Weights & Biases

Install the optional extra and authenticate once:

pip install "trloom[wandb]"
# or from source: pip install -e ".[wandb]"
wandb login

Enable in YAML

training:
  report_to: wandb   # optional; set automatically when wandb.enabled is true

wandb:
  enabled: true
  project: my-project
  entity: my-team
  run_name: qwen-sft-01
  group: experiments
  tags: [sft, lora]
  notes: First SFT baseline
  mode: online       # online | offline | disabled
  dir: null
  job_type: train
  init_kwargs: {}

When wandb.enabled is true, TRLoom:

1. Sets related environment variables when provided (WANDB_MODE, WANDB_PROJECT, WANDB_ENTITY, WANDB_DIR)
2. Ensures training.report_to includes wandb
3. Calls wandb.init(...) with your settings and the full config dump
4. Finishes the run when job.run() completes

Example

See examples/dpo_wandb.yaml for DPO + W&B:

trloom run examples/dpo_wandb.yaml

Modal + W&B

Create a Modal secret, then reference it:

python -m modal secret create wandb WANDB_API_KEY=...

wandb:
  enabled: true
  project: trloom-grpo

modal:
  enabled: true
  secrets:
    - wandb
    - huggingface

Troubleshooting

• Missing package → install trloom[wandb]
• Offline debugging → set mode: offline
• Disable without editing the rest of the block → enabled: false or mode: disabled