[merger] ci: add FSDP checkpoint merging test into CI#1266
[merger] ci: add FSDP checkpoint merging test into CI#12660x404 wants to merge 3 commits intoverl-project:mainfrom
Conversation
|
Hi @0x404 , thanks a lot for contribution! The testing option is mainly work for CI tests, could you help to add some testing cases in e2e_ppo_trainer.yml? You may need to store the checkpoint first for converter to compare, add some option to Your great contribution will be appreciated~ |
|
Sure! I will update this latter. |
|
This PR depends on #1288. Hi @ETOgaosion, could you help review these two changes? |
|
Appreciate a lot for contribution, please check my comments in #1288 , let's work together to merge that PR first. |
Sure, but I don't see a comment there yet. |
|
@0x404 Could you see these? And the reviewing content? |
I can see this comment #1288 (comment)
But I can't see the review comments, maybe you forgot to submit them? |
|
Sorry, my fault, @0x404 , actually submit button is needed. And I think that why can't we just add tests to that PR directly and test its function? So we can debug directly and save our time. Actually it would be better to test every function proposed in PR. So it would be better to merge this one to #1288 ~ |
|
No problem, I will merge this pr to it and close this latter tonight. |
|
Move to #1288 and close this. |
Currently
--testoption is only supported for merging megatron checkpoints, this PR support test compatibility for FSDP checkpoints.