blip3o/model/builder.py

import os
import warnings
import shutil

from transformers import AutoTokenizer, AutoModelForCausalLM, AutoConfig, BitsAndBytesConfig
import torch
from blip3o.model import *
from blip3o.constants import DEFAULT_IMAGE_PATCH_TOKEN, DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN
from blip3o.train.train import smart_tokenizer_and_embedding_resize


def load_pretrained_model(model_path, model_base, model_name, load_8bit=False, load_4bit=False, device_map="auto", device="cuda", use_flash_attn=False, **kwargs):
    kwargs = {"device_map": device_map, **kwargs}

    if device != "cuda":
        kwargs['device_map'] = {"": device}

    if load_8bit:
        kwargs['load_in_8bit'] = True
    elif load_4bit:
        kwargs['load_in_4bit'] = True
        kwargs['quantization_config'] = BitsAndBytesConfig(
            load_in_4bit=True,
            bnb_4bit_compute_dtype=torch.float16,
            bnb_4bit_use_double_quant=True,
            bnb_4bit_quant_type='nf4'
        )
    else:
        kwargs['torch_dtype'] = torch.float16

    if use_flash_attn:
        kwargs['attn_implementation'] = 'flash_attention_2'


    tokenizer = AutoTokenizer.from_pretrained(model_path, use_fast=False)

    model = blip3oQwenForCausalLM.from_pretrained(model_path, low_cpu_mem_usage=True, torch_dtype=torch.float16).to('cuda:0')

    image_processor = None

    mm_use_im_start_end = getattr(model.config, "mm_use_im_start_end", False)
    mm_use_im_patch_token = getattr(model.config, "mm_use_im_patch_token", True)
    if mm_use_im_patch_token:
        tokenizer.add_tokens([DEFAULT_IMAGE_PATCH_TOKEN], special_tokens=True)
    if mm_use_im_start_end:
        tokenizer.add_tokens([DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN], special_tokens=True)
    model.resize_token_embeddings(len(tokenizer))

    if hasattr(model.config, "max_sequence_length"):
        context_len = model.config.max_sequence_length
    else:
        context_len = 2048

    return tokenizer, model, context_len