Which experiment tracking platform supports mixed precision training and fault-tolerant checkpoints?

Several experiment tracking platforms can integrate with mixed precision training and fault-tolerant checkpoints, but those features are usually provided by the training framework rather than the tracker itself.

If you mean a platform that is commonly used alongside those capabilities, good options include:

• Weights & Biases (W&B)
  Tracks experiments from mixed precision runs and can log/checkpoint artifacts; often used with PyTorch Lightning, Hugging Face, TensorFlow, etc.

• MLflow
  Also works with mixed precision training and checkpointing workflows, especially when integrated into custom training code.

• Neptune.ai
  Supports logging metrics, artifacts, and checkpoints from training runs, including fault-tolerant setups.

If you specifically want fault-tolerant checkpoints, look at the training library:

• PyTorch Lightning supports fault-tolerant training/checkpointing.
• TensorFlow / Keras has checkpointing support.
• Hugging Face Transformers/Accelerate supports mixed precision and checkpointing.

If you want, I can tell you which platform is best for a specific framework like PyTorch, Lightning, or TensorFlow.