from dataclasses import dataclass, field from typing import Dict, Optional, Sequence import logging import os, sys import copy import torch import transformers from transformers import LlamaForCausalLM, LlamaTokenizer from torch.utils.data import Dataset from transformers import Trainer sys.path.append(os.path.dirname(__file__)) sys.path.append(os.path.dirname(os.path.dirname(__file__))) from utils.special_tok_llama2 import ( B_CODE, E_CODE, B_RESULT, E_RESULT, B_INST, E_INST, B_SYS, E_SYS, DEFAULT_PAD_TOKEN, DEFAULT_BOS_TOKEN, DEFAULT_EOS_TOKEN, DEFAULT_UNK_TOKEN, IGNORE_INDEX, ) from conversation_template import json_to_code_result_tok_temp @dataclass class ModelArguments: model_name_or_path: Optional[str] = field(default="./ckpt/llama-2-13b-chat") peft: bool = field(default=False) @dataclass class DataArguments: data_path: str = field( default=None, metadata={"help": "Path to the training data."} ) @dataclass class TrainingArguments(transformers.TrainingArguments): cache_dir: Optional[str] = field(default=None) optim: str = field(default="adamw_torch") model_max_length: int = field( default=4096, metadata={ "help": "Maximum sequence length. Sequences will be right padded (and possibly truncated)." }, ) def create_peft_config(model): from peft import ( get_peft_model, LoraConfig, TaskType, prepare_model_for_int8_training, ) peft_config = LoraConfig( task_type=TaskType.CAUSAL_LM, inference_mode=False, r=8, lora_alpha=16, lora_dropout=0.05, target_modules=["q_proj", "v_proj"], ) # prepare int-8 model for training model = prepare_model_for_int8_training(model) model = get_peft_model(model, peft_config) model.print_trainable_parameters() print(f"Using Peft") return model, peft_config 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 = [tokenized.input_ids[0] for tokenized in tokenized_list] input_ids_lens = [ tokenized.input_ids.ne(tokenizer.pad_token_id).sum().item() for tokenized in tokenized_list ] return dict( input_ids=input_ids, input_ids_lens=input_ids_lens, ) def find_all_sublist_end(main_list, sublist): """Find all the ending indices of a sublist in a main list.""" sublist_len = len(sublist) main_list = main_list.tolist() indices = [] for index in (i for i, e in enumerate(main_list) if e == sublist[0]): if main_list[index : index + sublist_len] == sublist: indices.append(index + sublist_len) return indices def find_all_sublist_start(main_list, sublist): """Find all the starting indices of a sublist in a main list.""" sublist_len = len(sublist) main_list = main_list.tolist() indices = [] for index in (i for i, e in enumerate(main_list) if e == sublist[0]): if main_list[index : index + sublist_len] == sublist: indices.append(index) return indices def preprocess( trajs: Sequence[str], tokenizer: transformers.PreTrainedTokenizer, ) -> Dict: INST_START_INDEX = tokenizer.encode(f"{B_INST}")[-1] INST_END_INDEX = tokenizer.encode(f"{E_INST}")[-1] RESULT_START_INDEX = tokenizer.encode(f"{B_RESULT}")[-1] RESULT_END_INDEX = tokenizer.encode(f"{E_RESULT}")[-1] """Preprocess the data by tokenizing.""" examples_tokenized = _tokenize_fn(trajs, tokenizer) input_ids_lens = examples_tokenized["input_ids_lens"] input_ids = examples_tokenized["input_ids"] # [torch.tensor , torch.tensor , ...] labels = copy.deepcopy(input_ids) # IGNORE INDEX SET for i, label in enumerate(labels): user_start_inds = find_all_sublist_start(label, [INST_START_INDEX]) assistant_start_inds = find_all_sublist_end(label, [INST_END_INDEX]) result_start_inds = find_all_sublist_start(label, [RESULT_START_INDEX]) result_end_inds = find_all_sublist_end(label, [RESULT_END_INDEX]) # for debug # for len_i, ind in enumerate(label): # print(f'{len_i}|{ind} -> "{tokenizer.decode(ind)}"') assert len(user_start_inds) == len( assistant_start_inds ), f"User and Assistant pair should be equal :: \n\tUser [{user_start_inds}]/\n\tAssistant [{assistant_start_inds}]\n\n Text : \n{trajs[i]}" assert len(result_start_inds) == len( result_end_inds ), f"Start and End indices pairs do not