Add CUDA 13 support with version-based package matching

Author: andriiryzhkov

data/ort_gpu.json

@@ -5,27 +5,61 @@
       "vendor": "nvidia",
       "platform": "linux",
       "arch": "x86_64",
+      "cuda_min": "13.0",
+      "cuda_max": "13.99",
+      "ort_version": "1.24.4",
+      "url": "https://github.com/microsoft/onnxruntime/releases/download/v1.24.4/onnxruntime-linux-x64-gpu_cuda13-1.24.4.tgz",
+      "sha256": "fdc6eb18317b4eaeda8b3b86595e5da7e853f72bac67ccac9b04ffc20c9f7fe0",
+      "format": "tgz",
+      "lib_pattern": "libonnxruntime",
+      "install_subdir": "onnxruntime-cuda",
+      "size_mb": 200,
+      "requirements": "CUDA 13.x, cuDNN 9.x"
+    },
+    {
+      "vendor": "nvidia",
+      "platform": "linux",
+      "arch": "x86_64",
+      "cuda_min": "12.0",
+      "cuda_max": "12.99",
       "ort_version": "1.24.4",
       "url": "https://github.com/microsoft/onnxruntime/releases/download/v1.24.4/onnxruntime-linux-x64-gpu-1.24.4.tgz",
       "sha256": "c5f804ff5d239b436fa59e9f2fb288a39f7eb9552f6a636c8b71e792e91a8808",
       "format": "tgz",
       "lib_pattern": "libonnxruntime",
       "install_subdir": "onnxruntime-cuda",
       "size_mb": 200,
-      "requirements": "NVIDIA driver 525+, CUDA 12.x, cuDNN 9.x"
+      "requirements": "CUDA 12.x, cuDNN 9.x"
+    },
+    {
+      "vendor": "nvidia",
+      "platform": "windows",
+      "arch": "x86_64",
+      "cuda_min": "13.0",
+      "cuda_max": "13.99",
+      "ort_version": "1.24.4",
+      "url": "https://github.com/microsoft/onnxruntime/releases/download/v1.24.4/onnxruntime-win-x64-gpu_cuda13-1.24.4.zip",
+      "sha256": "971be8cf984950672934a3173669590a8ece10b44746883420da8066ba836707",
+      "format": "zip",
+      "lib_pattern": "onnxruntime",
+      "install_subdir": "onnxruntime-cuda",
+      "size_mb": 200,
+      "requirements": "CUDA 13.x, cuDNN 9.x"
     },
     {
       "vendor": "nvidia",
       "platform": "windows",
       "arch": "x86_64",
+      "cuda_min": "12.0",
+      "cuda_max": "12.99",
       "ort_version": "1.24.4",
       "url": "https://github.com/microsoft/onnxruntime/releases/download/v1.24.4/onnxruntime-win-x64-gpu-1.24.4.zip",
       "sha256": "ef3337a0b8184eb8beec310f7c83bd50376b3eefc43aab84ac8e452f6987df0a",
       "format": "zip",
       "lib_pattern": "onnxruntime",
       "install_subdir": "onnxruntime-cuda",
       "size_mb": 200,
-      "requirements": "NVIDIA driver 525+, CUDA 12.x, cuDNN 9.x"
+      "requirements": "CUDA 12.x, cuDNN 9.x"
     },
     {
       "vendor": "amd",

src/ai/ort_install.c

@@ -50,6 +50,8 @@ typedef struct
   char *sha256;        // expected SHA-256 hash (optional but recommended)
   char *rocm_min;      // optional, AMD only
   char *rocm_max;      // optional, AMD only
+  char *cuda_min;      // optional, NVIDIA only
+  char *cuda_max;      // optional, NVIDIA only
   gsize size_mb;
   char *requirements;
 } _ort_package_t;
@@ -61,6 +63,7 @@ static void _package_free(_ort_package_t *p)
   g_free(p->url); g_free(p->format); g_free(p->lib_pattern);
   g_free(p->install_subdir); g_free(p->sha256);
   g_free(p->rocm_min); g_free(p->rocm_max);
+  g_free(p->cuda_min); g_free(p->cuda_max);
   g_free(p->requirements); g_free(p);
 }
 
