[monitor] Unify explorer and trainer wandb logging into a single run

[MengsD]

Problem

When monitor_type: wandb is configured, Trinity creates two separate wandb runs — {name}_explorer and {name}_trainer. This makes it harder to correlate
explorer and trainer metrics (e.g., comparing rollout accuracy against actor loss), and clutters the wandb project with twice as many runs.

Proposed Solution

A primary process in the launcher creates the wandb run, and both Explorer and Trainer join as
secondary writers via wandb.init(id=run_id, resume="allow", mode="shared").

Planned Changes:

• config.py: Add wandb_run_id field to MonitorConfig (auto-populated by launcher)
• monitor.py: Add init_wandb_primary() / finish_wandb_primary() helpers; modify WandbMonitor to support shared (secondary) mode
• launcher.py: In both(), create the primary run before spawning actors and pass wandb_run_id through config

Expected Outcome:

• In shared mode, all metrics are prefixed with their role (explorer/ or trainer/), producing two clean top-level sections in the wandb UI
• wandb.define_metric assigns independent step axes (explorer/step and trainer/step) so the two processes' step counters don't conflict
• for standalone mode, explore() and train() standalone entry points should remain unchanged — they would still create their own independent runs since wandb_run_id remains None.

[pan-x-c]

I completely understand the issue you raised—it’s indeed counterintuitive and takes up more space—but it’s currently an unavoidable trade-off for an RL framework like Trinity, where training and inference are fully decoupled.

More specifically:

1. Explorer and Trainer are entirely separate processes, and they may run on different machines. Forcing them to use the same run_id can lead to data inconsistency.
2. The meaning of “step” in the Explorer and Trainer processes is not equivalent. So the metrics you see at, for example, step 5 on the Explorer side do not necessarily correspond to step 5 on the Trainer side. Putting them under a single Weights & Biases run could actually create confusion and misunderstanding.

For these reasons, Trinity separates Explorer and Trainer logs across all monitor types. In Weights & Biases, it’s recommended to observe metrics for a specific task via filtering—but please keep in mind point (2): the same step number on the Explorer and Trainer sides does not necessarily align.

[MengsD]

Thanks for the concern. I'd like to add some additional context. Please let me know your thoughts.

1. wandb shared mode is designed for this — mode="shared" with resume="allow" is wandb's official multi-process pattern. Each process buffers locally
   and syncs to the server independently; the server handles merging. There's no shared filesystem or write lock between processes, even when they run
   on different machines.
2. Step axes are fully isolated — define_metric assigns explorer/step and trainer/step as independent x-axes, so the two processes never compete on
   step counters.
3. Proven in production — Slime uses the exact same pattern (primary creates run, secondaries join via init_wandb_secondary(run_id)) across separate nodes, and has been running without data inconsistency issues.

[pan-x-c]

Thanks for the helpful feedback. W&B does support having the explorer and trainer share a single run. Going forward, we’ll consider adding an extra monitor parameter to control whether the explorer and trainer use a shared run, so we can better support and stay compatible with other monitor backends such as TensorBoard, MLflow, and SwanLab.