Logging variants

[maharajamihir]

added support for multiple loggers:

wandb

tensorboard

local

console

[Copilot]

Pull Request Overview

This PR centralizes logging by introducing a flexible CompositeLogger that supports multiple backends (wandb, TensorBoard, local JSON/PNG, and console) and updating all training scripts to use it.

• Added utils/logger.py containing BaseLogger and concrete loggers (WandbLogger, TensorboardLogger, LocalLogger, ConsoleLogger, and CompositeLogger).
• Modified train_tokenizer.py, train_lam.py, and train_dynamics.py to remove direct wandb usage and configure logging via CompositeLogger.
• Updated each script’s Args to specify log_dir and a list of desired loggers.

Reviewed Changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 4 comments.

File	Description
utils/logger.py	New module defining an abstract logger interface and four implementations, plus a composite wrapper.
train_tokenizer.py	Replaced wandb calls with CompositeLogger; changed CLI args to loggers list instead of a flag.
train_lam.py	Same updates as tokenizer training: use CompositeLogger and new loggers/log_dir args.
train_dynamics.py	Aligned with other scripts to use CompositeLogger; removed direct wandb integration.

Comments suppressed due to low confidence (1)

utils/logger.py:233

• [nitpick] Using the key "tb" for TensorBoard may be unclear to new users. Consider renaming it to "tensorboard" or providing both aliases for clarity.

                            "tb": TensorboardLogger,

[emergenz]

Since we are going to do proper logging, do we want to switch from prints to proper python logging?

[emergenz]

Switch to python logging in consoleLogging