# Adopted from https://github.com/lm-sys/FastChat. Below is the original copyright: # Adopted from tatsu-lab@stanford_alpaca. Below is the original copyright: # Copyright 2023 Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li # # 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 io import copy from dataclasses import dataclass, field import json import logging import pathlib from typing import Dict, Optional, Sequence, List import glob import torch import transformers import tokenizers from llava.constants import IGNORE_INDEX, DEFAULT_IMAGE_TOKEN from torch.utils.data import Dataset from llava.train.llava_trainer import LLaVATrainer from llava import conversation as conversation_lib from llava.model import * from llava.mm_utils import tokenizer_image_token from PIL import Image, UnidentifiedImageError import random from datasets import load_dataset, concatenate_datasets, Image local_rank = None def rank0_print(*args): if local_rank == 0: print(*args) from packaging import version IS_TOKENIZER_GREATER_THAN_0_14 = version.parse(tokenizers.__version__) >= version.parse('0.14') @dataclass class ModelArguments: model_name_or_path: Optional[str] = field(default="facebook/opt-125m") version: Optional[str] = field(default="v0") freeze_backbone: bool = field(default=False) tune_mm_mlp_adapter: bool = field(default=False) vision_tower: Optional[str] = field(default=None) gen_vision_tower: Optional[str] = field(default=None) mm_vision_select_layer: Optional[int] = field(default=-1) # default to the last layer pretrain_mm_mlp_adapter: Optional[str] = field(default=None) pretrain_gen_mlp_adapter: Optional[str] = field(default=None) vision_tower_pretrained: Optional[str] = field(default=None) mm_projector_type: Optional[str] = field(default='linear') gen_projector_type: Optional[str] = field(default='linear') mm_use_im_start_end: bool = field(default=False) mm_use_im_patch_token: bool = field(default=True) mm_patch_merge_type: Optional[str] = field(default='flat') mm_vision_select_feature: Optional[str] = field(default="patch") n_query: Optional[int] = field(default=729) # clip 576, siglip 729 n_und_query: Optional[int] = field(default=729) # clip 576, siglip 729 gen_pooling: Optional[str] = field(default="all") # options are: pool2d_3, pool2d_9, seq_3, seq_9, seq_27 @dataclass class DataArguments: data_path: str = field(default=None, metadata={"help": "Path to the training data."}) lazy_preprocess: bool = False is_multimodal: bool = False image_folder: Optional[str] = field(default=None) pixelprose_image_folder: Optional[str] = field(default=None) datacomp_image_folder: Optional[str] = field(default=None) data_type: Optional[str] = field(default="webdataset") image_aspect_ratio: str = 'square' @dataclass class TrainingArguments(transformers.TrainingArguments): cache_dir: Optional[str] = field(default=None) optim: str = field(default="adamw_torch") remove_unused_columns: bool = field(default=False) freeze_mm_mlp_adapter: bool = field(default=False) mpt_attn_impl: Optional[str] = field(default="triton") model_max_length: int = field( default=512, metadata={ "help": "Maximum sequence length. Sequences will be right padded (and possibly truncated)." }, ) double_quant: bool = field( default=True, metadata={"help": "Compress the quantization statistics through double quantization."} ) quant_type: str = field( default="nf4", metadata={"help": "Quantization data type to use. Should be one of `fp4` or `nf4`."} ) bits: int = field( default=16, metadata={"help": "How many bits to use."} ) lora_enable: bool = False lora_r: int = 64 lora_alpha: int = 16 lora_dropout: float = 0.05 lora_weight_path: str = "" lora_bias: str = "none" mm_projector_lr: Optional[float] = None group_by_modality_length: bool = field(default=False) def maybe_zero_3(param, ignore_status=False, name=None): from deepspeed import zero from deepspeed.runtime.zero.partition_parameters import ZeroParamStatus if hasattr(param, "ds_id"): if param.ds_status == ZeroParamStatus.NOT_AVAILABLE: if not ignore_status: logging.warning(f"{name}: param.ds_status != ZeroParamStatus.NOT_AVAILABLE: {param.ds_status}") with zero.GatheredParameters([param]): param = param.data.detach().cpu().clone() else: param = param.detach().cpu().clone() return param # Borrowed from peft.utils.get_peft_model_state_dict def get_peft_state_maybe_zero_3(named_params, bias): if bias == "none": to_return = {k: t for k, t in named_params if "lora_" in k} elif bias == "all": to_return = {k: t for k, t in named_params if "lora_" in k or "bias" in k} elif bias == "lora_only": to_return = {} maybe_lora_bias = {} lora_bias_names = set() for k, t in named_params: if "lora_" in k: to_return[k] = t bias_name = k.split("lora_")[0] + "bias" lora_bias_names.add(bias_name) elif "bias" in k: maybe_lora_bias[k] = t for k, t in maybe_lora_bias: if bias_name in lora_bias_names: to_return[bias_name] = t else: raise NotImplementedError to_return = {k: maybe_zero_3(v, ignore_status=True) for k, v in to_return.items()} return to_return def get_peft_state_non_lora_maybe_zero_3(named_params, require_grad_only=True): to_return = {k: t for k, t in named_params if "lora_" not in k} if require_grad_only: to_return = {k: t for k, t in to_return.items() if t.requires_grad} to_return = {k: maybe_zero_3(v, ignore_status=True).cpu() for k, v in to_return.items()} return to_return def get_mm_adapter_state_maybe_zero_3(named_params, keys_to_match): to_return = {k: t for k, t in named_params if any(key_match in k for key_match in keys_to_match)} to_return = {k: maybe_zero_3(v, ignore_status=True).cpu() for k, v in to_return.items()} return to_return def get_vision_tower_state_maybe_zero_3(named_params, keys_to_match=['']): to_return = {k: t for k, t in named_params if any(key_match in k for key_match in keys_to_match)} to_return = {k: maybe_zero_3(v, ignore_status=True).cpu()for k, v in to_return.items()} return to_return def find_all_linear_names(model): cls = torch.nn.Linear lora_module_names = set() multimodal_keywords = ['mm_projector', 'vision_tower', 'vision_resampler'] for name, module in model.named_modules(): if any(mm_keyword in name for mm_keyword in multimodal_keywords): continue if isinstance(module, cls): names = name.split('.') lora_module_names.add(names[0] if len(names) == 1 else names[-1]) if 'lm_head' in lora_module_names: # needed for 16-bit lora_module_names.remove('lm_head') return list(lora_module_names) def safe_save_model_for_hf_trainer(trainer: transformers.Trainer, output_dir: str, vision_tower: str): """Collects the state dict and dump to disk.""" # if