setup: locate nvvm for CUDA 13

CUDA 13 split libnvvm off the nvcc component and moved its dll into an x64
subdirectory, so the release job installed no nvvm at all and the hardcoded
nvvm/bin copy aborted the setup.

Install the nvvm sub-package and accept both the CUDA 12 and CUDA 13 dll
locations, failing with a hint if none is present.

Author: PearCoding

.github/workflows/release-windows.yml

@@ -22,14 +22,15 @@ jobs:
 
       # The CUDA toolkit is needed so the installer ships the CUDA device
       # runtime (runtime_cuda + bundled nvrtc/nvvm). A network install with
-      # just the required sub-packages keeps this reasonably small: nvcc
-      # carries libnvvm, cudart the driver stub, nvrtc the runtime compiler.
+      # just the required sub-packages keeps this reasonably small: cudart
+      # provides the driver stub, nvrtc the runtime compiler. Since CUDA 13
+      # libnvvm is no longer part of nvcc and needs its own sub-package.
       - name: Setup CUDA toolkit
         uses: Jimver/cuda-toolkit@v0.2.35
         with:
           cuda: "13.2.0"
           method: network
-          sub-packages: '["nvcc", "cudart", "nvrtc", "nvrtc_dev"]'
+          sub-packages: '["nvcc", "nvvm", "cudart", "nvrtc", "nvrtc_dev"]'
 
       - name: Setup sccache
         uses: hendrikmuhs/ccache-action@v1.2

scripts/setup/setup_runtime.ps1

@@ -40,7 +40,12 @@ If ($IsLinux) {
     }
 
     If ($HasCuda) {
-        Copy-Item "$CUDA\nvvm\bin\nvvm*.dll" "$BIN_ROOT/" > $null
+        # CUDA 13 moved the nvvm dll into an x64 subdirectory and split nvvm off the nvcc component
+        $NVVM_BIN = @("$CUDA\nvvm\bin\x64", "$CUDA\nvvm\bin") | Where-Object { Test-Path -Path "$_\nvvm*.dll" } | Select-Object -First 1
+        If ($null -eq $NVVM_BIN) {
+            throw "No nvvm library found in '$CUDA\nvvm'. Since CUDA 13 it is provided by the separate 'nvvm' toolkit component."
+        }
+        Copy-Item "$NVVM_BIN\nvvm*.dll" "$BIN_ROOT/" > $null
     }
 }