You may find the thread at Strategy for randomizing large files via sysseek useful.

One thing that is not clear from your post is whether you want to actually shuffle the lines in your 4GB file, or just find a mechanism for allowing many processes to be given unique random subset of the lines from that file?

It may be that they amount to the same thing in that shuffling the large file is possibly the easiest way, but it may be possible to allow each process to select lines at random without overlaps more efficiently.

Also whether you would accept a "nearly random" distribution as opposed to statistically valid randomness , may allow a more efficient, slightly less than really random solution.