@@ -116,6 +119,10 @@ static GList *_load_manifest(void)
       pkg->rocm_min = g_strdup(json_object_get_string_member(p, "rocm_min"));
     if(json_object_has_member(p, "rocm_max"))
       pkg->rocm_max = g_strdup(json_object_get_string_member(p, "rocm_max"));
+    if(json_object_has_member(p, "cuda_min"))
+      pkg->cuda_min = g_strdup(json_object_get_string_member(p, "cuda_min"));
+    if(json_object_has_member(p, "cuda_max"))
+      pkg->cuda_max = g_strdup(json_object_get_string_member(p, "cuda_max"));
 
     result = g_list_append(result, pkg);
   }
@@ -135,6 +142,17 @@ static gchar *_version_major_minor(const char *version)
   return mm;
 }
 
+// numeric version compare: returns -1, 0, or 1
+static int _version_cmp(const char *a, const char *b)
+{
+  int ax = 0, ay = 0, bx = 0, by = 0;
+  if(a) sscanf(a, "%d.%d", &ax, &ay);
+  if(b) sscanf(b, "%d.%d", &bx, &by);
+  if(ax != bx) return ax < bx ? -1 : 1;
+  if(ay != by) return ay < by ? -1 : 1;
+  return 0;
+}
+
 // Find the best matching package for vendor + current platform + ROCm version
 static _ort_package_t *_find_package(GList *packages, dt_ort_gpu_vendor_t vendor)
 {
@@ -154,23 +172,52 @@ static _ort_package_t *_find_package(GList *packages, dt_ort_gpu_vendor_t vendor
     g_free(rocm_ver);
   }
 
+  // detect CUDA toolkit version via nvcc
+  gchar *cuda_mm = NULL;
+  if(vendor == DT_ORT_GPU_NVIDIA)
+  {
+    gchar *nvcc_out = NULL;
+    if(g_spawn_command_line_sync("nvcc --version", &nvcc_out,
+                                 NULL, NULL, NULL)
+       && nvcc_out)
+    {
+      // parse "Cuda compilation tools, release X.Y, VX.Y.Z"
+      gchar *v = g_strstr_len(nvcc_out, -1, ", V");
+      if(v)
+      {
+        v += 3; // skip ", V"
+        cuda_mm = _version_major_minor(v);
+      }
+      g_free(nvcc_out);
+    }
+  }
+
   _ort_package_t *best = NULL;
   for(GList *l = packages; l; l = g_list_next(l))
   {
     _ort_package_t *pkg = l->data;
     if(g_strcmp0(pkg->vendor, vendor_str) != 0) continue;
     if(g_strcmp0(pkg->platform, platform) != 0) continue;
 
-    // AMD: match ROCm version range (string compare works for single-digit majors)
+    // AMD: match ROCm version range
     if(rocm_mm && pkg->rocm_min)
     {
-      if(g_strcmp0(rocm_mm, pkg->rocm_min) < 0) continue;
-      if(pkg->rocm_max && g_strcmp0(rocm_mm, pkg->rocm_max) > 0) continue;
+      if(_version_cmp(rocm_mm, pkg->rocm_min) < 0) continue;
+      if(pkg->rocm_max && _version_cmp(rocm_mm, pkg->rocm_max) > 0) continue;
     }
+
+    // NVIDIA: match CUDA version range
+    if(cuda_mm && pkg->cuda_min)
+    {
+      if(_version_cmp(cuda_mm, pkg->cuda_min) < 0) continue;
+      if(pkg->cuda_max && _version_cmp(cuda_mm, pkg->cuda_max) > 0) continue;
+    }
+
     best = pkg;
     break;
   }
   g_free(rocm_mm);
+  g_free(cuda_mm);
   return best;
 }
 

tools/ai/install-ort-gpu.ps1

@@ -116,19 +116,36 @@ if ($Vendor) {
     }
 }
 
+# --- Detect CUDA version for NVIDIA package matching ---
+$CudaMM = ""
+if ($Selected.vendor -eq "nvidia") {
+    $nvcc = Get-Command nvcc -ErrorAction SilentlyContinue
+    if ($nvcc) {
+        $nvccOut = nvcc --version 2>$null | Select-String 'V(\d+\.\d+)' |
+            ForEach-Object { $_.Matches[0].Groups[1].Value }
+        if ($nvccOut) { $CudaMM = $nvccOut }
+    }
+}
+
 # --- Find matching package in manifest ---
 $Package = $null
 foreach ($p in $ManifestData.packages) {
     if ($p.vendor -ne $Selected.vendor) { continue }
     if ($p.platform -ne $Platform) { continue }
     if ($p.arch -and $p.arch -ne $Arch) { continue }
+    # NVIDIA: match CUDA version range
+    if ($CudaMM -and $p.cuda_min) {
+        if ([version]$CudaMM -lt [version]$p.cuda_min) { continue }
+        if ($p.cuda_max -and [version]$CudaMM -gt [version]$p.cuda_max) { continue }
+    }
     $Package = $p
     break
 }
 
