SFT does not work max_steps

[AtsunoriFujita]

SFT API ignores trainer.sft.max_steps in gpt_sft.yaml.
Always refer to trainer.sft.max_epochs only.

Test case:

• trainer.sft.max_steps=200 and trainer.sft.max_epochs=-1
  　- 0 step job finished
• trainer.sft.max_steps=200 and trainer.sft.max_epochs=0
  　- 0 step job finished
• trainer.sft.max_steps=200 and trainer.sft.max_epochs=1 (187 steps)
  　- 187 steps job finished
• ~trainer.sft.max_epochs
  　- error

[odelalleau]

Thanks for reporting this. I think the main issue is that we don't support undefined max_epochs (-1 or None -- btw I would suggest to treat both the same, contrary to PTL, because I don't think there's much value in using a default number of 1000 epochs when using None as in PTL). This definitely needs to be fixed.

In the meantime, you can probably work around this by simply setting max_epochs to a large value, so that only max_steps is taken into account (note that the only bugged run in your examples is the first one -- the other three work as intended).

[AtsunoriFujita]

Sorry, my testing was insufficient. I thought this issue was a similar situation to NeMo (does not work max_epochs consistently).
Certainly, that workaround worked.