Zero Bubble Schedules
The key of achieving zero bubble is to breaking a backward pass into a B pass and W pass. B on one stage will only depend on the B on its next stage, compared to depending on both B and W of in 1F1B.
Comparision of Schedules
- 1F1B
- ZB1P
- ZB2P
- ZBV - Each device is assigned to exactly 2 chunks (virtual stages), where white text colors represent the first chunk and black text colors represent the second chunk. The sequence of dependencies among model chunks follows a ”V” shape pattern for both the forward and backward passes.
| Comparison assuming T_F=T_B=T_W | 1F1B | ZB1P | ZB2P | ZBV (Recommended) |
|---|---|---|---|---|
| Bubble Rate | (p-1)/m | (p-1)/3m | 0 | 0 |
| Activation Memory (Compared to 1F1B) |
1x | 1x | 2x | 1x |
| Pipeline Communication Volume (Compared to 1F1B) |
1x | 1x | 1x | 2x |
Optimizer Post Validation
In most practices of PP there's an all-reduce cross all pipeline stages for numerical robustness, e.g. global gradient norm for gradient clipping. INF/NAN check for mixed precision training, etc. This all-reduce breaks parallelogram and makes zero bubble impossible. Under the observation that during a stable training both the gradient clipping and INF/NAN rarely triggers, we replace the before-hand synchronizations with a post update validation.
We eagerly step the optimizers assuming the grad cliping, INF/NAN conditions are not triggered. In case an amendment to the gradient is required, a rollback will be issued and then we redo the optimizer step based on the fully reduced global state.