yuze.zyz
|
a0bc5549a1
|
trainer support parallel_groups
Design doc: https://yuque.alibaba-inc.com/suluyan.sly/yh1rvu/yx0owblyebpa2b3l?singleDoc#flU3s
1. Add parallel_group field in trainer to support DP, TP, PP.
2. Move the construction of common hooks(except optimizer/lrscheduler hook) to trainer's init method to support after_init stage.
after_init is to support DP, TP, PP's initializing
https://aone.alibaba-inc.com/v2/workitem#viewIdentifier=1c46ee8637e0c978f115b6f7&openWorkitemIdentifier=48099986
3. Add before_eval/after_eval stage to support model wrapping.
to solve the order problem of apex amp & ddp wrapping.
https://aone.alibaba-inc.com/v2/workitem#viewIdentifier=1c46ee8637e0c978f115b6f7&openWorkitemIdentifier=48099986
4. Exporter supports lazy importing.
https://aone.alibaba-inc.com/v2/workitem#viewIdentifier=1c46ee8637e0c978f115b6f7&openWorkitemIdentifier=48122780
5. Fold all megatron imports to megatron hook.
https://aone.alibaba-inc.com/v2/workitem#viewIdentifier=1c46ee8637e0c978f115b6f7&openWorkitemIdentifier=48099986
6. Add compile method to TorchModel ,Pipeline,Trainer to support torch2.0
https://aone.alibaba-inc.com/v2/workitem#viewIdentifier=1c46ee8637e0c978f115b6f7&openWorkitemIdentifier=46869415
7. Fix bug: Lrscheduler builder does not support torch2.0
8. Add callbacks for trainer
https://aone.alibaba-inc.com/v2/workitem#viewIdentifier=1c46ee8637e0c978f115b6f7&openWorkitemIdentifier=48210342
Link: https://code.alibaba-inc.com/Ali-MaaS/MaaS-lib/codereview/11849932
|
2023-03-09 21:33:35 +08:00 |
|