accelerate + fabric

[hvgazula]

determine batch size automatically

[hvgazula]

https://lightning.ai/docs/pytorch/stable/advanced/training_tricks.html#batch-size-finder

one idea is to run this on one gpu..get the batch size and then initialize the fabric and all that…

[hvgazula]

the above function only works with models created with LightningModule 🤷‍♂️ . Another option is the snippet in https://towardsdatascience.com/a-batch-too-large-finding-the-batch-size-that-fits-on-gpus-aef70902a9f1 that can be coupled with trainer object.

[hvgazula]

closing this as fabric's features are limited at this point.