for everyone downvoting my comment

An “epoch” is one full pass through your training dataset. The number of optimization steps in one epoch is simply:

steps_per_epoch  =  ⌈dataset_sizebatch_size⌉\text{steps\_per\_epoch} \;=\;\Big\lceil \frac{\text{dataset\_size}}{\text{batch\_size}}\Big\rceil

— where

• dataset_size is the total number of training examples (or total number of tokens, if you’re counting in tokens),
• batch_size is the number of examples (or tokens) processed at each step.

If you’re using gradient‐accumulation over NN mini‐batches to form an effective batch, then:

steps_per_epoch=  ⌈dataset_sizebatch_size×N⌉\text{steps\_per\_epoch} =\; \Big\lceil \frac{\text{dataset\_size}} {\text{batch\_size}\times N} \Big\rceil

For example, 100 000 examples with a per‐device batch size of 32 (and no accumulation) gives

100 000/32≈3125 steps per epoch.100\,000 / 32 \approx 3125\text{ steps per epoch.}