import ast import gc import os from enum import Enum from itertools import chain import requests import torch from accelerate.state import PartialState from datasets import load_dataset, load_from_disk from huggingface_hub import HfApi from peft import LoraConfig, PeftModel, get_peft_model, prepare_model_for_kbit_training from transformers import AutoConfig, AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig from autotrain import is_unsloth_available, logger from autotrain.trainers.clm.callbacks import LoadBestPeftModelCallback, SavePeftModelCallback from autotrain.trainers.common import ( ALLOW_REMOTE_CODE, LossLoggingCallback, TrainStartCallback, UploadLogs, pause_space, remove_autotrain_data, save_training_params, ) DEFAULT_CHAT_TEMPLATE = "{% if not add_generation_prompt is defined %}{% set add_generation_prompt = false %}{% endif %}{% for message in messages %}{{'<|im_start|>' + message['role'] + '\n' + message['content'] + '<|im_end|>' + '\n'}}{% endfor %}{% if add_generation_prompt %}{{ '<|im_start|>assistant\n' }}{% endif %}" CHATML_CHAT_TEMPLATE = "{% for message in messages %}\n{{'<|im_start|>' + message['role'] + '\n' + message['content'] + '<|im_end|>' + '\n'}}{% if loop.last and add_generation_prompt %}{{'<|im_start|>assistant\n' }}{% endif %}{% endfor %}" ZEPHYR_CHAT_TEMPLATE = "{% for message in messages %}\n{% if message['role'] == 'user' %}\n{{ '<|user|>\n' + message['content'] + eos_token }}\n{% elif message['role'] == 'system' %}\n{{ '<|system|>\n' + message['content'] + eos_token }}\n{% elif message['role'] == 'assistant' %}\n{{ '<|assistant|>\n' + message['content'] + eos_token }}\n{% endif %}\n{% if loop.last and add_generation_prompt %}\n{{ '<|assistant|>' }}\n{% endif %}\n{% endfor %}" IGNORE_INDEX = -100 DEFAULT_PAD_TOKEN = "[PAD]" DEFAULT_EOS_TOKEN = "" DEFAULT_BOS_TOKEN = "" DEFAULT_UNK_TOKEN = "" TARGET_MODULES = { "Salesforce/codegen25-7b-multi": "q_proj,k_proj,v_proj,o_proj,down_proj,up_proj,gate_proj", } MODEL_CARD = """ --- tags: - autotrain - text-generation-inference - text-generation{peft} library_name: transformers{base_model} widget: - messages: - role: user content: What is your favorite condiment? license: other{dataset_tag} --- # Model Trained Using AutoTrain This model was trained using AutoTrain. For more information, please visit [AutoTrain](https://hf.co/docs/autotrain). # Usage ```python from transformers import AutoModelForCausalLM, AutoTokenizer model_path = "PATH_TO_THIS_REPO" tokenizer = AutoTokenizer.from_pretrained(model_path) model = AutoModelForCausalLM.from_pretrained( model_path, device_map="auto", torch_dtype='auto' ).eval() # Prompt content: "hi" messages = [ {{"role": "user", "content": "hi"}} ] input_ids = tokenizer.apply_chat_template(conversation=messages, tokenize=True, add_generation_prompt=True, return_tensors='pt') output_ids = model.generate(input_ids.to('cuda')) response = tokenizer.decode(output_ids[0][input_ids.shape[1]:], skip_special_tokens=True) # Model response: "Hello! How can I assist you today?" print(response) ``` """ class ZephyrSpecialTokens(str, Enum): USER = "<|user|>" ASSISTANT = "<|assistant|>" SYSTEM = "<|system|>" EOS_TOKEN = "" BOS_TOKEN = "" PAD_TOKEN = "" @classmethod def list(cls): return [c.value for c in cls] class ChatmlSpecialTokens(str, Enum): USER = "<|im_start|>user" ASSISTANT = "<|im_start|>assistant" SYSTEM = "<|im_start|>system" EOS_TOKEN = "<|im_end|>" BOS_TOKEN = "" PAD_TOKEN = "" @classmethod def list(cls): return [c.value for c in cls] def preprocess_reward(examples, tokenizer): """ Preprocesses the reward data by tokenizing the chosen and rejected examples. Args: examples (dict): A dictionary containing two keys, "chosen" and "rejected", each mapping to a list of text examples. tokenizer (PreTrainedTokenizer): A tokenizer instance from the Hugging Face library used to tokenize the text examples. Returns: dict: A dictionary with the following keys: - "input_ids_chosen": List of tokenized input IDs for the chosen examples. - "attention_mask_chosen": List of attention masks for the chosen examples. - "input_ids_rejected": List of tokenized input IDs for the rejected examples. - "attention_mask_rejected": List of attention masks for the rejected examples. """ new_examples = { "input_ids_chosen": [], "attention_mask_chosen": [], "input_ids_rejected": [], "attention_mask_rejected": [], } for chosen, rejected in zip(examples["chosen"], examples["rejected"]): tokenized_chosen = tokenizer(chosen, truncation=True) tokenized_rejected = tokenizer(rejected, truncation=True) new_examples["input_ids_chosen"].append(tokenized_chosen["input_ids"]) new_examples["attention_mask_chosen"].append(tokenized_chosen["attention_mask"]) new_examples["input_ids_rejected"].append(tokenized_rejected["input_ids"]) new_examples["attention_mask_rejected"].append(tokenized_rejected["attention_mask"]) return new_examples def get_target_modules(config): """ Determines the target modules based on the provided configuration. Args: config (object): Configuration object that contains the following attributes: - target_modules (str or None): Specifies the target modules. It can be: - None: Returns the default target modules for the model specified in the config. - An empty string: Returns the default target modules for the model specified in the config. - "all-linear": Returns the string "all-linear". - A comma-separated string: Returns a list of target modules split by commas. Returns: list or str: A list of target modules or a specific string ("all-linear") based on the configuration. """ if config.target_modules is None: return TARGET_MODULES.get(config.model) if config.target_modules.strip() == "": return TARGET_MODULES.get(config.model) if config.target_modules.strip().lower() == "all-linear": return "all-linear" return config.target_modules.split(",") def group_texts(examples, config): """ Groups texts into chunks of a specified block size. Args: examples (dict): A dictionary where keys are feature names and values are lists of lists containing text data. config (object): A configuration object that contains the block_size attribute. Returns: dict: A dictionary with the same keys as the input examples, where each value is a list of chunks of text data. Additionally, a "labels" key is added with the same value as the "input_ids" key. """ # Concatenate all texts. concatenated_examples = {k: list(chain(*examples[k])) for k in examples.keys()} total_length = len(concatenated_examples[list(examples.keys())[0]]) # We drop the small remainder, we could add padding if the model supported it instead of this drop, you can # customize this part to your needs. if total_length >= config.block_size: total_length = (total_length // config.block_size) * config.block_size else: total_length = 0 # Split by chunks of max_len. result = { k: [t[i : i + config.block_size] for i in range(0, total_length, config.block_size)] for k, t in concatenated_examples.items() } result["labels"] = result["input_ids"].copy() return result def tokenize(examples, tokenizer, config): """ Tokenizes the input examples using the provided tokenizer and configuration. Args: examples (dict): A dictionary containing the input examples to be tokenized. tokenizer (PreTrainedTokenizer): The tokenizer to be used for tokenizing the examples. config (object): Configuration object that contains the text column name. Returns: dict: A dictionary containing the tokenized output. """ output = tokenizer(examples[config.text_column]) return output def merge_adapter(base_model_path, target_model_path, adapter_path): """ Merges an adapter into a base model and saves the resulting model and tokenizer. Args: base_model_path (str): Path to the base model directory. target_model_path (str): Path to the directory where the merged model and tokenizer will be saved. adapter_path (str): Path to the adapter model directory. Raises: RuntimeError: If resizing token embeddings fails without padding to a multiple of 8. """ logger.info("Loading adapter...") model = AutoModelForCausalLM.from_pretrained( base_model_path, torch_dtype=torch.float16, low_cpu_mem_usage=True, trust_remote_code=ALLOW_REMOTE_CODE, ) tokenizer = AutoTokenizer.from_pretrained( target_model_path, trust_remote_code=ALLOW_REMOTE_CODE, ) try: model.resize_token_embeddings(len(tokenizer)) model = PeftModel.from_pretrained(model, adapter_path) except RuntimeError: model.resize_token_embeddings(len(tokenizer), pad_to_multiple_of=8) model = PeftModel.from_pretrained(model, adapter_path) model = model.merge_and_unload() logger.info("Saving target