Loss of Color and Vertical Formatting in Logs After Logging Refactor #570
Description
In order to effectively run and monitor training jobs, ML practitioners often track important metrics such as loss and gradnorm as they evolve in real-time. PR #500 refactored the logging system to enhance its robustness. However, the new logging implementation has changed the formatting and removed the previously available color highlights.
Below is a visual comparison for reference:
Old logs
|
New logs
|
Previously, logs were displayed in a vertical format with color highlights, allowing practitioners to quickly scan the logs vertically and easily identify the metrics of interest without external tools (e.g., TensorBoard, Wandb).
In the current system, logs are presented horizontally on a single line without color differentiation. This change requires practitioners to scan horizontally while the logs themselves move vertically, and also necessitates additional parsing to distinguish metric names from their corresponding values.