bloom-multilingual-chat / load_model.py
AstraBert
temp
df7e912
raw
history blame
343 Bytes
from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline
model = AutoModelForCausalLM.from_pretrained("bigscience/bloom-1b7")
tokenizer = AutoTokenizer.from_pretrained("bigscience/bloom-1b7")
pipe = pipeline("text-generation", model=model, tokenizer=tokenizer, max_new_tokens=2048, repetition_penalty=1.2, temperature=0.4)