model...") model.save_pretrained(target_model_path) tokenizer.save_pretrained(target_model_path) def create_model_card(config): """ Generates a model card string based on the provided configuration. Args: config (object): Configuration object with the following attributes: - peft (bool): Indicates if PEFT (Parameter-Efficient Fine-Tuning) is used. - data_path (str): Path to the dataset. - project_name (str): Name of the project. - model (str): Path or identifier of the model. Returns: str: A formatted model card string. """ if config.peft: peft = "\n- peft" else: peft = "" if config.data_path == f"{config.project_name}/autotrain-data" or os.path.isdir(config.data_path): dataset_tag = "" else: dataset_tag = f"\ndatasets:\n- {config.data_path}" if os.path.isdir(config.model): base_model = "" else: base_model = f"\nbase_model: {config.model}" model_card = MODEL_CARD.format( dataset_tag=dataset_tag, peft=peft, base_model=base_model, ) return model_card.strip() def pause_endpoint(params): """ Pauses a Hugging Face endpoint using the provided parameters. Args: params (object): An object containing the necessary parameters, including: - token (str): The authorization token to access the Hugging Face API. Returns: dict: The JSON response from the API call. Raises: KeyError: If the "ENDPOINT_ID" environment variable is not set. requests.exceptions.RequestException: If there is an issue with the API request. """ endpoint_id = os.environ["ENDPOINT_ID"] username = endpoint_id.split("/")[0] project_name = endpoint_id.split("/")[1] api_url = f"https://api.endpoints.huggingface.cloud/v2/endpoint/{username}/{project_name}/pause" headers = {"Authorization": f"Bearer {params.token}"} r = requests.post(api_url, headers=headers, timeout=30) return r.json() def apply_chat_template( example, tokenizer, config, ): """ Applies a chat template to the given example based on the specified configuration. Args: example (dict): The input example containing the text data to be processed. tokenizer (object): The tokenizer to be used for applying the chat template. config (object): Configuration object containing the following attributes: - trainer (str): Specifies the type of trainer. Can be "default", "sft", "reward", "dpo", or "orpo". - text_column (str): The key in the example dict that contains the text data. - chat_template (str): Specifies the chat template to be used. Relevant for "reward" and "dpo" trainers. Returns: dict: The modified example with the chat template applied. Raises: ValueError: If the required keys are not found in the example for "reward", "dpo", or "orpo" trainers. """ # kudos to Hugging Face H4 Team for this snippet if config.trainer in ("default", "sft"): messages = example[config.text_column] if isinstance(messages, str): messages = ast.literal_eval(messages) example[config.text_column] = tokenizer.apply_chat_template( messages, tokenize=False, add_generation_prompt=False ) elif config.trainer == "reward": if all(k in example.keys() for k in ("chosen", "rejected")): chosen_messages = example["chosen"] rejected_messages = example["rejected"] if isinstance(chosen_messages, str): chosen_messages = ast.literal_eval(chosen_messages) if isinstance(rejected_messages, str): rejected_messages = ast.literal_eval(rejected_messages) if config.chat_template == "zephyr" and chosen_messages[0]["role"] != "system": chosen_messages.insert(0, {"role": "system", "content": ""}) if config.chat_template == "zephyr" and rejected_messages[0]["role"] != "system": rejected_messages.insert(0, {"role": "system", "content": ""}) example["chosen"] = tokenizer.apply_chat_template(chosen_messages, tokenize=False) example["rejected"] = tokenizer.apply_chat_template(rejected_messages, tokenize=False) else: raise ValueError( f"Could not format example as dialogue for `rm/orpo` task! Require `[chosen, rejected]` keys but found {list(example.keys())}" ) elif config.trainer in ("dpo", "orpo"): if all(k in example.keys() for k in ("chosen", "rejected")): # For DPO, the inputs are triples of (prompt, chosen, rejected), where `chosen` and `rejected` are the final turn of a dialogue # We therefore need to extract the N-1 turns to form the prompt if isinstance(example["chosen"], str): example["chosen"] = ast.literal_eval(example["chosen"]) if