The loggingLevel setup in Transformers-CFG will polute the user program

[Saibo-creator]

The probelmatic line is here

transformers-CFG/transformers_cfg/logging_config.py

Line 18 in 651dca0

logging.basicConfig(level=log_level)

reproduce

import torch
import logging  

if __name__ == "__main__":


    from transformers import AutoModelForCausalLM, AutoTokenizer
    from transformers_cfg.grammar_utils import IncrementalGrammarConstraint
    from transformers_cfg.generation.logits_process import GrammarConstrainedLogitsProcessor


    # logging.basicConfig has already been called in the transformers_cfg library, so this will not work
    # the above function can only be called once in a python session
    # so we should NEVER call logging.basicConfig in a library! 
    logging.basicConfig(level=logging.INFO)
    logger = logging.getLogger(__name__)

    # you will not see this message in the console because the logging level is set to WARNING inside the transformers_cfg library
    logger.info("Loading model and tokenizer")

Something which works is below:

import torch
import logging  

if __name__ == "__main__":

    logging.basicConfig(level=logging.INFO)

    from transformers import AutoModelForCausalLM, AutoTokenizer
    # the function call inside the transformers_cfg will not work because we have already called logging.basicConfig above
    from transformers_cfg.grammar_utils import IncrementalGrammarConstraint
    from transformers_cfg.generation.logits_process import GrammarConstrainedLogitsProcessor



    logger = logging.getLogger(__name__)

    # you will SEE this log message because the logging level is set to INFO
    logger.info("Loading model and tokenizer")

Reason

import os
import logging


def setup_logging():
    log_level_name = os.getenv(
        "TCFG_LOG_LEVEL", "WARNING"
    ).upper()  # Default to WARNING if not set
    log_levels = {
        "DEBUG": logging.DEBUG,
        "INFO": logging.INFO,
        "WARNING": logging.WARNING,
        "ERROR": logging.ERROR,
        "CRITICAL": logging.CRITICAL,
    }
    log_level = log_levels.get(log_level_name, logging.WARNING)
   # this is BAD, this function configures the root logger of the entire program, so it will overwrite the user's logging level defined in their script when they import the library
    logging.basicConfig(level=log_level)

Solution

import logging

# Create a logger for the library
logger = logging.getLogger('transformers_cfg')
# the level will propagate to loggers with submodule scope
logger.setLevel(logging.WARNING)

[Saibo-creator]

solved in #51