Indexing: handle cpu & single-gpu without using multiprocessing & dist. data parallel #290
Conversation
Anmol6
changed the title
|
I think after the latest batch of fixes it looks good now! Thank you so much for doing this! |
colbert/infra/config/settings.py
Outdated
| @@ -30,6 +30,8 @@ class RunSettings: | |||
| total_visible_gpus = torch.cuda.device_count() | |||
| gpus: int = DefaultVal(total_visible_gpus) | |||
|
|
|||
| use_rank1_fork: bool = DefaultVal(False) | |||
There was a problem hiding this comment.
Just a quick note (well, two):
- Should this not be the opposite logic? i.e. defaults to
True(and the associated logic changes in the code)? I feel like it'd make more sense with the naming - Do we want this to be a
configitem rather than a flag passed to the function? I think either works fine and maybe having it as part of config is more fitting for how the rest of the code is structured.
There was a problem hiding this comment.
I was using fork in the sense of a code fork but will revert to fork as in multiprocessing! thanks for explaining the terminology!
There was a problem hiding this comment.
Thanks, all looking good to me now! Just needs Omar's final approval 👍
|
This fixed a problem I was having downstream in RAGatouille, always running in distributed mode on wsl2 with a single gpu. Thank you for the PR! bclavie/RAGatouille#30 (comment) edit: looks like the trainer is still always forcing distributed torch. the collection indexer fixed indexing, though. Definitely the right direction though. |
No description provided.