isinstance(example["rejected"], str): example["rejected"] = ast.literal_eval(example["rejected"]) prompt_messages = example["chosen"][:-1] if config.chat_template == "zephyr" and example["chosen"][0]["role"] != "system": prompt_messages.insert(0, {"role": "system", "content": ""}) chosen_messages = example["chosen"][-1:] rejected_messages = example["rejected"][-1:] example["chosen"] = tokenizer.apply_chat_template(chosen_messages, tokenize=False) example["rejected"] = tokenizer.apply_chat_template(rejected_messages, tokenize=False) example["prompt"] = tokenizer.apply_chat_template(prompt_messages, tokenize=False) else: raise ValueError( f"Could not format example as dialogue for `dpo` task! Require `[chosen, rejected]` keys but found {list(example.keys())}" ) return example def post_training_steps(config, trainer): """ Perform post-training steps including saving the model, creating a model card, merging adapter weights, and optionally pushing the model to the Hugging Face Hub. Args: config (object): Configuration object containing various settings and parameters. trainer (object): Trainer object used for training the model. Steps: 1. Save the trained model and set `use_cache` to True. 2. Create a model card and save it as README.md in the output directory. 3. If PEFT (Parameter-Efficient Fine-Tuning) and adapter merging are enabled: - Delete the trainer object and clear CUDA cache. - Merge adapter weights into the base model. - Remove adapter weight files from the output directory. 4. If pushing to the Hugging Face Hub is enabled: - Remove training data folder. - Push the model to the Hugging Face Hub repository. 5. Pause the space if the process index is 0. Raises: Exception: If merging adapter weights fails. """ logger.info("Finished training, saving model...") trainer.model.config.use_cache = True trainer.save_model(config.project_name) model_card = create_model_card(config) # save model card to output directory as README.md with open(f"{config.project_name}/README.md", "w", encoding="utf-8") as f: f.write(model_card) if config.peft and config.merge_adapter: del trainer gc.collect() torch.cuda.empty_cache() logger.info("Merging adapter weights...") try: merge_adapter( base_model_path=config.model, target_model_path=config.project_name, adapter_path=config.project_name, ) # remove adapter weights: adapter_* for file in os.listdir(config.project_name): if file.startswith("adapter_"): os.remove(f"{config.project_name}/{file}") except Exception as e: logger.warning(f"Failed to merge adapter weights: {e}") logger.warning("Skipping adapter merge. Only adapter weights will be saved.") if config.push_to_hub: if PartialState().process_index == 0: # remove data folder remove_autotrain_data(config) logger.info("Pushing model to hub...") save_training_params(config) api = HfApi(token=config.token) api.create_repo( repo_id=f"{config.username}/{config.project_name}", repo_type="model", private=True, exist_ok=True ) api.upload_folder( folder_path=config.project_name, repo_id=f"{config.username}/{config.project_name}", repo_type="model", ) if PartialState().process_index == 0: pause_space(config) def process_input_data(config): """ Processes input data based on the provided configuration. Args: config (object): Configuration object containing the following attributes: - data_path (str): Path to the dataset. - project_name (str): Name of the project. - train_split (str): Split name for training data. - valid_split (str, optional): Split name for validation data. - token (str, optional): Token for accessing the dataset. - text_column (str): Name of the text column. - rejected_text_column (str): Name of the rejected text column. - prompt_text_column (str): Name of the prompt text column. - trainer (str): Type of trainer (e.g., "dpo", "reward", "orpo"). Returns: tuple: A tuple containing: - train_data (Dataset): Processed training dataset. - valid_data (Dataset or None): Processed validation dataset if valid_split is provided, otherwise None. """ if config.data_path == f"{config.project_name}/autotrain-data": logger.info("loading dataset from disk") train_data = load_from_disk(config.data_path)[config.train_split] else: if ":" in config.train_split: dataset_config_name, split = config.train_split.split(":") train_data = load_dataset( config.data_path, name=dataset_config_name, split=split, token=config.token, trust_remote_code=ALLOW_REMOTE_CODE, ) else: train_data = load_dataset( config.data_path, split=config.train_split, token=config.token, trust_remote_code=ALLOW_REMOTE_CODE, ) # rename columns for reward trainer if config.trainer in ("dpo", "reward", "orpo"): if not (config.text_column == "chosen" and config.text_column in train_data.column_names): train_data = train_data.rename_column(config.text_column, "chosen") if not (config.rejected_text_column == "rejected" and config.rejected_text_column in train_data.column_names): train_data = train_data.rename_column(config.rejected_text_column, "rejected") if config.trainer in ("dpo", "orpo"): if not (config.prompt_text_column == "prompt" and config.prompt_text_column in train_data.column_names): train_data = train_data.rename_column(config.prompt_text_column, "prompt") if config.valid_split is not None: if config.data_path == f"{config.project_name}/autotrain-data": valid_data = load_from_disk(config.data_path)[config.valid_split] else: if ":" in config.valid_split: dataset_config_name, split = config.valid_split.split(":") valid_data = load_dataset( config.data_path, name=dataset_config_name, split=split, token=config.token, trust_remote_code=ALLOW_REMOTE_CODE, ) else: valid_data = load_dataset( config.data_path, split=config.valid_split, token=config.token, trust_remote_code=ALLOW_REMOTE_CODE, ) if config.trainer in ("dpo", "reward", "orpo"): if not (config.text_column == "chosen" and config.text_column in valid_data.column_names): valid_data = valid_data.rename_column(config.text_column, "chosen") if not ( config.rejected_text_column == "rejected" and config.rejected_text_column in valid_data.column_names ): valid_data = valid_data.rename_column(config.rejected_text_column, "rejected") if config.trainer in ("dpo", "reward"): if not (config.prompt_text_column == "prompt" and config.prompt_text_column in valid_data.column_names): valid_data = valid_data.rename_column(config.prompt_text_column, "prompt") else: valid_data = None logger.info(f"Train data: {train_data}") logger.info(f"Valid data: {valid_data}") return train_data, valid_data def get_tokenizer(config): """ Initializes and returns a tokenizer based on the provided configuration. Args: config (object): Configuration object containing the following attributes: - chat_template (str): The chat template type, either "chatml" or "zephyr". - model (str): The model identifier to load the tokenizer from. - token (str): The token to use for the tokenizer. - model_max_length (int): The maximum length of the model. - padding (str): The padding side, either "left" or "right". Returns: tokenizer (PreTrainedTokenizer): The initialized tokenizer with the specified configuration. """ special_tokens = None chat_template = None if config.chat_template == "chatml": special_tokens = ChatmlSpecialTokens chat_template = CHATML_CHAT_TEMPLATE elif config.chat_template == "zephyr": special_tokens = ZephyrSpecialTokens chat_template = ZEPHYR_CHAT_TEMPLATE if special_tokens is not None: tokenizer = AutoTokenizer.from_pretrained( config.model, pad_token=special_tokens.PAD_TOKEN.value, bos_token=special_tokens.BOS_TOKEN.value, eos_token=special_tokens.EOS_TOKEN.value, additional_special_tokens=special_tokens.list(), token=config.token, trust_remote_code=ALLOW_REMOTE_CODE, ) tokenizer.chat_template = chat_template else: tokenizer = AutoTokenizer.from_pretrained( config.model, token=config.token, trust_remote_code=ALLOW_REMOTE_CODE ) if tokenizer.chat_template is None: tokenizer.chat_template = DEFAULT_CHAT_TEMPLATE if tokenizer.model_max_length > 2048: tokenizer.model_max_length = config.model_max_length if getattr(tokenizer, "pad_token", None) is None: tokenizer.pad_token = tokenizer.eos_token if getattr(tokenizer, "pad_token_id", None) is None: tokenizer.pad_token_id = tokenizer.eos_token_id if config.padding in ("left", "right"): tokenizer.padding_side = config.padding return tokenizer def process_data_with_chat_template(config, tokenizer, train_data, valid_data): """ Processes training and validation data using a specified chat template. Args: config (object): Configuration object containing settings and parameters. tokenizer (object): Tokenizer object used