from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig
from peft import PeftModel
import torch

tokenizer = AutoTokenizer.from_pretrained("oudmaria/darija-translator")
bnb_config = BitsAndBytesConfig(load_in_4bit=True, bnb_4bit_compute_dtype=torch.bfloat16)
base = AutoModelForCausalLM.from_pretrained("unsloth/Llama-3.2-3B-Instruct", quantization_config=bnb_config, device_map="auto")
model = PeftModel.from_pretrained(base, "oudmaria/darija-translator")

messages = [
    {"role": "system", "content": "You are a translation assistant. Translate the Moroccan Darija text to English. Reply with ONLY the English translation, nothing else."},
    {"role": "user", "content": "Translate: labas 3lik?"},
]
inputs = tokenizer.apply_chat_template(messages, return_tensors="pt", add_generation_prompt=True).to("cuda")
out = model.generate(input_ids=inputs, max_new_tokens=64)
print(tokenizer.decode(out[0][inputs.shape[1]:], skip_special_tokens=True))
# → are you okay

Framework versions

PEFT 0.19.1