[Log] Optimize startup log

[yewentao256]

Purpose

Reduce duplicate logs when using dp/tp

[gemini-code-assist]

Code Review

This pull request optimizes startup logs to reduce duplication when running in a distributed environment (data parallelism / tensor parallelism). The changes correctly use logger.info_once and logger.warning_once with appropriate scopes (global or local) to ensure that messages are logged only once, or once per node, as intended. The implementation is correct and improves the logging behavior. For consistency and further log reduction, you might consider applying this pattern to other logger.info and logger.warning calls that are executed by multiple processes during startup.

[njhill]

Thanks @yewentao256!

[njhill]

I don't think we need to use get_attr here since it should always be an attribute.

Also we should maybe use data_parallel_rank_local instead so that this is logged in each node?
e.g.

Suggested change

	if getattr(self.parallel_config, "data_parallel_rank", 0) == 0:
	if self.parallel_config.data_parallel_rank_local in (None, 0):