Jasper Lu

Upsert does not update SequenceField

1

Is this intended behavior? I ran into a bug where when I called `update_one` with`upsert=True`, a sequencefield was not created for the object on creation.

When I enter in text such as "Tomorrow evening," it seems to only be parsing the tomorrow, and not the evening part. Can you check this?

Why does torchtune FSDP have a lower memory footprint than accelerate?

3

I am trying to fine-tune Qwen2.5-32B using huggingface transformers with FSDP and a max context length of 20k on 4xA100-80GB. I'm hitting OOM even when I use all optimizations. However,...