[PyTorch] transformer engine modifies root logger

[Marks101]

Hello everyone,

we noticed that since #1226, transformer engine modifies the root logger of Python in the global scope:

TransformerEngine/transformer_engine/pytorch/attention.py

Lines 100 to 101 in 7fb22c3

	fa_logger = logging.getLogger()
	fa_logger.setLevel(_log_level)

That means that importing transformer-engine already modifies the logger setup of the application.

Is it maybe possible to use a named logger? fa_logger = logging.getLogger("FlashAttention") for example?

(related to this issue #1065)

Thank you 😄

[cyanguwa]

Hi Markus!

I can certainly add a name to the logger! I guess I meant for these package availability messages (e.g. this and this) to be global messages that are not tied to a particular module.

What do you mean by importing transformer-engine already modifies the logger setup of the application? Did this fa_logger change the formatting of other loggers in your application? I think the formatter is only applied here right?

[Marks101]

Hi Charlene,

Did this fa_logger change the formatting of other loggers in your application?

Not the format, but the log level. I was surprised as well 😄 By omitting the name argument when calling getLogger(), the fa_logger is assigned to the root logger of the logging framework. Modifying the logging level of the root logger influences other loggers. This is from the documentation of logging:

Logged messages to the module-level logger get forwarded to handlers of loggers in higher-level modules, all the way up to the highest-level logger known as the root logger

So modyfing the root logger with these two lines:

TransformerEngine/transformer_engine/pytorch/attention.py

Lines 100 to 101 in 7fb22c3

	fa_logger = logging.getLogger()
	fa_logger.setLevel(_log_level)

modifies all other loggers.

I double checked that with the following snippet:

import logging

logger = logging.getLogger("MyApp")
logging.basicConfig(level=logging.DEBUG)

# when this gets uncommented, there is no logging output to the console anymore
#import transformer_engine

logger.debug("This is a debug message.")

Can you reproduce this?