Open QPHutu opened 8 months ago
The paper been accepted by ICLR 2024.
The key idea is to split the backward computation into two parts, one that computes gradient for the input and another that computes for the parameters. By rescheduling the parameters' gradient computation, we can have get a better efficiency without scrificing anything.
May I ask what led you to commit to this repository over the original one? Just curious about your thoughts! @QPHutu
Thanks for the reply. There are 2 main reasons.
thanks for the PR!
for merging we'd like to understand the impact a bit better. did you verify how model parallel training of the current models supported here (such as llama2) is impacted by your change? (in terms of speed, stability and also verify model behavior is unchanged?)
indeed could be nice to also hear the feedback from the Nvidia/Megatron-LM team if you get a chance
The change is a quick implementation to replace 1F1B with ZB-H1 proposed in Zero Bubble Pipeline Parallelism, which reduces the bubbles in pipeline parallelism.