackages; [
+    paths = with buildPackages.cudaPackages; [
       cuda_cudart
       cuda_nvcc
       cuda_nvml_dev.include
@@ -158,6 +171,10 @@ stdenv.mkDerivation (finalAttrs: {
     (lib.cmakeFeature "CUDA_cupti_LIBRARY" "${cudatoolkit_joined}/lib/libcupti.so")
   ];
 
+  passthru.setupVars =
+    lib.optionalAttrs rocmSupport { ROCM_PATH = rocmPath; }
+    // lib.optionalAttrs cudaSupport { CUDA_PATH = cudatoolkit_joined; };
+
   passthru.backends =
     lib.optionals levelZeroSupport [
       "level_zero"