 if (-not $Package) {
     Write-Host ""
     Write-Host "Error: no matching package found for $($Selected.vendor)/$Platform/$Arch" -ForegroundColor Red
+    if ($CudaMM) { Write-Host "  CUDA version: $CudaMM" }
     exit 1
 }
 

tools/ai/install-ort-gpu.sh

@@ -65,6 +65,7 @@ GPU_LABEL_AMD=""
 DRIVER_VERSION_AMD=""
 GPU_LABEL_INTEL=""
 ROCM_MM=""
+CUDA_MM=""
 
 if [ -n "$VENDOR_OVERRIDE" ]; then
   case "$VENDOR_OVERRIDE" in
@@ -75,11 +76,14 @@ if [ -n "$VENDOR_OVERRIDE" ]; then
       ;;
   esac
   echo "Vendor override: $VENDOR_OVERRIDE (skipping GPU detection)"
-  # still collect ROCm version for AMD package matching
+  # still collect version info for package matching
   if [ "$VENDOR_OVERRIDE" = "amd" ] && [ -f "/opt/rocm/.info/version" ]; then
     ROCM_VERSION=$(cat /opt/rocm/.info/version | tr -d '[:space:]')
     ROCM_MM=$(echo "$ROCM_VERSION" | grep -oP '^\d+\.\d+')
   fi
+  if [ "$VENDOR_OVERRIDE" = "nvidia" ] && command -v nvcc &>/dev/null; then
+    CUDA_MM=$(nvcc --version 2>/dev/null | grep -oP 'V\K\d+\.\d+' || true)
+  fi
 else
   # NVIDIA
   if command -v nvidia-smi &>/dev/null; then
@@ -88,6 +92,10 @@ else
       GPU_LABEL=$(echo "$NVIDIA_INFO" | cut -d, -f1 | xargs)
       DRIVER_VERSION=$(echo "$NVIDIA_INFO" | cut -d, -f2 | xargs)
       VENDOR="nvidia"
+      # detect CUDA toolkit version
+      if command -v nvcc &>/dev/null; then
+        CUDA_MM=$(nvcc --version 2>/dev/null | grep -oP 'V\K\d+\.\d+' || true)
+      fi
     fi
   fi
 
@@ -168,15 +176,18 @@ if ! command -v jq &>/dev/null; then
   exit 1
 fi
 
-PKG_JSON=$(jq -c --arg v "$SELECTED" --arg p "$PLATFORM" --arg a "$ARCH" --arg r "${ROCM_MM:-}" \
+PKG_JSON=$(jq -c --arg v "$SELECTED" --arg p "$PLATFORM" --arg a "$ARCH" \
+  --arg r "${ROCM_MM:-}" --arg c "${CUDA_MM:-}" \
   '[.packages[] | select(.vendor==$v and .platform==$p and (.arch // "x86_64")==$a) |
-   if ($r != "" and .rocm_min != null) then select(.rocm_min <= $r and (.rocm_max // "99") >= $r) else . end] | first' \
+   if ($r != "" and .rocm_min != null) then select(.rocm_min <= $r and (.rocm_max // "99") >= $r) else . end |
+   if ($c != "" and .cuda_min != null) then select(.cuda_min <= $c and (.cuda_max // "99") >= $c) else . end] | first' \
   "$MANIFEST" 2>/dev/null || true)
 
 if [ -z "$PKG_JSON" ]; then
   echo ""
   echo "Error: no matching package found in manifest for $SELECTED/$PLATFORM/$ARCH"
   [ -n "${ROCM_MM:-}" ] && echo "  ROCm version: $ROCM_MM"
+  [ -n "${CUDA_MM:-}" ] && echo "  CUDA version: $CUDA_MM"
   echo "  Check ort_gpu.json for supported configurations."
   echo ""
   exit 1