getattr(trainer.args, "tune_vision_model", False): if trainer.deepspeed: torch.cuda.synchronize() # ## hardcode for siglip save # if 'siglip' in vision_tower: # preprocessor_config_path = os.path.join(vision_tower, 'preprocessor_config.json') # config_path = os.path.join(vision_tower, 'config.json') # special_tokens_map = os.path.join(vision_tower, 'special_tokens_map.json') # tokenizer_config = os.path.join(vision_tower, 'tokenizer_config.json') # vision_tower_dir = os.path.join(output_dir, 'vision_tower') # os.makedirs(vision_tower_dir, exist_ok=True) # import shutil # shutil.copy(preprocessor_config_path, os.path.join(vision_tower_dir, 'preprocessor_config.json')) # shutil.copy(config_path, os.path.join(vision_tower_dir, 'config.json')) # shutil.copy(special_tokens_map, os.path.join(vision_tower_dir, 'special_tokens_map.json')) # shutil.copy(tokenizer_config, os.path.join(vision_tower_dir, 'tokenizer_config.json')) # weight_to_save = get_vision_tower_state_maybe_zero_3( # trainer.model.get_vision_tower().vision_tower.named_parameters()) # if trainer.args.local_rank == 0 or trainer.args.local_rank == -1: # torch.save(weight_to_save, os.path.join( # output_dir, 'vision_tower/pytorch_model.bin')) # else: trainer.model.get_vision_tower().image_processor.save_pretrained( os.path.join(output_dir, 'vision_tower')) trainer.model.get_vision_tower().config.save_pretrained( os.path.join(output_dir, 'vision_tower')) weight_to_save = get_vision_tower_state_maybe_zero_3( trainer.model.get_vision_tower().vision_tower.named_parameters()) if trainer.args.local_rank == 0 or trainer.args.local_rank == -1: torch.save(weight_to_save, os.path.join( output_dir, 'vision_tower/pytorch_model.bin')) # if getattr(trainer.args, "tune_mm_mlp_adapter", False): # # Only save Adapter # keys_to_match = ['mm_projector'] # if getattr(trainer.args, "use_im_start_end", False): # keys_to_match.extend(['embed_tokens', 'embed_in']) # weight_to_save = get_mm_adapter_state_maybe_zero_3(trainer.model.named_parameters(), keys_to_match) # trainer.model.config.save_pretrained(output_dir) # current_folder = output_dir.split('/')[-1] # parent_folder = os.path.dirname(output_dir) # if trainer.args.local_rank == 0 or trainer.args.local_rank == -1: # if current_folder.startswith('checkpoint-'): # mm_projector_folder = os.path.join(parent_folder, "mm_projector") # os.makedirs(mm_projector_folder, exist_ok=True) # torch.save(weight_to_save, os.path.join(mm_projector_folder, f'{current_folder}.bin')) # else: # torch.save(weight_to_save, os.path.join(output_dir, f'mm_projector.bin')) # return # Only save Adapter keys_to_match = ['mm_projector'] if getattr(trainer.args, "use_im_start_end", False): keys_to_match.extend(['embed_tokens', 'embed_in']) weight_to_save = get_mm_adapter_state_maybe_zero_3(trainer.model.named_parameters(), keys_to_match) trainer.model.config.save_pretrained(output_dir) current_folder = output_dir.split('/')[-1] parent_folder = os.path.dirname(output_dir) if trainer.args.local_rank == 0 or trainer.args.local_rank == -1: if current_folder.startswith('checkpoint-'): mm_projector_folder = os.path.join(parent_folder, "mm_projector") os.makedirs(mm_projector_folder, exist_ok=True) torch.save(weight_to_save, os.path.join(mm_projector_folder, f'{current_folder}.bin')) else: torch.save(weight_to_save, os.path.join(output_dir, f'mm_projector.bin')) keys_to_match = ['gen_projector'] if getattr(trainer.args, "use_im_start_end", False): keys_to_match.extend(['embed_tokens', 'embed_in']) weight_to_save = get_mm_adapter_state_maybe_zero_3(trainer.model.named_parameters(), keys_to_match) trainer.model.config.save_pretrained(output_dir) current_folder = output_dir.split('/')[-1] parent_folder = os.path.dirname(output_dir) if trainer.args.local_rank == 0 or trainer.args.local_rank == -1: if current_folder.startswith('checkpoint-'): mm_projector_folder = os.path.join(parent_folder, "gen_projector") os.makedirs(mm_projector_folder, exist_ok=True) torch.save(weight_to_save, os.path.join(mm_projector_folder, f'{current_folder}.bin')) else: torch.save(weight_to_save, os.path.join(output_dir, f'gen_projector.bin')) if trainer.deepspeed: torch.cuda.synchronize() trainer.save_model(output_dir) return state_dict = trainer.model.state_dict() if trainer.args.should_save: cpu_state_dict = { key: value.cpu() for key, value in state_dict.items() } del state_dict trainer._save(output_dir, state_dict=cpu_state_dict) # noqa def smart_tokenizer_and_embedding_resize( special_tokens_dict: Dict, tokenizer: transformers.PreTrainedTokenizer, model: transformers.PreTrainedModel, ): """Resize tokenizer and embedding. Note: This is the unoptimized version that may make your embedding size not be divisible by 64. """ num_new_tokens = tokenizer.add_special_tokens(special_tokens_dict) model.resize_token_embeddings(len(tokenizer)) if num_new_tokens > 0: input_embeddings = model.get_input_embeddings().weight.data output_embeddings = model.get_output_embeddings().weight.data input_embeddings_avg = input_embeddings[:-num_new_tokens].mean( dim=0, keepdim=True) output_embeddings_avg = output_embeddings[:-num_new_tokens].mean( dim=0, keepdim=True) input_embeddings[-num_new_tokens:] = input_embeddings_avg output_embeddings[-num_new_tokens:] = output_embeddings_avg def _tokenize_fn(strings: Sequence[str], tokenizer: transformers.PreTrainedTokenizer) -> Dict: """Tokenize a list of strings.""" tokenized_list = [ tokenizer( text, return_tensors="pt", padding="longest", max_length=tokenizer.model_max_length, truncation=True, ) for text in strings ] input_ids = labels = [ tokenized.input_ids[0] for tokenized in tokenized_list ] input_ids_lens = labels_lens = [ tokenized.input_ids.ne(tokenizer.pad_token_id).sum().item() for tokenized in tokenized_list ] return dict( input_ids=input_ids, labels=labels, input_ids_lens=input_ids_lens, labels_lens=labels_lens, ) def _mask_targets(target, tokenized_lens, speakers): # cur_idx = 0 cur_idx = tokenized_lens[0] tokenized_lens = tokenized_lens[1:] target[:cur_idx] = IGNORE_INDEX for tokenized_len, speaker in zip(tokenized_lens, speakers): if speaker == "human": target[cur_idx + 2:cur_idx + tokenized_len] = IGNORE_INDEX cur_idx += tokenized_len def _add_speaker_and_signal(header, source, get_conversation=True): """Add speaker and start/end signal on each round.""" BEGIN_SIGNAL = "### " END_SIGNAL = "\n" conversation = header for sentence in source: from_str = sentence["from"] if from_str.lower() == "human": from_str = conversation_lib.default_conversation.roles[0] elif from_str.lower() == "gpt": from_str = conversation_lib.default_conversation.roles[1] else: from_str = 'unknown' sentence["value"] = (BEGIN_SIGNAL + from_str + ": " + sentence["value"] + END_SIGNAL) if get_conversation: conversation += sentence["value"] conversation += BEGIN_SIGNAL return conversation def preprocess_multimodal( sources: Sequence[str], data_args: DataArguments ) -> Dict: is_multimodal = data_args.is_multimodal if not is_multimodal: return sources und_placeholder = "[IMG]" + "" * data_args.n_und_query + "[/IMG]" gen_placeholder = "[IMG]" + "" * data_args.n_query + "[/IMG]" inst_type = None for source in sources: # [instance] for sentence in source: if sentence['from'] == "human" and '' in sentence['value']: sentence['value'] = sentence['value'].replace( DEFAULT_IMAGE_TOKEN, und_placeholder).strip() inst_type = "und" elif sentence['from'] == "gpt" and '' in sentence['value']: sentence['value'] = sentence['value'].replace( DEFAULT_IMAGE_TOKEN, gen_placeholder).strip() inst_type = "gen" return sources, inst_type def preprocess_llama_2( sources, tokenizer: transformers.PreTrainedTokenizer, has_image: bool = False ) -> Dict: conv = conversation_lib.default_conversation.copy() roles = {"human": conv.roles[0], "gpt": conv.roles[1]} # Apply prompt templates conversations = [] for i, source in enumerate(sources): if roles[source[0]["from"]] != conv.roles[0]: # Skip the first one if it is not from human source = source[1:] conv.messages = [] for j, sentence in enumerate(source): role = roles[sentence["from"]] assert role == conv.roles[j % 2], f"{i}" conv.append_message(role, sentence["value"]) conversations.append(conv.get_prompt()) # Tokenize conversations if has_image: input_ids = torch.stack([tokenizer_image_token(prompt, tokenizer, return_tensors='pt') for prompt in conversations], dim=0) else: input_ids = tokenizer( conversations, return_tensors="pt", padding="longest", max_length=tokenizer.model_max_length, truncation=True, ).input_ids targets = input_ids.clone() assert conv.sep_style == conversation_lib.SeparatorStyle.LLAMA_2 # Mask targets sep = "[/INST] " for conversation, target in zip(conversations, targets): total_len = int(target.ne(tokenizer.pad_token_id).sum()) rounds = conversation.split(conv.sep2) cur_len = 1 target[:cur_len] = IGNORE_INDEX for i, rou in enumerate(rounds): if rou == "": break parts = rou.split(sep) if len(parts) != 2: break parts[0] += sep if has_image: round_len = len(tokenizer_image_token(rou, tokenizer)) instruction_len = len(tokenizer_image_token(parts[0], tokenizer)) - 2 else: round_len = len(tokenizer(rou).input_ids) instruction_len = len(tokenizer(parts[0]).input_ids) - 2 target[cur_len : cur_len + instruction_len] = IGNORE_INDEX cur_len += round_len target[cur_len:] = IGNORE_INDEX if cur_len < tokenizer.model_max_length: if cur_len != total_len: target[:] = IGNORE_INDEX print( f"WARNING: tokenization mismatch: {cur_len} vs. {total_len}." f" (ignored)" ) return dict( input_ids=input_ids, labels=targets, ) def preprocess_v1( sources, tokenizer: transformers.PreTrainedTokenizer, has_image: bool = False ) -> Dict: conv = conversation_lib.default_conversation.copy() roles = {"human": conv.roles[0], "gpt": conv.roles[1]} # Apply prompt templates conversations = [] for i, source in enumerate(sources): if roles[source[0]["from"]] != conv.roles[0]: # Skip the first one if it is not from human source = source[1:] conv.messages = [] for j, sentence in enumerate(source): role = roles[sentence["from"]] assert role == conv.roles[j % 2], f"{i}" conv.append_message(role, sentence["value"]) conversations.append(conv.get_prompt()) # Tokenize conversations if has_image: input_ids = torch.stack([tokenizer_image_token(prompt, tokenizer, return_tensors='pt') for prompt in conversations], dim=0) else: input_ids = tokenizer( conversations, return_tensors="pt", padding="longest", max_length=tokenizer.model_max_length, truncation=True, ).input_ids targets = input_ids.clone() assert conv.sep_style == conversation_lib.SeparatorStyle.TWO # Mask targets sep = conv.sep + conv.roles[1] + ": " for conversation, target in zip(conversations, targets): total_len = int(target.ne(tokenizer.pad_token_id).sum()) rounds = conversation.split(conv.sep2) cur_len = 1 target[:cur_len] = IGNORE_INDEX for i, rou in enumerate(rounds): if rou == "": break parts = rou.split(sep) if len(parts) != 2: break parts[0] += sep if has_image: round_len = len(tokenizer_image_token(rou, tokenizer)) instruction_len = len(tokenizer_image_token(parts[0], tokenizer)) - 2 else: round_len = len(tokenizer(rou).input_ids) instruction_len = len(tokenizer(parts[0]).input_ids) - 2 if i != 0 and not tokenizer.legacy and IS_TOKENIZER_GREATER_THAN_0_14: round_len -= 1 instruction_len -= 1 target[cur_len : cur_len + instruction_len] = IGNORE_INDEX cur_len += round_len target[cur_len:] = IGNORE_INDEX if cur_len < tokenizer.model_max_length: if cur_len != total_len: target[:] = IGNORE_INDEX print( f"WARNING: tokenization mismatch: {cur_len} vs. {total_len}." f" (ignored)" ) return dict( input_ids=input_ids, labels=targets, ) def preprocess_mpt( sources, tokenizer: transformers.PreTrainedTokenizer, has_image: bool = False ) -> Dict: conv = conversation_lib.default_conversation.copy() roles = {"human": conv.roles[0], "gpt": conv.roles[1]} # Apply prompt templates conversations = [] for i, source in enumerate(sources): if roles[source[0]["from"]] != conv.roles[0]: # Skip the first one if it is not from human source = source[1:] conv.messages = [] for j, sentence in enumerate(source): role = roles[sentence["from"]] assert role == conv.roles[j % 2], f"{i}" conv.append_message(role, sentence["value"]) conversations.append(conv.get_prompt()) # Tokenize conversations if has_image: input_ids = torch.stack([tokenizer_image_token(prompt, tokenizer, return_tensors='pt') for prompt in conversations], dim=0) else: input_ids = tokenizer( conversations, return_tensors="pt", padding="longest", max_length=tokenizer.model_max_length, truncation=True, ).input_ids targets = input_ids.clone() assert conv.sep_style == conversation_lib.SeparatorStyle.MPT # Mask targets sep = conv.sep + conv.roles[1] for conversation, target in zip(conversations, targets): total_len = int(target.ne(tokenizer.pad_token_id).sum()) rounds = conversation.split(conv.sep) re_rounds = [conv.sep.join(rounds[:3])] # system + user + gpt for conv_idx in range(3, len(rounds), 2): re_rounds.append(conv.sep.join(rounds[conv_idx:conv_idx+2])) # user + gpt cur_len = 0 target[:cur_len] = IGNORE_INDEX for i, rou in enumerate(re_rounds): if rou == "": break parts = rou.split(sep) if