The good thing is that it seems it can recover, so the steps after converge ok and the calculation seems to proceed OK. However, looking into the convergence might speed it up quite a bit, so I would take out the geometry of one of those "weird" steps, run it separately, and see if you can make it converge better/faster by tuning some parameters; maybe you need more k-points, higher temperature, different mixing. Even if more k-points is slower nominally, if it uses considerably fewer SCF steps it's faster overall anyway (and more accurate).