Troubleshooting

This page covers common issues and their solutions when working with MIP.

Loss Doesn't Work

Problem: Mean Flow, Lagrangian Self-Distillation, Euler Self-Distillation don't work with transformer.

Solutions: This is a know issue with torch, consider use UNet instead.

CUDA Graphs Not Working

Problem: use_cudagraphs=true has no effect or crashes

Reasons:

1. Image observations: CUDA graphs require static tensor shapes, not supported for image observations

   • Solution: Use state observations or disable CUDA graphs

2. Dynamic shapes in network: Some networks may have dynamic computation

   • Solution: Use simpler networks like MLP

3. Inconsistent batch sizes: DataLoader must have drop_last=True

   • Already set in training scripts, verify if using custom script

Dataset Not Found

Problem: FileNotFoundError: Dataset not found

Solutions:

1. For HuggingFace datasets: Ensure you're connected to internet and have huggingface_hub installed

2. For local datasets: Download and process dataset first:

   uv run python examples/process_robomimic_dataset.py --skip_upload

3. Override dataset path:

   uv run examples/train_robomimic.py +task.dataset_path=/path/to/dataset.hdf5