Stancld

/

longt5-tglobal-large-16384-pubmed-3k_steps

like 20

hey, thanks for the great upload! my question is related to model inference.

I was able to train using the trainer API on Colab, but in trying to run some inference in a notebook on non-dataset text, I get the following error even with LongT5ForConditionalGeneration:

TypeError: forward() got an unexpected keyword argument 'global_attention_mask'

(perhaps unrelated, but that same notebook works fine with this checkpoint for example). I then turned to your example in the model card to see if I could replicate that and I am doing something wrong, but I can't find where you define global_attention_mask in the example:

import torch
from transformers import AutoTokenizer, LongT5ForConditionalGeneration

tokenizer = AutoTokenizer.from_pretrained("Stancld/longt5-tglobal-large-16384-pubmed-3k_steps")

input_ids = tokenizer(LONG_ARTICLE, return_tensors="pt").input_ids.to("cuda")

model = LongT5ForConditionalGeneration.from_pretrained("Stancld/longt5-tglobal-large-16384-pubmed-3k_steps", return_dict_in_generate=True).to("cuda")

sequences = model.generate(input_ids, global_attention_mask=global_attention_mask).sequences

summary = tokenizer.batch_decode(sequences)

any help would be appreciated :)