Extremely slow initialization and errors during test

[YanZehong]

Hi, thanks for releasing DGM4 and HAMMER!

I encountered two issues when initializing the HAMMER model.

1. Extremely slow initialization

When I run the following line:

model = HAMMER(args=args, config=config, text_encoder=args.text_encoder, tokenizer=tokenizer, init_deit=True)'

it takes around 30–40 minutes to finish before it reaches: checkpoint = torch.load(checkpoint_dir, map_location='cpu').

Note

Is this long initialization time expected? Has anyone else run into this, or is there something I might be misconfigured on my side?

2. Subsequent error

After that, I encounter the following error (see attached screenshot for the full stack trace):