Metric Logging

[calvinpelletier]

Build metric logger:

metric_logger = get_metric_logger(**config.metrics)

Log a metric:

metric_logger.log("name", value, step)

If you want to check if this is a log step before calculating the metric:

if metric_logger.is_log_step(name, step):
    value = calculate_metric()
    metric_logger.log(name, value, step)

Open questions:

• multiple loggers? e.g. both wandb and stdout
• non-constant log frequencies? (e.g. for metrics that take a long time to calculate, I like to use a logarithmic frequency so that i get quick data points early in training and fewer points as training becomes more stable)

Follow up PRs:

• put metric logger inside an actor

[ebsmothers]

A few more small comments, after that I think this is good!

[ebsmothers]

@calvinpelletier bumping this PR. Looks like there are some merge conflicts. Let's resolve those and address the remaining comments, then we should be good to land

[calvinpelletier]

@ebsmothers addressed the comments. I'll leave the config class for the next PR after adding metric logging for sft_v2 and toy_rl

[ebsmothers]

Two more nits, otherwise looks good!

[ebsmothers]

Suggested change

	INFO:torchtune.utils._logging:Hello world!
	INFO:forge.util.logging:Hello world!

[ebsmothers]

Suggested change

	project (str): WandB project name. Default is `torchtune`.
	project (str): WandB project name.

[DNXie]

Shouldn't log rank be configurable like how titan did? But I guess we can always improve later.