len(parts) != 2: break parts[0] += sep if has_image: round_len = len(tokenizer_image_token(rou, tokenizer)) instruction_len = len(tokenizer_image_token(parts[0], tokenizer)) - 1 else: round_len = len(tokenizer(rou).input_ids) instruction_len = len(tokenizer(parts[0]).input_ids) - 1 if i != 0 and getattr(tokenizer, 'legacy', False) and IS_TOKENIZER_GREATER_THAN_0_14: round_len += 1 instruction_len += 1 target[cur_len : cur_len + instruction_len] = IGNORE_INDEX cur_len += round_len target[cur_len:] = IGNORE_INDEX if cur_len < tokenizer.model_max_length: if cur_len != total_len: target[:] = IGNORE_INDEX print( f"WARNING: tokenization mismatch: {cur_len} vs. {total_len}." f" (ignored)" ) return dict( input_ids=input_ids, labels=targets, ) def preprocess_llama3( sources, tokenizer: transformers.PreTrainedTokenizer, has_image: bool = False ) -> Dict: conv = conversation_lib.default_conversation.copy() roles = {"human": conv.roles[0], "gpt": conv.roles[1]} # Apply prompt templates conversations = [] for i, source in enumerate(sources): if roles[source[0]["from"]] != conv.roles[0]: # Skip the first one if it is not from human source = source[1:] conv.messages = [] for j, sentence in enumerate(source): role = roles[sentence["from"]] assert role == conv.roles[j % 2], f"{i}" conv.append_message(role, sentence["value"]) conversations.append(conv.get_prompt()) # Tokenize conversations if has_image: input_ids = torch.stack( [tokenizer_image_token(prompt, tokenizer, return_tensors='pt') for prompt in conversations], dim=0) else: input_ids = tokenizer( conversations, return_tensors="pt", padding="longest", max_length=tokenizer.model_max_length, truncation=True, ).input_ids targets = input_ids.clone() assert conv.sep_style == conversation_lib.SeparatorStyle.MPT # Mask targets sep = conv.sep + conv.roles[1] for conversation, target in zip(conversations, targets): total_len = int(target.ne(tokenizer.pad_token_id).sum()) rounds = conversation.split(conv.sep) re_rounds = [conv.sep.join(rounds[:3])] for conv_idx in range(3, len(rounds), 2): re_rounds.append(conv.sep.join(rounds[conv_idx:conv_idx + 2])) cur_len = 0 target[:cur_len] = IGNORE_INDEX for i, rou in enumerate(re_rounds): if rou == "": break parts = rou.split(sep) if len(parts) != 2: break parts[0] += sep if has_image: round_len = len(tokenizer_image_token(rou, tokenizer)) + 1 instruction_len = len(tokenizer_image_token(parts[0], tokenizer)) else: round_len = len(tokenizer(rou).input_ids) + 1 instruction_len = len(tokenizer(parts[0]).input_ids) target[cur_len: cur_len + instruction_len] = IGNORE_INDEX cur_len += round_len target[cur_len:] = IGNORE_INDEX if cur_len < tokenizer.model_max_length: if cur_len != total_len: target[:] = IGNORE_INDEX print( f"WARNING: tokenization mismatch: {cur_len} vs. {total_len}." f" (ignored)" ) return dict( input_ids=input_ids, labels=targets, ) def preprocess_plain( sources: Sequence[str], tokenizer: transformers.PreTrainedTokenizer, ) -> Dict: # add end signal and concatenate together conversations = [] for source in sources: assert len(source) == 2 assert DEFAULT_IMAGE_TOKEN in source[0]['value'] or DEFAULT_IMAGE_TOKEN in source[1]['value'] conversation = source[0]['value'] + source[1]['value'] + conversation_lib.default_conversation.sep conversations.append(conversation) # tokenize conversations input_ids = [tokenizer_image_token(prompt, tokenizer, return_tensors='pt') for prompt in conversations] targets = copy.deepcopy(input_ids) for target, source in zip(targets, sources): tokenized_len = len(tokenizer_image_token(source[0]['value'], tokenizer)) target[:tokenized_len] = IGNORE_INDEX return dict(input_ids=input_ids, labels=targets) def preprocess( sources: Sequence[str], tokenizer: transformers.PreTrainedTokenizer, has_image: bool = False ) -> Dict: """ Given a list of sources, each is a conversation list. This transform: 1. Add signal '### ' at the beginning each sentence, with end signal '\n'; 2. Concatenate conversations together; 3. Tokenize the concatenated conversation; 4. Make a deepcopy as the target. Mask human words with IGNORE_INDEX. """ if conversation_lib.default_conversation.sep_style == conversation_lib.SeparatorStyle.PLAIN: return preprocess_plain(sources, tokenizer) if conversation_lib.default_conversation.sep_style == conversation_lib.SeparatorStyle.LLAMA_2: return preprocess_llama_2(sources, tokenizer, has_image=has_image) if conversation_lib.default_conversation.version.startswith("v1"): return preprocess_v1(sources, tokenizer, has_image=has_image) if conversation_lib.default_conversation.version == "mpt": return preprocess_mpt(sources, tokenizer, has_image=has_image) if conversation_lib.default_conversation.version == "llama3": return preprocess_llama3(sources, tokenizer, has_image=has_image) # add end signal and concatenate together conversations = [] for source in sources: header = f"{conversation_lib.default_conversation.system}\n\n" conversation = _add_speaker_and_signal(header, source) conversations.append(conversation) # tokenize conversations def get_tokenize_len(prompts): return [len(tokenizer_image_token(prompt, tokenizer)) for prompt in prompts] if has_image: input_ids = [tokenizer_image_token(prompt, tokenizer, return_tensors='pt') for prompt in conversations] else: conversations_tokenized = _tokenize_fn(conversations, tokenizer) input_ids = conversations_tokenized["input_ids"] targets = copy.deepcopy(input_ids) for target, source in zip(targets, sources): if has_image: tokenized_lens = get_tokenize_len([header] + [s["value"] for s in source]) else: tokenized_lens = _tokenize_fn([header] + [s["value"] for s in source], tokenizer)["input_ids_lens"] speakers = [sentence["from"] for sentence in source] _mask_targets(target, tokenized_lens, speakers) return dict(input_ids=input_ids, labels=targets) class LazySupervisedWebDataset(Dataset): """Dataset for supervised fine-tuning.""" def __init__(self, data_path: str, tokenizer: transformers.PreTrainedTokenizer, data_args: DataArguments): super(LazySupervisedWebDataset, self).