MLflow Integration Complete

MLflow Integration Complete | Generated by AI

Home 2026.06

No code changes needed. The integration is already complete in the codebase. Here’s why:

CLI args (base_train.py lines 43-45): Already defined:

--tracker with choices wandb|mlflow|none
--mlflow-uri for remote server
--mlflow-experiment for experiment name

Init logic (lines 101-110): Already handles the switch:

if args.tracker == "mlflow":
    wandb_run = MLflowLogger(...)

Training loop: All logging goes through wandb_run.log(...) which works identically for both wandb and the MLflow logger (same interface).

MLflowLogger class: Already implements log() and finish() — numeric metrics get logged, strings become tags, config gets logged as params.

So yes, just adding these three flags to any existing training command is all it takes:

--tracker=mlflow --mlflow-experiment=<name> --run=<run-name>

Everything else (val_bpb every N steps, train loss every 100 steps, CORE metrics, checkpoints) flows through the same wandb_run.log() calls and gets captured by MLflow automatically.

Back Donate