llms.py

from dotenv import load_dotenv
import logging

import torch
from transformers import (
    AutoModelForCausalLM,
    AutoTokenizer,
    BitsAndBytesConfig,
    AutoTokenizer,
    BitsAndBytesConfig,
    pipeline,
)

from langchain_huggingface import HuggingFacePipeline
from langchain.globals import set_debug
from langchain.globals import set_verbose

from config import HF_MODEL_ID
from config import LLM_VERBOSE

set_verbose(LLM_VERBOSE)
set_debug(LLM_VERBOSE)

logger = logging.getLogger(__name__)
load_dotenv()

cuda_check = torch.cuda.is_available()
logger.info(f"torch.cuda.is_available : {cuda_check}")
print(f"> torch.cuda.is_available : {cuda_check}")

# Load Llama3 model and tokenizer
model_id = HF_MODEL_ID

tokenizer = AutoTokenizer.from_pretrained(model_id)

# BitsAndBytesConfig int-4 config
# device_map = {"": 0}
device_map = "auto"
compute_dtype = getattr(torch, "float16")
bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_use_double_quant=False,
    bnb_4bit_quant_type="nf4",
    # bnb_4bit_compute_dtype=torch.bfloat16
    bnb_4bit_compute_dtype=compute_dtype,
    # bnb_4bit_use_double_quant=False,
)

model = AutoModelForCausalLM.from_pretrained(
    model_id,
    device_map=device_map,
    # attn_implementation="flash_attention_2",
    quantization_config=bnb_config,
)

model.generation_config.pad_token_id = tokenizer.eos_token_id

pipe = pipeline(
    "text-generation", 
    model=model, 
    tokenizer=tokenizer, 
    max_new_tokens=50,
    return_full_text=False,
    num_return_sequences=1,
    eos_token_id=tokenizer.eos_token_id,
    temperature=0.0001,
    do_sample=True,
)

llm = HuggingFacePipeline(pipeline=pipe)