In the training loop, an accumulated loss is computing in an additive manner here across all steps in the num_accumulation_rounds. This lose should be divided by the value of num_accumulation_rounds to get the mean rather than the sum of loss to be recorded to wandb:
loss_accum += loss/num_accumulation_rounds
Version
0.5.0
On which installation method(s) does this occur?
Pip
Describe the issue
In the training loop, an accumulated loss is computing in an additive manner here across all steps in the
num_accumulation_rounds
. This lose should be divided by the value ofnum_accumulation_rounds
to get the mean rather than the sum of loss to be recorded to wandb:loss_accum += loss/num_accumulation_rounds
Minimum reproducible example
No response
Relevant log output
No response
Environment details
No response