import datetime import logging import logging.handlers import os import sys import torch import requests from transformers import StoppingCriteria from .constants import LOGDIR server_error_msg = "**NETWORK ERROR DUE TO HIGH TRAFFIC. PLEASE REGENERATE OR REFRESH THIS PAGE.**" moderation_msg = "YOUR INPUT VIOLATES OUR CONTENT MODERATION GUIDELINES. PLEASE TRY AGAIN." handler = None def build_logger(logger_name, logger_filename): global handler formatter = logging.Formatter( fmt="%(asctime)s | %(levelname)s | %(name)s | %(message)s", datefmt="%Y-%m-%d %H:%M:%S", ) # Set the format of root handlers if not logging.getLogger().handlers: logging.basicConfig(level=logging.INFO) logging.getLogger().handlers[0].setFormatter(formatter) # Redirect stdout and stderr to loggers stdout_logger = logging.getLogger("stdout") stdout_logger.setLevel(logging.INFO) sl = StreamToLogger(stdout_logger, logging.INFO) sys.stdout = sl stderr_logger = logging.getLogger("stderr") stderr_logger.setLevel(logging.ERROR) sl = StreamToLogger(stderr_logger, logging.ERROR) sys.stderr = sl # Get logger logger = logging.getLogger(logger_name) logger.setLevel(logging.INFO) # Add a file handler for all loggers if handler is None: os.makedirs(LOGDIR, exist_ok=True) filename = os.path.join(LOGDIR, logger_filename) handler = logging.handlers.TimedRotatingFileHandler( filename, when='D', utc=True) handler.setFormatter(formatter) for name, item in logging.root.manager.loggerDict.items(): if isinstance(item, logging.Logger): item.addHandler(handler) return logger class StreamToLogger(object): """ Fake file-like stream object that redirects writes to a logger instance. """ def __init__(self, logger, log_level=logging.INFO): self.terminal = sys.stdout self.logger = logger self.log_level = log_level self.linebuf = '' def __getattr__(self, attr): return getattr(self.terminal, attr) def write(self, buf): temp_linebuf = self.linebuf + buf self.linebuf = '' for line in temp_linebuf.splitlines(True): # From the io.TextIOWrapper docs: # On output, if newline is None, any '\n' characters written # are translated to the system default line separator. # By default sys.stdout.write() expects '\n' newlines and then # translates them so this is still cross platform. if line[-1] == '\n': self.logger.log(self.log_level, line.rstrip()) else: self.linebuf += line def flush(self): if self.linebuf != '': self.logger.log(self.log_level, self.linebuf.rstrip()) self.linebuf = '' def disable_torch_init(): """ Disable the redundant torch default initialization to accelerate model creation. """ import torch setattr(torch.nn.Linear, "reset_parameters", lambda self: None) setattr(torch.nn.LayerNorm, "reset_parameters", lambda self: None) def violates_moderation(text): """ Check whether the text violates OpenAI moderation API. """ url = "https://api.openai.com/v1/moderations" headers = {"Content-Type": "application/json", "Authorization": "Bearer " + os.environ["OPENAI_API_KEY"]} text = text.replace("\n", "") data = "{" + '"input": ' + f'"{text}"' + "}" data = data.encode("utf-8") try: ret = requests.post(url, headers=headers, data=data, timeout=5) flagged = ret.json()["results"][0]["flagged"] except requests.exceptions.RequestException as e: flagged = False except KeyError as e: flagged = False return flagged def pretty_print_semaphore(semaphore): if semaphore is None: return "None" return f"Semaphore(value={semaphore._value}, locked={semaphore.locked()})" class KeywordsStoppingCriteria(StoppingCriteria): def __init__(self, keywords, tokenizer, input_ids): self.keywords = keywords self.keyword_ids = [tokenizer(keyword).input_ids for keyword in keywords] self.keyword_ids = [keyword_id[0] for keyword_id in self.keyword_ids if type(keyword_id) is list and len(keyword_id) == 1] self.tokenizer = tokenizer self.start_len = None self.input_ids = input_ids def __call__(self, output_ids: torch.LongTensor, scores: torch.FloatTensor, **kwargs) -> bool: if self.start_len is None: self.start_len = self.input_ids.shape[1] else: for keyword_id in self.keyword_ids: if output_ids[0, -1] == keyword_id: return True outputs = self.tokenizer.batch_decode(output_ids[:, self.start_len:], skip_special_tokens=True)[0] for keyword in self.keywords: if keyword in outputs: return True return False def smart_tokenizer_and_embedding_resize(special_tokens_dict, tokenizer, model): """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) # # num_new_tokens = 1 # # tokenizer.add_tokens(special_tokens_dict, special_tokens=True) # model.resize_token_embeddings(len(tokenizer)) 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 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, name=k) 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 find_all_linear_names(model): cls = torch.nn.Linear lora_module_names = set() for name, module in model.named_modules(): if isinstance(module, cls) and 'vision_model' not in name and 'mm_projector' not in name and 'vision_encoder' not in name and 'conv_final' not in name and'lm_head' not in name: lora_module_names.add(name) print(lora_module_names) return list(lora_module_names)