__init__() self.data_args = data_args list_data_dict = [] ## cc12m ## image to text # train_dataset = load_dataset("webdataset", data_files=glob.glob(os.path.join(self.data_args.pixelprose_image_folder, 'cc12m_part*/*.tar')), split="train", num_proc=128) # pixelprose_json = train_dataset['json'][0] # json_with_none_values = {key: None for key in pixelprose_json} # train_dataset = train_dataset.rename_column("jpg", "image") # train_dataset = train_dataset.add_column('type', len(train_dataset) * ['Pixelprose_I2T']) # train_dataset = train_dataset.remove_columns([col for col in train_dataset.column_names if not col in ( # ["image", "json", "type", "txt"])]) # list_data_dict.append(train_dataset) # ## text to image # train_dataset = load_dataset("webdataset", data_files=glob.glob(os.path.join(self.data_args.pixelprose_image_folder, 'cc12m_part*/*.tar')), split="train", num_proc=128) # train_dataset = train_dataset.rename_column("jpg", "image") # train_dataset = train_dataset.add_column('type', len(train_dataset) * ['Pixelprose_T2I']) # train_dataset = train_dataset.remove_columns([col for col in train_dataset.column_names if not col in ( # ["image", "json", "type", "txt"])]) # list_data_dict.append(train_dataset) # train_dataset = load_dataset("webdataset", data_files=glob.glob(os.path.join(self.data_args.pixelprose_image_folder, 'commonpool_*/*.tar')), split="train", num_proc=128) # train_dataset = train_dataset.rename_column("jpg", "image") # train_dataset = train_dataset.add_column('type', len(train_dataset) * ['Pixelprose_I2T']) # train_dataset = train_dataset.remove_columns([col for col in train_dataset.column_names if not col in ( # ["image", "json", "type", "txt"])]) # list_data_dict.append(train_dataset) # ## text to image # train_dataset = load_dataset("webdataset", data_files=glob.glob(os.path.join(self.data_args.pixelprose_image_folder, 'commonpool_*/*.tar')), split="train", num_proc=128) # train_dataset = train_dataset.rename_column("jpg", "image") # train_dataset = train_dataset.add_column('type', len(train_dataset) * ['Pixelprose_T2I']) # train_dataset = train_dataset.remove_columns([col for col in train_dataset.column_names if not col in ( # ["image", "json", "type", "txt"])]) # list_data_dict.append(train_dataset) ## commonpool # train_dataset = load_dataset("webdataset", data_files=glob.glob(os.path.join(self.data_args.pixelprose_image_folder, 'commonpool_node0_*/*.tar')), split="train", num_proc=128) # train_dataset = train_dataset.rename_column("jpg", "image") # train_dataset = train_dataset.add_column('type', len(train_dataset) * ['Pixelprose_I2T']) # train_dataset = train_dataset.remove_columns([col for col in train_dataset.column_names if not col in ( # ["image", "json", "type", "txt"])]) # list_data_dict.append(train_dataset) # ## text to image # train_dataset = load_dataset("webdataset", data_files=glob.glob(os.path.join(self.data_args.pixelprose_image_folder, 'commonpool_node0_*/*.tar')), split="train", num_proc=128) # train_dataset = train_dataset.rename_column("jpg", "image") # train_dataset = train_dataset.add_column('type', len(train_dataset) * ['Pixelprose_T2I']) # train_dataset = train_dataset.remove_columns([col for col in train_dataset.column_names if not col in ( # ["image", "json", "type", "txt"])]) # list_data_dict.append(train_dataset) ## readcaps ## image to text train_dataset = load_dataset("webdataset", data_files=glob.glob(os.path.join(self.data_args.pixelprose_image_folder, 'redcaps_part0/*.tar')), split="train", num_proc=128) train_dataset = train_dataset.rename_column("jpg", "image") train_dataset = train_dataset.add_column('type', len(train_dataset) * ['Pixelprose_I2T']) train_dataset = train_dataset.remove_columns([col for col in train_dataset.column_names if not col in ( ["image", "json", "type", "txt"])]) list_data_dict.append(train_dataset) ## text to image train_dataset = load_dataset("webdataset", data_files=glob.glob(os.path.join(self.data_args.pixelprose_image_folder, 'redcaps_part0/*.tar')), split="train", num_proc=128) train_dataset = train_dataset.rename_column("jpg", "image") train_dataset = train_dataset.add_column('type', len(train_dataset) * ['Pixelprose_T2I']) train_dataset = train_dataset.remove_columns([col for col in train_dataset.column_names if not col in ( ["image", "json", "type", "txt"])]) list_data_dict.append(train_dataset) ## datacomp ## text to image # train_dataset = load_dataset("webdataset", data_files=glob.glob(os.path.join(self.data_args.datacomp_image_folder, '*.tar')), split="train", num_proc=128) # train_dataset = train_dataset.rename_column("jpg", "image") # train_dataset = train_dataset.add_column('type', len(train_dataset) * ['Datacomp_T2I']) # train_dataset = train_dataset.remove_columns(['json']) # train_dataset = train_dataset.add_column('json', len(train_dataset) * [json_with_none_values]) # train_dataset = train_dataset.remove_columns([col for col in train_dataset.column_names if not col in ( # ["image", "json", "type", "txt"])]) # list_data_dict.append(train_dataset) if len(list_data_dict) > 1: list_data_dict = concatenate_datasets(list_data_dict) else: list_data_dict = list_data_dict[0] # list_data_dict = list_data_dict.shuffle(seed=42) rank0_print(f"Totoal number of training instance: {len(list_data_dict)}") self.tokenizer = tokenizer self.list_data_dict = list_data_dict def __len__(self): return len(self.list_data_dict) @property def lengths(self): length_list = [] for sample in self.list_data_dict: img_tokens = 128 if 'image' in sample else 0 length_list.append(sum(len(conv['value'].split()) for conv in sample['conversations']) + img_tokens) return length_list @property def modality_lengths(self): length_list = [] for sample in self.list_data_dict: cur_len = sum(len(conv['value'].split()) for conv in sample['conversations']) cur_len = cur_len if 'image' in sample else -cur_len length_list.append(cur_len) return length_list def __getitem__(self, i) -> Dict[str, torch.Tensor]: print(i) while True: sources = self.list_data_dict[i] # Make the conversations based on the 'type' in 'sources' if sources['type'] == 'Pixelprose_T2I': sources['conversations'] = [ { "from": "human", "value": f"Please generate image based on the following detailed caption: {sources['json']['vlm_caption']}" }, { "from": "gpt", "value": "\n" } ] elif sources['type'] == 'Pixelprose_I2T': sources['conversations'] = [ { "from": "human", "value": f"\nProvide a detailed description of the given image." }, { "from": "gpt", "value": f"{sources['json']['vlm_caption']}" } ] elif sources['type'] == 'Datacomp_T2I': sources['conversations'] = [ { "from": "human", "value": f"Please generate image based on the following short caption: {sources['txt']}" }, { "from": "gpt", "value": "\n" } ] else: raise ValueError("Unknown source type. Please check the 'type' in 'sources'.") # if 'pixelprose' in sources['__url__'].lower(): # if random.choice([True, False]): # sources['conversations'] = [ # { # "from": "human", # "value": f"Please generate image based on the following detailed caption: {sources['json']['vlm_caption']}" # }, # { # "from": "gpt", # "value": "\n" # } # ] # else: # sources['conversations'] = [ # { # "from": "human", # "value": f"\nProvide a detailed description of the given image." # }, # { # "from": "gpt", # "value": f"{sources['json']['vlm_caption']}" # } # ] # elif 'datacomp' in sources['__url__'].lower(): # sources['conversations'] = [ # { # "from": "human", # "value": f"Please generate image based on the following short caption: {sources['json']['caption']}" # }, # { # "from": "gpt", # "value": "\n" # } # ] # else: # print(sources) # raise ValueError("Unknown data type. Please check the '__url__' in 'sources'.") if 'image' in sources: def img_process(images, processor, image_aspect_ratio): if image_aspect_ratio == 'pad': def expand2square(pil_img, background_color): width, height = pil_img.size if width == height: return pil_img elif width > height: result = Image.new(pil_img.mode, (width, width), background_color) result.paste(pil_img, (0, (width - height) // 2)) return result else: result = Image.new(pil_img.mode, (height, height), background_color) result.paste(pil_img, ((height - width) // 2, 0)) return result images = [expand2square(img, tuple(int(x * 255) for x in processor.image_mean)) for img in images] images = processor.preprocess(images, return_tensors='pt')['pixel_values'] else: images = processor.preprocess(images, return_tensors='pt')['pixel_values'] return images image_files = self.list_data_dict[i]['image'] if not isinstance(image_files, list): image_files = [image_files] images = [] for img in image_files: try: images.append(img.convert('RGB')) except Exception as e: print(f"Error opening image {img}: {e}") images=None break # Skip to the next image if there's an error if not images is None: try: temp = img_process(images, self.data_args.gen_image_processor, self.data_args.image_aspect_ratio) temp = img_process(images, self.data_args.image_processor, self.data_args.image_aspect_ratio) except ValueError as e: print(f"Error wrong number of channels: {e}") images=None # If no valid images were found, randomly pick another item if images is None: print(f"warning false image!!!!!!") i = random.randint(0, len(self.list_data_dict) - 1) continue # Processor logic sources, inst_type = preprocess_multimodal( copy.deepcopy([sources["conversations"]]), self.data_args) else: sources = copy.deepcopy([sources["conversations"]]) data_dict = preprocess( sources, self.tokenizer, has_image=('image' in self.list_data_dict[i])) if isinstance(i, int): data_dict = dict(input_ids=data_dict["input_ids"][0], labels=data_dict["labels"][0]) # image exist in the data if 'image' in self.list_data_dict[i]: if inst_type == 'gen': data_dict['gen_image'] = img_process(images, self.data_args.gen_image_processor, self.data_args.image_aspect_ratio) elif inst_type == 'und': data_dict['und_image'] = img_process(images, self.data_args.image_processor, self.data_args.image_aspect_ratio) elif self.data_args.is_multimodal: crop_size = self.data_args.image_processor.crop_size data_dict['image'] = torch.zeros(3, crop_size['height'], crop_size['width']) data_dict['ids'] = self.list_data_dict[i]['id'] if 'id' in self.list_data_dict[i] else 'unk' return data_dict class LazySupervisedJsonDataset(Dataset): """Dataset for supervised fine-tuning.""" def __init__(self, data_path: str, tokenizer: transformers.PreTrainedTokenizer, data_args: DataArguments): super(LazySupervisedJsonDataset, self).__init__() list_data_dict_raw = json.load(open(data_path, "r")) # filter out instance that does not have images list_data_dict = [] for inst in list_data_dict_raw: if not 'image' in inst: continue num_img_token = 0 for conv in inst['conversations']: num_img_token += conv['value'].count('') if not num_img_token == 1: continue list_data_dict.append(inst) rank0_print(f"Totoal number of training instance: {len(list_data_dict)}, filtered {len(list_data_dict_raw) - len(list_data_dict)}") rank0_print("Formatting inputs...Skip in lazy mode") self.tokenizer = tokenizer self.list_data_dict = list_data_dict self.data_args = data_args def __len__(self): return len(self.list_data_dict) @property def lengths(self): length_list = [] for sample in self.list_data_dict: img_tokens = 128 if 'image' in sample else 0 length_list.append(sum(len(conv['value'].split()) for conv in sample['conversations']) + img_tokens) return length_list @property def modality_lengths(self): length_list = [] for sample in self.list_data_dict: cur_len = sum(len(conv['value'].split()) for conv in sample['conversations']) cur_len = cur_len if 'image' in sample else -cur_len length_list.append(cur_len) return length_list def __getitem__(self, i) -> Dict[str, torch.Tensor]: while True: sources = self.list_data_dict[i] if isinstance(i, int): sources = [sources] assert len(sources) == 1, "Don't know why it is wrapped to a list" # FIXME if 'image' in sources[0]: def img_process(images, processor, image_aspect_ratio): if image_aspect_ratio == 'pad': def expand2square(pil_img, background_color): width, height = pil_img.size if width == height: return pil_img elif width > height: result = Image.new(pil_img.mode, (width, width), background_color) result.paste(pil_img, (0, (width - height) // 2)) return result else: result = Image.new(pil_img.mode, (height, height), background_color) result.paste(pil_img, ((height - width) // 2, 0)) return result images = [expand2square(img, tuple(int(x * 255) for x in processor.image_mean)) for img in images] images = processor.preprocess(images, return_tensors='pt')['pixel_values'] else: images = processor.preprocess(images, return_tensors='pt')['pixel_values'] return images image_files = self.list_data_dict[i]['image'] if isinstance(image_files, str): image_files = [image_files] image_folder = self.data_args.image_folder images = [] for img in image_files: try: image_path = os.path.join(image_folder, img) img = Image.open(image_path).convert('RGB') images.append(img) except Exception as e: print(f"Error opening image {img}: {e}") images=None break # Skip to the next image if there's an error if not images is None: try: temp = img_process(images, self.data_args.gen_image_processor, self.data_args.image_aspect_ratio) temp = img_process(images, self.data_args.image_processor, self.data_args.image_aspect_ratio) except ValueError as e: print(f"Error wrong number of channels: {e}") images=None # If no valid images were found, randomly pick another item if images is None: print(f"warning false image!!!!!!") i = random.randint(0, len(self.list_data_dict) - 1) continue # Processor logic sources, inst_type = preprocess_multimodal( copy.deepcopy([e["conversations"] for