for tokenizing the data. train_data (Dataset): Training dataset to be processed. valid_data (Dataset): Validation dataset to be processed. Returns: tuple: A tuple containing the processed training and validation datasets. Notes: - If `config.chat_template` is one of ("chatml", "zephyr", "tokenizer"), the chat template will be applied. - Logs information about the application of the chat template. - For ORPO/DPO, the `prompt` will be extracted from chosen messages. - If `config.valid_split` is not None, the validation data will also be processed. """ valid_data = None if config.chat_template in ("chatml", "zephyr", "tokenizer"): logger.info("Applying chat template") logger.info("For ORPO/DPO, `prompt` will be extracted from chosen messages") train_data = train_data.map( apply_chat_template, fn_kwargs={ "tokenizer": tokenizer, "config": config, }, ) if config.valid_split is not None: valid_data = valid_data.map( apply_chat_template, fn_kwargs={ "tokenizer": tokenizer, "config": config, }, ) return train_data, valid_data def configure_logging_steps(config, train_data, valid_data): """ Configures the logging steps for training based on the provided configuration and data. Parameters: config (object): Configuration object containing training parameters, including `logging_steps`, `valid_split`, and `batch_size`. train_data (iterable): Training dataset. valid_data (iterable): Validation dataset. Returns: int: The number of logging steps to be used during training. Notes: - If `config.logging_steps` is set to -1, the function calculates logging steps based on 20% of the length of the validation data (if `valid_split` is provided) or the training data. - The calculated logging steps are constrained to be between 1 and 25. - If `config.logging_steps` is not -1, the function uses the provided value. """ logger.info("configuring logging steps") if config.logging_steps == -1: if config.valid_split is not None: logging_steps = int(0.2 * len(valid_data) / config.batch_size) else: logging_steps = int(0.2 * len(train_data) / config.batch_size) if logging_steps == 0: logging_steps = 1 if logging_steps > 25: logging_steps = 25 config.logging_steps = logging_steps else: logging_steps = config.logging_steps logger.info(f"Logging steps: {logging_steps}") return logging_steps def configure_training_args(config, logging_steps): """ Configures the training arguments for a language model based on the provided configuration. Args: config (object): Configuration object containing various training parameters. logging_steps (int): Number of steps between logging events. Returns: dict: A dictionary containing the configured training arguments. The configuration object `config` should have the following attributes: - project_name (str): The name of the project, used as the output directory. - batch_size (int): Batch size for both training and evaluation. - lr (float): Learning rate. - epochs (int): Number of training epochs. - eval_strategy (str): Evaluation strategy, e.g., "steps" or "epoch". - valid_split (float or None): Validation split ratio. If None, evaluation is disabled. - save_total_limit (int): Maximum number of checkpoints to save. - gradient_accumulation (int): Number of gradient accumulation steps. - log (str): Logging destination, e.g., "tensorboard". - auto_find_batch_size (bool): Whether to automatically find the optimal batch size. - scheduler (str): Learning rate scheduler type. - optimizer (str): Optimizer type. - warmup_ratio (float): Warmup ratio for learning rate scheduling. - weight_decay (float): Weight decay for the optimizer. - max_grad_norm (float): Maximum gradient norm for clipping. - disable_gradient_checkpointing (bool): Whether to disable gradient checkpointing. - peft (bool): Whether to use Parameter-Efficient Fine-Tuning (PEFT). - quantization (str): Quantization type, e.g., "int4" or "int8". - mixed_precision (str): Mixed precision type, e.g., "fp16" or "bf16". The function also sets additional training arguments based on the provided configuration, such as enabling gradient checkpointing and mixed precision training. """ logger.info("configuring training args") training_args = dict( output_dir=config.project_name, per_device_train_batch_size=config.batch_size, per_device_eval_batch_size=config.batch_size, learning_rate=config.lr, num_train_epochs=config.epochs, eval_strategy=config.eval_strategy