[Perf] Conditionally enable SWAP AB for speculative decoding #5403
Description
It is a common scenario in speculative decoding for the draft model to handle a small number of tokens. When the value of q_seq_len * head_group_size is small, enabling SWAP AB provides a considerable performance improvement.
The performance results on H20 are shown in the chart below.
