Spaces:
Running
on
Zero
Running
on
Zero
| # 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 shutil | |
| from transformers import ( | |
| AutoTokenizer, | |
| AutoModelForCausalLM, | |
| AutoConfig, | |
| BitsAndBytesConfig, | |
| ) | |
| import torch | |
| from llava.model import * | |
| from llava.constants import ( | |
| DEFAULT_IMAGE_PATCH_TOKEN, | |
| DEFAULT_IM_START_TOKEN, | |
| DEFAULT_IM_END_TOKEN, | |
| ) | |
| def load_pretrained_model( | |
| model_path, | |
| model_base, | |
| model_name, | |
| load_8bit=False, | |
| load_4bit=False, | |
| device_map="auto", | |
| load_bf16=False, | |
| ): | |
| 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", | |
| ) | |
| elif load_bf16: | |
| kwargs["torch_dtype"] = torch.bfloat16 | |
| else: | |
| kwargs["torch_dtype"] = torch.float16 | |
| if "llava" in model_name.lower(): | |
| # Load LLaVA model | |
| 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_path, use_fast=False) | |
| print("Loading LLaVA from base model...") | |
| model = LlavaLlamaForCausalLM.from_pretrained( | |
| model_base, low_cpu_mem_usage=True, config=lora_cfg_pretrained, **kwargs | |
| ) | |
| if model.get_vision_tower() is not None and not model.get_vision_tower().is_loaded: | |
| model.get_vision_tower().load_model() | |
| # if the parameters have been ever modified during model training, | |
| # then for some reason, the layer will have the correct shape | |
| # but the weight will have a wrong shape | |
| # the code below fix this weird shape mismatch issue | |
| 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) | |
| ) | |
| # if the parameters have been ever modified during model training, | |
| # then for some reason, the layer will have the correct shape | |
| # but the weight will have a wrong shape | |
| # the code below fix this weird shape mismatch issue | |
| if model.get_vision_tower() is not None: | |
| mm_projector_in, mm_projector_out = ( | |
| model.model.mm_projector.in_features, | |
| model.model.mm_projector.out_features, | |
| ) | |
| if ( | |
| model.model.mm_projector.weight.shape[1] != mm_projector_in | |
| or model.model.mm_projector.weight.shape[0] != mm_projector_out | |
| ): | |
| model.model.mm_projector.weight = torch.nn.Parameter( | |
| torch.empty( | |
| mm_projector_out, | |
| mm_projector_in, | |
| device=model.device, | |
| dtype=model.dtype, | |
| ) | |
| ) | |
| model.model.mm_projector.bias = torch.nn.Parameter( | |
| torch.empty(mm_projector_out, device=model.device, dtype=model.dtype) | |
| ) | |
| print("Loading additional LLaVA 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 from base model...") | |
| if "mpt" in model_name.lower(): | |
| if not os.path.isfile(os.path.join(model_path, "configuration_mpt.py")): | |
| shutil.copyfile( | |
| os.path.join(model_base, "configuration_mpt.py"), | |
| os.path.join(model_path, "configuration_mpt.py"), | |
| ) | |
| tokenizer = AutoTokenizer.from_pretrained(model_path, use_fast=True) | |
| cfg_pretrained = AutoConfig.from_pretrained(model_path, trust_remote_code=True) | |
| model = LlavaMPTForCausalLM.from_pretrained( | |
| model_base, low_cpu_mem_usage=True, config=cfg_pretrained, **kwargs | |
| ) | |
| else: | |
| tokenizer = AutoTokenizer.from_pretrained(model_path, use_fast=False) | |
| cfg_pretrained = AutoConfig.from_pretrained(model_path) | |
| model = LlavaLlamaForCausalLM.from_pretrained( | |
| model_base, low_cpu_mem_usage=True, config=cfg_pretrained, **kwargs | |
| ) | |
| # load mm projector weights (this include the vision tower weights too) | |
| if model.get_vision_tower() is not None: | |
| if not model.get_vision_tower().is_loaded: | |
| model.get_vision_tower().load_model() | |
| mm_projector_weights = torch.load( | |
| os.path.join(model_path, "mm_projector.bin"), map_location="cpu" | |
| ) | |
| mm_projector_weights = {k: v for k, v in mm_projector_weights.items()} | |
| model.load_state_dict( | |
| mm_projector_weights, strict=False | |
| ) # for 3d point cloud, this will load the vision tower too. | |
| else: | |
| if "mpt" in model_name.lower(): | |
| tokenizer = AutoTokenizer.from_pretrained(model_path, use_fast=True) | |
| model = LlavaMPTForCausalLM.from_pretrained( | |
| model_path, low_cpu_mem_usage=True, **kwargs | |
| ) | |
| else: | |
| tokenizer = AutoTokenizer.from_pretrained(model_path, use_fast=False) | |
| model = LlavaLlamaForCausalLM.from_pretrained( | |
| model_path, low_cpu_mem_usage=True, **kwargs | |
| ) | |
| else: | |
| # Load language model | |
| if model_base is not None: | |
| # PEFT model | |
| from peft import PeftModel | |
| tokenizer = AutoTokenizer.from_pretrained(model_path, use_fast=False) | |
| model = AutoModelForCausalLM.from_pretrained( | |
| model_base, | |
| torch_dtype=torch.bfloat16, | |
| 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 BF16...") | |
| model.to(torch.bfloat16) | |
| else: | |
| use_fast = False | |
| if "mpt" in model_name.lower(): | |
| tokenizer = AutoTokenizer.from_pretrained(model_path, use_fast=True) | |
| model = AutoModelForCausalLM.from_pretrained( | |
| model_path, low_cpu_mem_usage=True, trust_remote_code=True, **kwargs | |
| ) | |
| 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(): | |
| 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)) | |
| vision_tower = model.get_vision_tower() | |
| if vision_tower is not None: | |
| if not vision_tower.is_loaded: | |
| vision_tower.load_model() | |
| 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 | |