Karina Zadorozhny

Issue #91 Example usage ```py from lobster.callbacks import RandomNeighborScoreCallback from lobster.model import Ume callback = RandomNeighborScoreCallback( biological_dataset_name="AMPLIFY", k=3, biological_dataset_limit=500, split="test" ) model = Ume(model_name="UME_mini", use_flash_attn=False) metrics = callback.evaluate(model, trainer=None) rns_score...

When running ESM models on CPU with flash_attn installed, inference fails with CUDA-related errors despite explicitly setting the device to CPU. **Steps to reproduce:** On a CPU machine with flash-attention...

multi-gpu training with torchdata nodes

2

What's the best way to handle multi-gpu and multi-node training with torchdata nodes? @keunwoochoi and I have the following nodes: * `FileListNode` which lists all files (shards) * `StreamNode` which...

Supporting samplers for creating balanced batches

1

Hey everyone, We're working with a dataset where items belong to different clusters. Example: item_1, cluster_A item_2, cluster_A item_3, cluster_B I was wondering if there's a way to create batches...

https://github.com/oxpig/ANARCI/blob/master/LICENCE

## Description Brief description of changes made ## Type of Change - [ ] Bug fix - [ ] New feature - [ ] Documentation update - [ ] Performance...

Add checksum verification for UME checkpoint caching

1

Currently, `UME.from_pretrained()` only checks if there is a local version of a checkpoint present. But it doesn't take into account that the checkpoints could have been updated already.