# Copyright 2024 Zhenwei Shao and MILVLG team. # Licensed under the Apache License, Version 2.0. # Adopted from https://github.com/haotian-liu/LLaVA. Below is the original copyright: # Copyright 2023 Haotian Liu # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import os import warnings import shutil from transformers import AutoTokenizer, AutoModelForCausalLM, AutoConfig, BitsAndBytesConfig import torch from flashsloth.model import * from flashsloth.constants import DEFAULT_IMAGE_PATCH_TOKEN, DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN from flashsloth.model.multimodal_encoder.builder import build_vision_tower from .multimodal_projector.builder import build_vision_projector from flashsloth import logger def load_pretrained_model(model_path, model_base, model_name, load_8bit=False, load_4bit=False, device_map="auto", device="cuda"): kwargs = {"device_map": device_map} 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 logger.info(f'load cfg kwargs: {kwargs}') if 'llava' in model_name.lower() or 'flashsloth' in model_name.lower(): 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. Detailed instruction: https://github.com/haotian-liu/LLaVA#launch-a-model-worker-lora-weights-unmerged.') exit() if 'lora' in model_name.lower() and model_base is not None: # Load model trained with LoRA logger.info(f'Load model name trained with LoRA, model base: {model_base}') assert 'flashsloth' in model_name.lower(), 'The model name must contain `flashsloth`' tokenizer = AutoTokenizer.from_pretrained(model_base, use_fast=False, trust_remote_code=True) if 'phi-2' in model_name.lower() or 'phi2' in model_name.lower(): lora_cfg_pretrained = FlashSlothConfig.from_pretrained(model_path) model = FlashSlothForCausalLM.from_pretrained(model_base, config=lora_cfg_pretrained, **kwargs) else: lora_cfg_pretrained = FlashSlothConfig.from_pretrained(model_path) model = FlashSlothForCausalLM.from_pretrained(model_base, config=lora_cfg_pretrained, **kwargs) logger.info('Loading additional 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(f'model.{model.base_model_prefix}.') 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()} logger.info(f'Loading additional weights: f{[*non_lora_trainables.keys()]}') model.load_state_dict(non_lora_trainables, strict=False) from peft import PeftModel logger.info('Loading LoRA weights...') model = PeftModel.from_pretrained(model, model_path) logger.info('Merging LoRA weights...') model = model.merge_and_unload() logger.info('Model is loaded...') elif model_base is not None: logger.info('Load mm projector only model...') if 'phi2' in model_name.lower() or 'phi-2' in model_name.lower(): logger.info(f'model_base:, {model_base}') config = FlashSlothConfig.from_pretrained(model_path, trust_remote_code=True) model = FlashSlothForCausalLM.from_pretrained(model_base, **kwargs) model.model.vision_tower = build_vision_tower(config) model.model.mm_projector = build_vision_projector(config) tokenizer = AutoTokenizer.from_pretrained(model_base) else: logger.info(f'model_base:, {model_base}') config = FlashSlothConfig.from_pretrained(model_path, trust_remote_code=True) model = FlashSlothForCausalLM.from_pretrained(model_base, **kwargs) model.model.vision_tower = build_vision_tower(config) model.model.mm_projector = build_vision_projector(config) tokenizer = AutoTokenizer.from_pretrained(model_base) 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()} logger.info(f'loading mm projector weights: {[*mm_projector_weights.keys()]}') model.load_state_dict(mm_projector_weights, strict=False) # model.to(device) logger.info('Model is loaded...') else: logger.info(f'load fully fine-tuned model or HF Hub model: {model_path}') #hg version if 'phi2' in model_name.lower() or 'phi-2' in model_name.lower(): tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True) model = FlashSlothForCausalLM.from_pretrained(model_path, **kwargs) logger.info('Model is loaded...') else: tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True) model = FlashSlothForCausalLM.from_pretrained(model_path, **kwargs) logger.info('Model is loaded...') else: raise NotImplementedError # 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, low_cpu_mem_usage=True, **kwargs) logger.info(f"Loading LoRA weights from {model_path}") model = PeftModel.from_pretrained(model, model_path) logger.info(f"Merging weights") model = model.merge_and_unload() logger.info('Convert to FP16...') model.to(torch.float16) else: if 'phi2' in model_name.lower() or 'flashsloth' in model_name.lower(): raise NotImplementedError else: tokenizer = AutoTokenizer.from_pretrained(model_path, use_fast=False) model = AutoModelForCausalLM.from_pretrained(model_path, low_cpu_mem_usage=True, **kwargs) image_processor = None if 'llava' in model_name.lower() or 'flashsloth' 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) 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) # FIXME: phi-2 has unused embeddings. # [Edited by zhenwei - 2024-01-31 13:50] # model.resize_token_embeddings(len(tokenizer)) vision_tower = model.get_vision_tower() if not vision_tower.is_loaded: vision_tower.load_model() logger.info('Delayed vision tower loaded.') vision_tower.to(device=model.device, dtype=model.dtype) image_processor = vision_tower.image_processor if hasattr(model.config, "max_sequence_length"): context_len = model.config.max_sequence_length else: context_len = 2048 return tokenizer, model, image_processor, context_len