An experiment tracker that works offline and speaks Chinese silicon

What it does

SwanLab logs metrics, hyperparameters, and hardware telemetry from deep-learning runs, then renders them in a web dashboard for comparison and analysis. It plugs into 50+ frameworks—including PyTorch, HuggingFace Transformers, and LLaMA Factory—so you can add it to existing code without rewriting training loops. You can use the hosted cloud service or keep everything inside your network via Docker or Kubernetes.

The interesting bit

Most experiment trackers treat NVIDIA as the only hardware that matters; SwanLab also monitors AMD ROCm and at least seven China-specific accelerators such as Cambricon MLU, Hygon DCU, and Kunlun XPU. A local-first sync mode lets you train air-gapped and push logs to the server later, which is useful when your cluster lacks outbound internet.

Key highlights

• 50+ framework integrations, including official support in HuggingFace Transformers and Accelerate
• Hardware monitoring for NVIDIA, AMD, and Chinese chips (MetaX, Moore Threads, Iluvatar, etc.)
• Fully self-hostable via Docker or Kubernetes; offline training with later cloud sync
• LightningBoard dashboard designed for “super-large chart counts”
• Embeddable charts, 3D point clouds, molecule visualizations, and custom ECharts

Caveats

• Documentation and community support are primarily Chinese-language; English/Japanese/Russian docs exist but may lag behind
• The README is essentially a dense changelog, making it hard to judge architectural stability or long-term roadmap

Verdict

Worth a look if you train models on mixed or non-NVIDIA hardware and need a self-hosted alternative to Weights & Biases. If you are already invested in TensorBoard or W&B and use only standard NVIDIA GPUs, the switching cost may outweigh the gains.