llama3-8b-instruct 4bit quantize

main.py

from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig
import torch
from time import time
model_id = "meta-llama/Meta-Llama-3-8B-Instruct"

config = BitsAndBytesConfig(load_in_4bit=True,
    llm_int8_enable_fp32_cpu_offload=True,
    bnb_4bit_compute_dtype = torch.float16) 

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    quantization_config=config,
    # torch_dtype=torch.bfloat16,
    # max_memory={i: '6000MB' for i in range(torch.cuda.device_count())},
    device_map="auto",
)

messages = [
    {"role": "system", "content": "You are a helpful decision aid that makes a decision if a scene is a marina, large dock, or airport"},
    {"role": "user", "content": "I have an image that has 20 small boats and no large vessels in it. There is also a small biplane in the water. what type of scene is this image most likely describing"},
]
start = time()
input_ids = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    return_tensors="pt"
).to(model.device)

terminators = [
    tokenizer.eos_token_id,
    tokenizer.convert_tokens_to_ids("<|eot_id|>")
]

outputs = model.generate(
    input_ids.to(torch.float32).long(),
    max_new_tokens=256,
    eos_token_id=terminators,
    do_sample=True,
    temperature=0.6,
    top_p=0.9,
)
response = outputs[0][input_ids.shape[-1]:]
print(tokenizer.decode(response, skip_special_tokens=True))

print(time() - start)