In reply to @hexa:lossy.network
tf-cpu also builds on aarch64-darwin

well it mat not be pretty, buth this deos give me what I need:

          cudaPackages = pkgs.cudaPackages_11_4;

          tensorrt = cudaPackages.tensorrt.overrideAttrs (old: let
            fullVersion = "8.2.1.8";
            fileVersionCuda = "11.4";
            fileVersionCudnn = "8.2";
            sha256 = "sha256-PpqcxK0OWuY3MX2STc3fZjgfTbBOJXHw8ubtWipR8kc=";
          in rec {
            pname = "cudatoolkit-${cudaPackages.cudatoolkit.majorVersion}-tensorrt";
            version = fullVersion;
            src = pkgs.requireFile rec {
              name = "TensorRT-${fullVersion}.Linux.x86_64-gnu.cuda-${fileVersionCuda}.cudnn${fileVersionCudnn}.tar.gz";
              inherit sha256;
              message = ''
                To use the TensorRT derivation, you must join the NVIDIA Developer Program and
                download the ${version} Linux x86_64 TAR package for CUDA ${cudaPackages.cudaVersion} from
                ${old.meta.homepage}.

                Once you have downloaded the file, add it to the store with the following
                command, and try building this derivation again.

                $ nix-store --add-fixed sha256 ${name}
              '';
            };

            sourceRoot = "TensorRT-${version}";

            # Tell autoPatchelf about runtime dependencies.
            # (postFixup phase is run before autoPatchelfHook.)
            postFixup =
              let
                mostOfVersion = builtins.concatStringsSep "."
                  (pkgs.lib.take 3 (pkgs.lib.versions.splitVersion version));
              in
              ''
                echo 'Patching RPATH of libnvinfer libs'
                patchelf --debug --add-needed libnvinfer.so \
                  "$out/lib/libnvinfer.so.${mostOfVersion}" \
                  "$out/lib/libnvinfer_plugin.so.${mostOfVersion}" \
                  "$out/lib/libnvinfer_builder_resource.so.${mostOfVersion}"
              '';
          });