Pytorch should not log to python root logger

[ppwwyyxx]

🐛 Describe the bug

Some components of pytorch currently log to python's root logger, e.g.:

$ag -s 'logging.info'
distributed/optim/zero_redundancy_optimizer.py
1387:                logging.info(
1392:                    logging.info(
1506:                logging.info(

distributed/elastic/timer/local_timer.py
118:            logging.info(f"Process with pid={worker_id} does not exist. Skipping")

distributed/elastic/timer/api.py
191:            logging.info(
196:                logging.info(f"Successfully reaped worker=[{worker_id}]")
208:        logging.info(
216:        logging.info("Starting watchdog thread...")
220:        logging.info(f"Stopping {type(self).__name__}")
223:            logging.info("Stopping watchdog thread...")
227:            logging.info("No watchdog thread running, doing nothing")
239:    logging.info(f"Timer client configured to: {type(_timer_client).__name__}")

distributed/distributed_c10d.py
3232:                logging.info("Rank {} is assigned to subgroup {}".format(rank, ranks))

I believe this isn't a good practice. As a result user cannot easily control the behavior of pytorch's logs / split it from user's own logs.

The issue is most frequent under torch.distributed.

These code should log to logger.getLogger(__name__) instead. Same for warning and error.

Versions

latest master

cc @pietern @mrshenli @pritamdamania87 @zhaojuanmao @satgera @rohan-varma @gqchen @aazzolini @osalpekar @jiayisuse @SciPioneer @H-Huang