we write 0 to /sys/kernel/mm/hugepages/hugepages-${SIZE}/nr_hugepages which should release them after the vm is shutdown
of course you could be unlucky and the memory fragmentation does not allow for a 1gb hugepage to be allocated again

sub hugepages_default_nr_hugepages {
    my ($size) = @_;

    my $cmdline = PVE::Tools::file_read_firstline("/proc/cmdline");
    my $args = PVE::Tools::split_args($cmdline);

    my $parsed_size = 2; # default is 2M

    foreach my $arg (@$args) {
        if ($arg eq "hugepagesz=2M") {
            $parsed_size = 2;
        } elsif ($arg eq "hugepagesz=1G") {
            $parsed_size = 1024;
        } elsif ($arg =~ m/^hugepages=(\d+)?$/) {
            if ($parsed_size == $size) {
                return $1;
            }
        }
    }

    return 0;
}

sub hugepages_pre_deallocate {
    my ($hugepages_topology) = @_;

    foreach my $size (sort keys %$hugepages_topology) {

        my $hugepages_size = $size * 1024;
        my $path = "/sys/kernel/mm/hugepages/hugepages-${hugepages_size}kB/";
        my $hugepages_nr = hugepages_default_nr_hugepages($size);
        PVE::ProcFSTools::write_proc_entry($path."nr_hugepages", $hugepages_nr);
    }
}