Summary

• Add CometLogger class in nemo_automodel/components/loggers/comet_utils.py — mirrors the existing MLflowLogger pattern
• Wire Comet logging into TrainFinetuneRecipeForNextTokenPrediction:
  • Training metrics (loss, TPS, grad_norm, lr, mem) via log_train_metrics()
  • Validation metrics (val_loss) via log_val_metrics()
  • MoE load balance metrics via _log_moe_metrics()
• Enable by adding a comet: block to the training YAML config:

comet:
  project_name: "my-project"
  experiment_name: "my-run"      # optional, auto-generated from model name
  workspace: "my-workspace"      # optional, uses COMET_WORKSPACE env var
  api_key: null                  # optional, uses COMET_API_KEY env var
  tags: ["finetune", "llama"]    # optional

Motivation

Currently, Comet ML users must rely on Comet's wandb auto-patcher, which does not
reliably intercept wandb.log() calls because NeMo gates all wandb logging behind
if wandb.run is not None — requiring a valid wandb API key and active run even when
wandb is only used as a bridge to Comet. Native Comet support (matching the existing
MLflow integration pattern) eliminates this fragile dependency.

Test plan

• Run a fine-tuning job with comet: config and verify metrics appear in Comet dashboard
• Verify training metrics (loss, TPS, grad_norm, lr, mem) are logged at log_remote_every_steps frequency
• Verify validation metrics (val_loss) are logged on each validation step
• Verify MoE load balance metrics are logged (on MoE models)
• Verify that omitting comet: config has no effect (backward compatible)
• Verify comet_ml not installed raises clear ImportError