With the BBMap package:
dedupe.sh in=reads.fq out=nodupes.fq
The output will contain exactly 1 copy of every unique sequence. It's extremely fast, but may take more memory than other solutions - the amount of memory is proportional to the number of unique reads (rather than, say, the total input size).
PRINSEQ should solve your problem. Check it out here: http://prinseq.sourceforge.net/manual.html#QCDUPLICATION
A bit of digging should get you the command line options for the feature.