You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
This may be related to #9675. We can reduce the copy of split op and limit the max size of each parameter/gradient block, and send those blocks using an "IOthreadpool" ( which should contain much more threads than machine CPU cores).
And, in earlier tests result #7944, we see that split copying is not the major time-consuming operator, so we delayed this improvement.
max_block_size is not used in this function:
Paddle/python/paddle/fluid/distribute_transpiler.py
Lines 91 to 95 in f53beed
The text was updated successfully, but these errors were encountered: