Add the Torch versions to build for and note.

flake.lock

@@ -98,11 +98,11 @@
         ]
       },
       "locked": {
-        "lastModified": 1750860635,
-        "narHash": "sha256-Wzz7UQRuzSAQCvmPclYN/xmAqpMR19QDgmsCzknAzeo=",
+        "lastModified": 1751014803,
+        "narHash": "sha256-9Xfq2k3uPfB602NwQF+zAY2GQZiKUN1G7Q6XiDCUR8Y=",
         "owner": "huggingface",
         "repo": "kernel-builder",
-        "rev": "7e2d7303d9f2d49d470bddcbb8379fffeee38dcf",
+        "rev": "bbc4e712ff2046e217818e97de2201e2b996756e",
         "type": "github"
       },
       "original": {

flake.nix

@@ -13,5 +13,39 @@
     kernel-builder.lib.genFlakeOutputs {
       path = ./.;
       rev = self.shortRev or self.dirtyShortRev or self.lastModifiedDate;
+      # Building with CDUA later than 12.4 fails with:
+      #
+      # error: 'ptxas' died due to signal 11 (Invalid memory reference)
+      #
+      # So, build for 12.4 only and copy to all the other build variants
+      # by hand (which works fine thanks to backward compat).
+      #
+      # Still need to check if upstream FA3 has the same issue.
+      torchVersions = [
+        {
+          torchVersion = "2.6";
+          cudaVersion = "12.4";
+          cxx11Abi = false;
+          systems = [ "x86_64-linux" ];
+          upstreamVariant = true;
+        }
+        {
+          torchVersion = "2.6";
+          cudaVersion = "12.4";
+          cxx11Abi = true;
+          systems = [ "x86_64-linux" ];
+          upstreamVariant = true;
+        }
+        {
+          torchVersion = "2.7";
+          cudaVersion = "12.4";
+          cxx11Abi = true;
+          systems = [
+            "x86_64-linux"
+            "aarch64-linux"
+          ];
+          upstreamVariant = true;
+        }
+      ];
     };
 }