pfnet

/

plamo-13b

like 81

Follow

Preferred Networks, Inc. 22

Hello, I would like to report an unintended behavior that I have encountered.

Summary:

In environments with transformers > 4.32.0, loading a Tokenizer results in an AttributeError: 'PlamoTokenizer' object has no attribute 'sp_model'. I have confirmed this issue in an environment with transformers 4.34.0, tokenizers 0.14.0, and sentencepiece 0.1.99. It is unclear whether this issue originates from the transformers library or from other dependent packages.

from transformers import AutoTokenizer, AutoModelForCausalLM
tokenizer = AutoTokenizer.from_pretrained("pfnet/plamo-13b", trust_remote_code=True) # raise error
# model = AutoModelForCausalLM.from_pretrained("pfnet/plamo-13b", trust_remote_code=True)

Strangely enough, text generation works without any errors with transformers version 4.34.0 when using the pipeline.

import transformers
pipeline = transformers.pipeline("text-generation", model="pfnet/plamo-13b", trust_remote_code=True)
print(pipeline("The future of artificial intelligence technology is ", max_new_tokens=32))

Workaround:

pip install transformers==4.32.0