Merge pull request #670 from ROCm/jukorhon/atomic-counter

persistent kernel version of the flash attention forward + FLOP calculation fix when seqlen_q != seqlen_k

Author: juuso-oskari

python/perf-kernels/README.md

@@ -42,6 +42,7 @@ This script contains the Flash Attention kernel with the following support
 - Multi and Grouped Query attention
 - ALiBi bias
 - Matrix bias
+- Persistent kernels. Useful when the sequence lengths are up to a moderate length and especially when doing causal attention.
 - Int8 quantization
 
 These are currently supported for the forward kernel only.