pytscope — run  (runs/exp2_good)

No steps recorded.

Communication overlap (from kernel trace):
  comm 177.2 ms · overlapped 12% · exposed 155.4 ms (62% of wall)
  by collective: all_reduce 177.2ms

Findings (1):
  [HIGH] Communication is not overlapped with compute  (DIST.EXPOSED_COMM)
        62% of wall time is *exposed* communication — collective time that does not overlap any compute kernel. Only 12% of the 177.2 ms of communication is hidden behind compute.
        -> Improve compute/communication overlap: enable DDP gradient bucketing (tune `bucket_cap_mb`), avoid `find_unused_parameters` stalls, overlap optimizer/all-reduce, or increase per-GPU compute so backward lasts long enough to hide the all-reduce.
