Very slow at the end (>=1.30)

I have a problem:
ocl64 and cuda64 v.130 and 1.31

Sometimes for attack methods Combinator and Hybryd
(1 dict smaller then 2-nd or mask keyspace; MD5, NTLM about 100 hashes)
(may be another too)
at the end of work (after ~96-97%)
very power decrease of speed (last 1% especially).
Fore example sometimes i have time of 95% work smaller then
time of last 5 % work.
(and 1.01 total time very smaller then total time 1.31)


At listing i have this string:
INFO: approaching final keyspace, workload adjusted

I used all combinations of workload_profile, gpu_accel and gpu_loops.

Can I deside this bug?
Can I block "workload adjusted" ?