Training Visualisation#
RLinf support for real-time experiment tracking. You can stream loss curves, accuracy, GPU utilization and arbitrary custom metrics to one or more of the following backends:
TensorBoard: A widely used, open-source visualization tool (from TensorFlow, also works with PyTorch, Hugging Face, etc.) that lets you track metrics like loss and accuracy, visualize model graphs, embeddings, images, and more.
Weights & Biases (W&B): A SaaS-based platform offering experiment tracking, hyperparameter sweeps, artifacts (for model and data versioning), reporting, and collaborative features for teams.
SwanLab: An open-source, lightweight experiment logging and visualization tool designed for local or self-hosted use. It provides intuitive Python APIs, logs metrics, hyperparameters, hardware and code info, and supports experiment comparison through a clean UI — ideal for privacy-focused workflows.
Enabling a back-end#
Add the desired logger(s) to runner.logger.logger_backends in your YAML:
runner:
task_type: math
logger:
log_path: ${runner.output_dir}/${runner.experiment_name}
project_name: rlinf
experiment_name: ${runner.experiment_name}
logger_backends: ["tensorboard", "wandb", "swanlab"] # <─ choose any subset
experiment_name: grpo-1.5b
output_dir: ./logs
RLinf creates a sub-directory for each active back-end:
logs/grpo-1.5b/
├── checkpoints/
├── converted_ckpts/
├── log/
├── swanlab/ # SwanLab event files
├── tensorboard/ # TensorBoard event files
└── wandb/ # WandB run directory
TensorBoard#
tensorboard --logdir ./logs/grpo-1.5b/tensorboard --port 6006
Open http://localhost:6006 in your browser
to inspect scalar plots, histograms and the computation graph.
Weights & Biases (WandB)#
Create a free account at wandb.ai and copy your API key.
Authenticate once per machine:
wandb login # paste API key when prompted
From now on RLinf will automatically start a new run and stream all metrics. You can check the metrics through your dashboard.
SwanLab#
Register at swanlab.ai and obtain an access token.
Authenticate:
swanlab login # paste access token
From now on RLinf will automatically start a new run and stream all metrics. You can check the metrics through your dashboard.
Tip
All three loggers run in parallel; feel free to mix and match.