does not have support for mistral, gemma, etc and generate error [BUG] ?

[NickyDark1]

model_id = "h2oai/h2o-danube-1.8b-chat"#

Upvote & Fund

• We're using Polar.sh so you can upvote and help fund this issue.
• We receive the funding once the issue is completed & confirmed by you.
• Thank you in advance for helping prioritize & fund our backlog.

[NickyDark1]

version: 4.36.2 new -> transformers==4.38.0 (no support)

[NickyDark1]

only support this model?

Load a model from Hugging Face's Transformers

model_name = "bert-base-uncased"

[NickyDark1]

no support:

• cuda()
• to("cuda:0")

[sanjeev-bhandari]

@NickyDark1, I ran that model in colab and it work

Without quanitizing

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("h2oai/h2o-danube-1.8b-chat")
model = AutoModelForCausalLM.from_pretrained("h2oai/h2o-danube-1.8b-chat")

# from transformers import pipeline

pipe = pipeline("text2text-generation", model=model, tokenizer=tokenizer)
pipe("Hello, How")

Output:

[{'generated_text': 'Hello, How are you?\n\n"I\'m doing well, thank you. How about'}]

After replacing Linear layer with bitnet

from bitnet import replace_linears_in_hf

replace_linears_in_hf(model)
# change model back to device cuda
model.to("cuda")
pipe_1_bit = pipeline("text-generation", model=model, tokenizer=tokenizer)
pipe_1_bit("Hello, How")

Output is:

[{'generated_text': 'Hello, How島 waters everyoneürgen Mess till revel馬 Vitt officials ambos">< czł plusieurs ap riv居'}]

But it takes ages to give this answer(8 min in my case in free colab).

[github-actions[bot]]

Stale issue message