match.: : \nText : \n{trajs[i]}" for user_start_ind, assistant_start_ind in zip( user_start_inds, assistant_start_inds ): label[user_start_ind + 1 : assistant_start_ind - 1] = IGNORE_INDEX for start, end in zip(result_start_inds, result_end_inds): label[start + 1 : end - 1] = IGNORE_INDEX # cut max length input_ids = [i[:1500] for i in input_ids] labels = [i[:1500] for i in labels] return dict(input_ids=input_ids, labels=labels) class SupervisedDataset(Dataset): """Dataset for supervised fine-tuning.""" def __init__(self, data_path: str, tokenizer: transformers.PreTrainedTokenizer): super(SupervisedDataset, self).__init__() logging.warning(f"Loading data from data path : {data_path}") all_json = os.listdir(data_path) trajs = list() for json_file_name in all_json: traj = json_to_code_result_tok_temp(json_file_name=json_file_name) trajs.append(traj) logging.warning("Tokenizing inputs... This may take some time...") data_dict = preprocess(trajs, tokenizer) self.input_ids = data_dict["input_ids"] self.labels = data_dict["labels"] def __len__(self): return len(self.input_ids) def __getitem__(self, i) -> Dict[str, torch.Tensor]: return dict(input_ids=self.input_ids[i], labels=self.labels[i]) @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 = tuple( [instance[key] for instance in instances] for key in ("input_ids", "labels") ) 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 ) return dict( input_ids=input_ids, labels=labels, attention_mask=input_ids.ne(self.tokenizer.pad_token_id), ) def make_supervised_data_module( tokenizer: transformers.PreTrainedTokenizer, data_args ) -> Dict: """Make dataset and collator for supervised fine-tuning.""" train_dataset = SupervisedDataset( tokenizer=tokenizer, data_path=data_args.data_path ) data_collator = DataCollatorForSupervisedDataset(tokenizer=tokenizer) return dict( train_dataset=train_dataset, eval_dataset=None, data_collator=data_collator ) def build_model_from_hf_path( hf_model_path: str = "./ckpt/llama-2-13b-chat", peft: bool = False ): # build tokenizer tokenizer = LlamaTokenizer.from_pretrained( hf_model_path, padding_side="right", use_fast=False, ) special_tokens_dict = dict() if tokenizer.pad_token is None: special_tokens_dict["pad_token"] = DEFAULT_PAD_TOKEN # 32000 if tokenizer.eos_token is None: special_tokens_dict["eos_token"] = DEFAULT_EOS_TOKEN # 2 if tokenizer.bos_token is None: special_tokens_dict["bos_token"] = DEFAULT_BOS_TOKEN # 1 if tokenizer.unk_token is None: special_tokens_dict["unk_token"] = DEFAULT_UNK_TOKEN tokenizer.add_special_tokens(special_tokens_dict) tokenizer.add_tokens( [ B_CODE, # 32001 E_CODE, # 32002 B_RESULT, # 32003 E_RESULT, # 32004 B_INST, E_INST, B_SYS, E_SYS, # 32008 ], special_tokens=True, ) # build model if peft: model = LlamaForCausalLM.from_pretrained( hf_model_path, load_in_8bit=True, device_map="auto", ignore_mismatched_sizes=True, torch_dtype=torch.float16, ) else: # for llama # model = LlamaForCausalLM.from_pretrained( # hf_model_path, ignore_mismatched_sizes=True # ) # for codellama from codellama_wrapper import CodeLlamaForCausalLM model = CodeLlamaForCausalLM.from_pretrained(hf_model_path) model.resize_token_embeddings(len(tokenizer)) return {"tokenizer": tokenizer, "model": model} def train(): parser = transformers.HfArgumentParser( (ModelArguments, DataArguments, TrainingArguments) ) model_args, data_args, training_args = parser.parse_args_into_dataclasses() model_dict = build_model_from_hf_path( hf_model_path=model_args.model_name_or_path, peft=model_args.peft ) model, tokenizer = model_dict["model"], model_dict["tokenizer"] # peft setting model.train() if model_args.peft: model, lora_config = create_peft_config(model) # make dataset data_module = make_supervised_data_module(tokenizer=tokenizer, data_args=data_args) trainer = Trainer( model=model, tokenizer=tokenizer, args=training_args, **data_module ) # train trainer.train() trainer.save_state() trainer.save_model(output_dir=training_args.output_dir) if __name__ == "__main__": train()