inference/model/builder.py

import os
import warnings

from transformers import (
    AutoTokenizer,
    AutoModelForCausalLM,
    AutoConfig,
    BitsAndBytesConfig,
    CLIPImageProcessor,
)
import torch
from .language_model.llava_phi import LlavaPhiForCausalLM
from .language_model.configuration_llava_phi import LlavaPhiConfig

IGNORE_INDEX = -100
IMAGE_TOKEN_INDEX = -200
DEFAULT_IMAGE_TOKEN = "<image>"
DEFAULT_IMAGE_PATCH_TOKEN = "<im_patch>"
DEFAULT_IM_START_TOKEN = "<im_start>"
DEFAULT_IM_END_TOKEN = "<im_end>"


def load_pretrained_model(
    model_path,
    model_base,
    model_name,
    load_8bit=False,
    load_4bit=False,
    device_map="cuda",
    device="cuda",
):
    kwargs = {"device_map": device_map}
    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:  # TODO: after fine-tuning LLava-Phi, load the model weights with fp16 will pose nan
    #     kwargs['torch_dtype'] = torch.float16

    if "phi" in model_name.lower():
        # Load LLaVA-Phi model
        if "lora" in model_name.lower() and model_base is None:
            warnings.warn(
                "There is `lora` in model name but no `model_base` is provided. If you are loading a LoRA model, please provide the `model_base` argument."
            )
        if "lora" in model_name.lower() and model_base is not None:
            lora_cfg_pretrained = AutoConfig.from_pretrained(model_path)
            tokenizer = AutoTokenizer.from_pretrained(model_base, use_fast=False)
            print("Loading LLaVA-Phi from base model...")
            model = LlavaPhiForCausalLM.from_pretrained(
                model_base, low_cpu_mem_usage=True, config=lora_cfg_pretrained, **kwargs
            )
            token_num, tokem_dim = model.lm_head.out_features, model.lm_head.in_features
            if model.lm_head.weight.shape[0] != token_num:
                model.lm_head.weight = torch.nn.Parameter(
                    torch.empty(
                        token_num, tokem_dim, device=model.device, dtype=model.dtype
                    )
                )
                model.model.embed_tokens.weight = torch.nn.Parameter(
                    torch.empty(
                        token_num, tokem_dim, device=model.device, dtype=model.dtype
                    )
                )

            print("Loading additional LLaVA-Phi weights...")
            if os.path.exists(os.path.join(model_path, "non_lora_trainables.bin")):
                non_lora_trainables = torch.load(
                    os.path.join(model_path, "non_lora_trainables.bin"),
                    map_location="cpu",
                )
            else:
                # this is probably from HF Hub
                from huggingface_hub import hf_hub_download

                def load_from_hf(repo_id, filename, subfolder=None):
                    cache_file = hf_hub_download(
                        repo_id=repo_id, filename=filename, subfolder=subfolder
                    )
                    return torch.load(cache_file, map_location="cpu")

                non_lora_trainables = load_from_hf(
                    model_path, "non_lora_trainables.bin"
                )
            non_lora_trainables = {
                (k[11:] if k.startswith("base_model.") else k): v
                for k, v in non_lora_trainables.items()
            }
            if any(k.startswith("model.model.") for k in non_lora_trainables):
                non_lora_trainables = {
                    (k[6:] if k.startswith("model.") else k): v
                    for k, v in non_lora_trainables.items()
                }
            model.load_state_dict(non_lora_trainables, strict=False)

            from peft import PeftModel

            print("Loading LoRA weights...")
            model = PeftModel.from_pretrained(model, model_path)
            print("Merging LoRA weights...")
            model = model.merge_and_unload()
            print("Model is loaded...")
        elif model_base is not None:
            # this may be mm projector only
            print("Loading LLaVA-Phi from base model...")
            tokenizer = AutoTokenizer.from_pretrained(model_base, use_fast=False)
            cfg_pretrained = AutoConfig.from_pretrained(model_path)
            model = LlavaPhiForCausalLM.from_pretrained(
                model_base, low_cpu_mem_usage=True, config=cfg_pretrained, **kwargs
            )

            mm_projector_weights = torch.load(
                os.path.join(model_path, "mm_projector.bin"), map_location="cpu"
            )
            mm_projector_weights = {
                k: v.to(torch.float16) for k, v in mm_projector_weights.items()
            }
            model.load_state_dict(mm_projector_weights, strict=False)
        else:
            print("load llaVA-Phi MLLM!!!")
            config = LlavaPhiConfig.from_pretrained(model_path, trust_remote_code=True)
            tokenizer = AutoTokenizer.from_pretrained(model_path, use_fast=True)
            model = LlavaPhiForCausalLM.from_pretrained(
                model_path, config=config, use_safetensors=True, **kwargs
            ).to("cuda")
    else:
        # Load language model
        if model_base is not None:
            # PEFT model
            from peft import PeftModel

            tokenizer = AutoTokenizer.from_pretrained(model_base, use_fast=False)
            model = AutoModelForCausalLM.from_pretrained(
                model_base,
                torch_dtype=torch.float16,
                low_cpu_mem_usage=True,
                device_map="auto",
            )
            print(f"Loading LoRA weights from {model_path}")
            model = PeftModel.from_pretrained(model, model_path)
            print(f"Merging weights")
            model = model.merge_and_unload()
            print("Convert to FP16...")
            model.to(torch.float16)
        else:
            tokenizer = AutoTokenizer.from_pretrained(model_path, use_fast=False)
            model = AutoModelForCausalLM.from_pretrained(
                model_path, low_cpu_mem_usage=True, **kwargs
            )

    image_processor = CLIPImageProcessor.from_pretrained(model_path)

    if "phi" in model_name.lower():
        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)

        # TODO: the tokenizer length of phi-2 is 50295, but the output class of lm_head is 51200
        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))
    else:
        raise ValueError(f"Unsupported model name: {model_name}")

    if hasattr(model.config, "max_sequence_length"):
        context_len = model.config.max_sequence_length
    else:
        context_len = 2048
    model.to(device="cuda")
    print(kwargs)
    return tokenizer, model, image_processor, context_len