if config.valid_split is not None else "no", logging_steps=logging_steps, save_total_limit=config.save_total_limit, save_strategy=config.eval_strategy if config.valid_split is not None else "no", gradient_accumulation_steps=config.gradient_accumulation, report_to=config.log, auto_find_batch_size=config.auto_find_batch_size, lr_scheduler_type=config.scheduler, optim=config.optimizer, warmup_ratio=config.warmup_ratio, weight_decay=config.weight_decay, max_grad_norm=config.max_grad_norm, push_to_hub=False, load_best_model_at_end=True if config.valid_split is not None else False, ddp_find_unused_parameters=False, gradient_checkpointing=not config.disable_gradient_checkpointing, remove_unused_columns=False, ) if not config.disable_gradient_checkpointing: if config.peft and config.quantization in ("int4", "int8"): training_args["gradient_checkpointing_kwargs"] = {"use_reentrant": True} else: training_args["gradient_checkpointing_kwargs"] = {"use_reentrant": False} if config.mixed_precision == "fp16": training_args["fp16"] = True if config.mixed_precision == "bf16": training_args["bf16"] = True return training_args def configure_block_size(config, tokenizer): """ Configures the block size for the given configuration and tokenizer. This function sets the `block_size` attribute in the `config` object based on the `tokenizer`'s maximum model length. If `config.block_size` is -1, it is set to None. If `config.block_size` is None, it defaults to the tokenizer's `model_max_length` but not exceeding 1024. If `config.block_size` is specified and exceeds the tokenizer's `model_max_length`, a warning is logged and the block size is set to the tokenizer's `model_max_length`. Args: config (object): Configuration object that contains the `block_size` attribute. tokenizer (object): Tokenizer object that contains the `model_max_length` attribute. Returns: object: The updated configuration object with the `block_size` attribute set. """ if config.block_size == -1: config.block_size = None if config.block_size is None: block_size = tokenizer.model_max_length if block_size > 1024: logger.warning( "The chosen tokenizer supports a `model_max_length` that is longer than the default `block_size` value" " of 1024. If you would like to use a longer `block_size` up to `tokenizer.model_max_length` you can" " override this default with `--block_size xxx`." ) block_size = 1024 else: if config.block_size > tokenizer.model_max_length: logger.warning( f"The block_size passed ({config.block_size}) is larger than the maximum length for the model" f"({tokenizer.model_max_length}). Using block_size={tokenizer.model_max_length}." ) block_size = min(config.block_size, tokenizer.model_max_length) config.block_size = block_size logger.info(f"Using block size {block_size}") return config def get_callbacks(config): """ Generate a list of callback instances based on the provided configuration. This function creates a list of callback instances that are used during the training process. It includes default callbacks for logging and training start, and conditionally adds callbacks for saving and loading PEFT models based on the configuration and environment settings. Args: config (object): Configuration object containing training settings and parameters. Returns: list: A list of callback instances to be used during training. """ is_deepspeed_enabled = os.environ.get("ACCELERATE_USE_DEEPSPEED", "False").lower() == "true" callbacks = [UploadLogs(config=config), LossLoggingCallback(), TrainStartCallback()] if config.peft and not is_deepspeed_enabled: callbacks.append(SavePeftModelCallback) if config.valid_split is not None: callbacks.append(LoadBestPeftModelCallback) return callbacks def get_model(config, tokenizer): """ Loads and configures a language model based on the provided configuration and tokenizer. Args: config (Namespace): Configuration object containing model parameters and settings. - model (str): The model name or path. - token (str): Token for accessing the model. - unsloth (bool): Flag to determine if unsloth is used. - trainer (str): Type of trainer to use. - target_modules (str): Target modules for unsloth. - peft (bool): Flag to determine if PEFT (Parameter-Efficient Fine-Tuning) is used. - quantization (str): Quantization type, either "int4" or "int8". - mixed_precision (str): Mixed precision type, either "fp16" or "bf16". - block_size (int): Maximum sequence length. - lora_r (int): LoRA rank. - lora_alpha (int): LoRA alpha. - lora_dropout (float): LoRA dropout rate. - seed (int): Random seed. - disable_gradient_checkpointing (bool): Flag to disable gradient checkpointing. - use_flash_attention_2 (bool): Flag to use flash attention 2. tokenizer (PreTrainedTokenizer): Tokenizer to use with the model. Returns: PreTrainedModel: The configured language model. Raises: ImportError: If unsloth is not available when required. """ model_config = AutoConfig.from_pretrained( config.model, token=config.token, trust_remote_code=ALLOW_REMOTE_CODE, ) model_type = model_config.model_type unsloth_target_modules = None can_use_unloth = False if config.unsloth and is_unsloth_available() and config.trainer in ("default", "sft"): can_use_unloth = True if model_type in ("llama", "mistral", "gemma", "qwen2") and config.unsloth: if config.target_modules.strip().lower() == "all-linear": unsloth_target_modules = ["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"] else: unsloth_target_modules = get_target_modules(config) else: can_use_unloth = False logger.info(f"Can use unsloth: {can_use_unloth}") if can_use_unloth: from unsloth import FastLanguageModel load_in_4bit = False load_in_8bit = False if config.peft and config.quantization == "int4": load_in_4bit = True elif config.peft and config.quantization == "int8": load_in_8bit = True dtype = None if config.mixed_precision == "fp16": dtype = torch.float16 elif config.mixed_precision == "bf16": dtype = torch.bfloat16 model, _ = FastLanguageModel.from_pretrained( model_name=config.model, token=config.token, trust_remote_code=ALLOW_REMOTE_CODE, load_in_4bit=load_in_4bit, load_in_8bit=load_in_8bit, max_seq_length=config.block_size, dtype=dtype, ) if config.peft: model = FastLanguageModel.get_peft_model( model, r=config.lora_r, target_modules=unsloth_target_modules, lora_alpha=config.lora_alpha, lora_dropout=config.lora_dropout, bias="none", use_gradient_checkpointing="unsloth", random_state=config.seed, max_seq_length=config.block_size, use_rslora=False, loftq_config=None, ) return model else: logger.warning("Unsloth not available, continuing without it...") logger.info("loading model config...") model_config = AutoConfig.from_pretrained( config.model, token=config.token, trust_remote_code=ALLOW_REMOTE_CODE, use_cache=config.disable_gradient_checkpointing, ) logger.info("loading model...") if config.peft: if config.quantization == "int4": bnb_config = BitsAndBytesConfig( load_in_4bit=True, bnb_4bit_quant_type="nf4", bnb_4bit_compute_dtype=torch.float16, bnb_4bit_use_double_quant=False, ) elif config.quantization == "int8": bnb_config = BitsAndBytesConfig(load_in_8bit=True) else: bnb_config = None model = AutoModelForCausalLM.from_pretrained( config.model, config=model_config, token=config.token, quantization_config=bnb_config, trust_remote_code=ALLOW_REMOTE_CODE, use_flash_attention_2=config.use_flash_attention_2, ) else: model = AutoModelForCausalLM.from_pretrained( config.model, config=model_config, token=config.token, trust_remote_code=ALLOW_REMOTE_CODE, use_flash_attention_2=config.use_flash_attention_2, ) logger.info(f"model dtype: {model.dtype}") model.resize_token_embeddings(len(tokenizer)) if config.trainer != "default": return model if config.peft: logger.info("preparing peft model...") if config.quantization is not None: gradient_checkpointing_kwargs = {} if not config.disable_gradient_checkpointing: if config.quantization in ("int4", "int8"): gradient_checkpointing_kwargs = {"use_reentrant": True} else: gradient_checkpointing_kwargs = {"use_reentrant": False} model = prepare_model_for_kbit_training( model, use_gradient_checkpointing=not config.disable_gradient_checkpointing, gradient_checkpointing_kwargs=gradient_checkpointing_kwargs, ) else: model.enable_input_require_grads() peft_config = LoraConfig( r=config.lora_r, lora_alpha=config.lora_alpha, lora_dropout=config.lora_dropout, bias="none", task_type="CAUSAL_LM", target_modules=get_target_modules(config), ) model = get_peft_model(model, peft_config) return model