1. Run System Health Check

moro doctor
  

2. CUDA Out Of Memory (OOM) Errors

If training experiences memory pressure despite auto-recovery, adjust your configuration:

• Enable gradient checkpointing: gradient_checkpointing: true in moro.yaml.
• Reduce micro-batch size to 1 and double gradient accumulation: micro_batch_size: 1.
• Clamp sequence length: max_seq_length: 1024.

3. MPS Accelerator (Apple Silicon)

On Apple Silicon M-series Macs, ensure PyTorch uses MPS:

export PYTORCH_ENABLE_MPS_FALLBACK=1
moro diagnose