The reason for introducing the RPKM and other measures was to account for the fact that long transcripts produce more sequencing reads at the same abundance levels.
For miRNA a single read will cover an entire miRNA so there is no need for length based normalization methods.
Of course whether or not simple counting and normalizing by coverage (and how that should be done) is source of a lot of disagreement.
As you normalize, be aware of high duplication rates and serious microRNA:adapter preferences that really introduce significant bias:
Jayaprakash, A. D., Jabado, O., Brown, B. D., & Sachidanandam, R. (2011). Identification and remediation of biases in the activity of RNA ligases in small-RNA deep sequencing.
Sorefan, K., Pais, H., Hall, A. E., Kozomara, A., Griffiths-Jones, S., Moulton, V., & Dalmay, T. (2012). Reducing ligation bias of small RNAs in libraries for next generation sequencing.
Zhang, Z., Lee, J. E., Riemondy, K., Anderson, E. M., & Yi, R. (2013). High-efficiency RNA cloning enables accurate quantification of miRNA expression by deep sequencing.
You should normalize. Each sample has a different sequencing depth. The workflow I suggest :
htseq-count using miRNA gtf on each bam file. It gives you an expression matrix
DESeq or edgeR for normalization and DE analysis