e in sources]), self.data_args) else: sources = copy.deepcopy([e["conversations"] for e in sources]) data_dict = preprocess( sources, self.tokenizer, has_image=('image' in self.list_data_dict[i])) if isinstance(i, int): data_dict = dict(input_ids=data_dict["input_ids"][0], labels=data_dict["labels"][0]) # image exist in the data if 'image' in self.list_data_dict[i]: if inst_type == 'gen': data_dict['gen_image'] = img_process(images, self.data_args.gen_image_processor, self.data_args.image_aspect_ratio) elif inst_type == 'und': data_dict['und_image'] = img_process(images, self.data_args.image_processor, self.data_args.image_aspect_ratio) elif self.data_args.is_multimodal: crop_size = self.data_args.image_processor.crop_size data_dict['image'] = torch.zeros(3, crop_size['height'], crop_size['width']) data_dict['ids'] = self.list_data_dict[i]['id'] if 'id' in self.list_data_dict[i] else 'unk' return data_dict @dataclass class DataCollatorForSupervisedDataset(object): """Collate examples for supervised fine-tuning.""" tokenizer: transformers.PreTrainedTokenizer def __call__(self, instances: Sequence[Dict]) -> Dict[str, torch.Tensor]: input_ids, labels, ids = tuple([instance[key] for instance in instances] for key in ("input_ids", "labels", "ids")) input_ids = torch.nn.utils.rnn.pad_sequence( input_ids, batch_first=True, padding_value=self.tokenizer.pad_token_id) labels = torch.nn.utils.rnn.pad_sequence(labels, batch_first=True, padding_value=IGNORE_INDEX) if input_ids.shape[1] > self.tokenizer.model_max_length: print(f"Warning input with length {input_ids.shape[1]} is longer than max length {self.tokenizer.model_max_length}") input_ids = input_ids[:, :self.tokenizer.model_max_length] labels = labels[:, :self.tokenizer.model_max_length] batch = dict( input_ids=input_ids, labels=labels, attention_mask=input_ids.ne(self.tokenizer.pad_token_id), ) batch_gen_images = [] batch_und_images = [] # if 'gen_image' in instances[0]: for instance in instances: # batch_gen_images.append(instance['gen_image']) # rank0_print(f"instance.keys() {instance.keys()}") if 'gen_image' in instance: # print(f"instance['gen_image'] {instance['gen_image']}") batch_gen_images.append(instance['gen_image']) # print(f"batch_gen_images {batch_gen_images}") if len(batch_gen_images) > 0: if all(x is not None and y.shape == batch_gen_images[0][0].shape for x in batch_gen_images for y in x): # rank0_print(f"[images for images in batch_gen_images] {[images for images in batch_gen_images]}") batch['gen_image'] = torch.cat( [images for images in batch_gen_images], dim=0) # rank0_print(batch['gen_image'].shape) # rank0_print(f"batch['gen_image'], {batch['gen_image']}") else: batch['gen_image'] = batch_gen_images else: batch['gen_image'] = None # if 'und_image' in instances[0]: for instance in instances: # rank0_print(f"instance.keys() {instance.keys()}") if 'und_image' in instance: # print(f"instance['und_image'] {instance['und_image']}") batch_und_images.append(instance['und_image']) # print(f"batch_und_images {batch_und_images}") if len(batch_und_images) > 0: if all(x is not None and y.shape == batch_und_images[0][0].shape for x in batch_und_images for y in x): batch['und_image'] = torch.cat( [images for images in batch_und_images], dim=0) # rank0_print(f"batch['und_image'], {batch['und_image']}") else: batch['und_image'] = batch_und_images else: batch['und_image'] = None batch['ids'] = ids return batch def make_supervised_data_module(tokenizer: transformers.PreTrainedTokenizer, data_args) -> Dict: """Make dataset and collator for supervised fine-tuning.""" if data_args.data_type == 'webdataset': train_dataset = LazySupervisedWebDataset(tokenizer=tokenizer, data_path=data_args.data_path, data_args=data_args) elif data_args.data_type == 'json': train_dataset = LazySupervisedJsonDataset(tokenizer=tokenizer, data_path=data_args.data_path, data_args=data_args) else: raise ValueError("Unknown data type. Please check the Dataloader type.") data_collator = DataCollatorForSupervisedDataset(tokenizer=tokenizer) return dict(train_dataset=train_dataset, eval_dataset=None, data_collator=data_collator) def unlock_vit(training_args, model_args, vision_tower): for n, p in vision_tower.named_parameters(): p.requires_grad = True def train(attn_implementation=None): global local_rank parser = transformers.HfArgumentParser( (ModelArguments, DataArguments, TrainingArguments)) model_args, data_args, training_args = parser.parse_args_into_dataclasses() print(model_args, data_args, training_args) local_rank = training_args.local_rank compute_dtype = (torch.float16 if training_args.fp16 else (torch.bfloat16 if training_args.bf16 else torch.float32)) bnb_model_from_pretrained_args = {} if training_args.bits in [4, 8]: from transformers import BitsAndBytesConfig bnb_model_from_pretrained_args.update(dict( device_map={"": training_args.device}, load_in_4bit=training_args.bits == 4, load_in_8bit=training_args.bits == 8, quantization_config=BitsAndBytesConfig( load_in_4bit=training_args.bits == 4, load_in_8bit=training_args.bits == 8, llm_int8_skip_modules=["mm_projector"], llm_int8_threshold=6.0, llm_int8_has_fp16_weight=False, bnb_4bit_compute_dtype=compute_dtype, bnb_4bit_use_double_quant=training_args.double_quant, bnb_4bit_quant_type=training_args.quant_type # {'fp4', 'nf4'} ) )) if model_args.vision_tower is not None: if 'mpt' in model_args.model_name_or_path: config = transformers.AutoConfig.from_pretrained(model_args.model_name_or_path, trust_remote_code=True) config.attn_config['attn_impl'] = training_args.mpt_attn_impl model = LlavaMptForCausalLM.from_pretrained( model_args.model_name_or_path, config=config, cache_dir=training_args.cache_dir, **bnb_model_from_pretrained_args ) else: model = LlavaLlamaForCausalLM.from_pretrained( model_args.model_name_or_path, cache_dir=training_args.cache_dir, attn_implementation=attn_implementation, torch_dtype=(torch.bfloat16 if training_args.bf16 else None), **bnb_model_from_pretrained_args ) else: model = transformers.LlamaForCausalLM.from_pretrained( model_args.model_name_or_path, cache_dir=training_args.cache_dir, attn_implementation=attn_implementation, torch_dtype=(torch.bfloat16 if training_args.bf16 else None), **bnb_model_from_pretrained_args ) model.config.use_cache = False if model_args.freeze_backbone: model.model.requires_grad_(False) if training_args.bits in [4, 8]: from peft import prepare_model_for_kbit_training model.config.torch_dtype = ( torch.float32 if training_args.fp16 else (torch.bfloat16 if training_args.bf16 else torch.float32)) model = prepare_model_for_kbit_training(model, use_gradient_checkpointing=training_args.gradient_checkpointing) if training_args.gradient_checkpointing: if hasattr(model, "enable_input_require_grads"): model.enable_input_require_grads() else: def make_inputs_require_grad(module, input, output): output.requires_grad_(True) model.get_input_embeddings().register_forward_hook(make_inputs_require_grad) if training_args.lora_enable: from peft import LoraConfig, get_peft_model lora_config = LoraConfig( r=training_args.lora_r, lora_alpha=training_args.lora_alpha, target_modules=find_all_linear_names(model), lora_dropout=training_args.lora_dropout, bias=training_args.lora_bias, task_type="CAUSAL_LM", ) if training_args.bits == 16: if training_args.bf16: model.to(torch.bfloat16) if training_args.fp16: model.to(torch.float16) rank0_print("Adding LoRA adapters...") model = get_peft_model(model, lora_config) if 'mpt' in model_args.model_name_or_path: tokenizer = transformers.AutoTokenizer.from_pretrained( model_args.model_name_or_path, cache_dir=training_args.cache_dir, model_max_length=training_args.model_max_length, padding_side="right" ) else: tokenizer = transformers.AutoTokenizer.from_pretrained( model_args.model_name_or_path, cache_dir=training_args.cache_dir, model_max_length=training_args.model_max_length, padding_side="right", use_fast=False, ) if model_args.version == "v0": if tokenizer.pad_token is None: print(f"Adding pad token as '[PAD]'") smart_tokenizer_and_embedding_resize( special_tokens_dict=dict(pad_token="[PAD]"), tokenizer=tokenizer, model=model, ) elif model_args.version == "v0.5": tokenizer.pad_token = tokenizer.unk_token elif model_args.version == "v1": tokenizer.pad_token = tokenizer.unk_token if model_args.version in conversation_lib.conv_templates: conversation_lib.default_conversation = conversation_lib.conv_templates[model_args.version] else: conversation_lib.default_conversation = conversation_lib.conv_templates["vicuna_v1"] else: # tokenizer.pad_token = tokenizer.unk_token if tokenizer.pad_token is None: smart_tokenizer_and_embedding_resize( special_tokens_dict=dict(pad_token="",additional_special_tokens=['[IMG]','[/IMG]','']), tokenizer=tokenizer, model=model, ) elif not '' in tokenizer.get_added_vocab(): smart_tokenizer_and_embedding_resize( special_tokens_dict=dict(additional_special_tokens=['[IMG]','[/IMG]','']), tokenizer=tokenizer, model=model, ) if model_args.version in conversation_lib.conv_templates: conversation_lib.default_conversation = conversation_lib.conv_templates[model_args.version] else: conversation_lib.default_conversation = conversation_lib.conv_templates["llama3"] rank0_print(f"Using conversation format: {conversation_lib.default_conversation.version}") if model_args.vision_tower is not None: model.get_model().initialize_vision_modules( model_args=model_args, fsdp=training_args.fsdp ) vision_tower = model.get_vision_tower() vision_tower.to(dtype=torch.bfloat16 if training_args.bf16 else torch.float16, device=training_args.device) vision_tower.requires_grad_(False) ## generation vision tower gen_vision_tower = model.get_gen_vision_tower() gen_vision_tower.to(dtype=torch.bfloat16 if training_args.bf16 else torch.float16, device=training_args.device) gen_vision_tower.requires_grad_(False) data_args.image_processor = vision_tower.image_processor data_args.gen_image_processor = gen_vision_tower.image_processor data_args.is_multimodal = True data_args.n_query = model_args.n_query data_args.n_und_query = model_args.n_und_query model.config.image_aspect_ratio = data_args.image_aspect_ratio model.config.tokenizer_padding_side = tokenizer.padding_side model.config.tokenizer_model_max_length = tokenizer.model_max_length model.config.tune_mm_mlp_adapter = training_args.tune_mm_mlp_adapter = model_args.tune_mm_mlp_adapter if model_args.tune_mm_mlp_adapter: model.requires_grad_(False) for p in model.get_model().mm_projector.parameters(): p.requires_grad = True for p in model.get_model().gen_projector.parameters(): p.requires_grad = True for p in model.get_model().down_projector.parameters(): p.requires_grad = True model.config.freeze_mm_mlp_adapter = training_args.freeze_mm_mlp_adapter if training_args.freeze_mm_mlp_adapter: for p in model.get_model().mm_projector.parameters(): p.requires_grad = False # Calculate total parameters and trainable parameters total_params = sum(p.numel() for p in model.get_model().parameters()) trainable_params = sum(p.numel() for p in model.get_model().parameters() if p.requires_grad) print(f"Total parameters: {total_params}") print(f"Trainable parameters: {trainable_params}") if training_args.bits in [4, 8]: model.get_model().mm_projector.to(dtype=compute_dtype, device=training_args.device) model.config.mm_use_im_start_end = data_args.mm_use_im_start_end = model_args.mm_use_im_start_end model.config.mm_projector_lr = training_args.mm_projector_lr training_args.use_im_start_end = model_args.mm_use_im_start_end model.config.mm_use_im_patch_token = model_args.mm_use_im_patch_token model.initialize_vision_tokenizer(model_args, tokenizer=tokenizer) model.config.pad_token_id = tokenizer.pad_token_id if training_args.bits in [4, 8]: from peft.tuners.lora import LoraLayer for name, module in model.named_modules(): if isinstance(module, LoraLayer): if training_args.bf16: module = module.to(torch.bfloat16) if 'norm' in name: module = module.to(torch.float32) if 'lm_head' in name or 'embed_tokens' in name: if hasattr(module, 'weight'): if training_args.bf16 and module.weight.dtype == torch.float32: module = module.to(torch.bfloat16) data_module = make_supervised_data_module(tokenizer=tokenizer, data_args=data_args) trainer = LLaVATrainer(model=model, tokenizer=tokenizer, args=training_args, **data_module) if list(pathlib.Path(training_args.output_dir).glob("checkpoint-*")): trainer.train(resume_from_checkpoint=True) else: trainer.train() trainer.save_state() model.config.use_cache = True if training_args.lora_enable: state_dict = get_peft_state_maybe_zero_3( model.named_parameters(), training_args.lora_bias ) non_lora_state_dict = get_peft_state_non_lora_maybe_zero_3( model.named_parameters() ) if training_args.local_rank == 0 or training_args.local_rank == -1: model.config.save_pretrained(training_args.output_dir) model.save_pretrained(training_args.output_dir, state_dict=state_dict) torch.save(non_lora_state_dict, os.path.join(training_args.output_dir, 'non_lora_trainables.bin')) else: safe_save_model_for_hf_trainer(trainer=trainer, output_dir=training_args.output_dir, vision_tower=model_args.vision_tower) if __name__ == "__main__": train()