Carlos Mocholí
Carlos Mocholí
Do you see any specific differences in the modeling?
It's not clear how to make this more robust. Perhaps the best way is to drop support for `python litgpt/finetune/lora.py` since we no longer advertise it
If you use `ddp_fork`, that will not shard your model. If you want this, I suggest that you give up using a notebook for training as it cannot support FSDP....
I'm sorry but I have no idea about what you are talking about :(
@awaelchli Improved the error in #1241. Still, we could set a default fraction
I believe this is not currently supported by looking at the piece of code that checks satisfiability https://github.com/tianhaoz95/check-group/blob/df61154e69ffd9d54a43207781839ce24a6867db/src/utils/satisfy_expected_checks.ts#L33 as there's no regex lookup
You can do `Trainer(reload_dataloaders_every_n_epochs=1)` to accomplish this
This is "working as expected" given the current design of `setup_data`, which doesn't run if the data is already setup and the trainer flag is not configured, see this early...
I also have this problem in https://github.com/Lightning-AI/lightning/pull/15043
My hunch here is that pre 2.0 it was using 16-bit precision by default and 2.0 is using 32 bit. You should be able to verify this by manually setting...