Troubleshooting Guide
Common issues and solutions.
Import Errors
Dimension Mismatch
Problem: Dimensions don’t match between models
Solution: Use SVD projection
projector = DimensionProjector(method="svd", variance_threshold=0.95)
Architecture Not Detected
Problem: Model architecture not recognized
Solution: Add custom binder or specify manually
ual = UniversalAdapter(model, tokenizer, architecture="custom")
Training Issues
OOM Errors
Problem: Out of memory during training
Solutions:
Reduce batch size
Use gradient accumulation
Lower LoRA rank
Enable gradient checkpointing
Slow Training
Problem: Training takes too long
Solutions:
Increase batch size
Use fewer target modules
Reduce dataset size
Use GPU if available
Dispatcher Issues
Poor Routing Accuracy
Problem: Queries routed to wrong domains
Solutions:
Add more diverse examples
Adjust confidence threshold
Analyze domain overlap
Retrain router with better examples
Low Confidence Scores
Problem: All confidence scores below threshold
Solutions:
Lower confidence threshold
Improve domain examples
Add fallback domain
Check query preprocessing
Quality Issues
Poor Transfer Quality
Problem: Transferred adapter performs poorly
Solutions:
Use SVD projection method
Increase LoRA rank
Check dimension compatibility
Verify source adapter quality
Getting Help
If issues persist:
Enable debug logging
Check GitHub issues
Provide minimal reproduction
Include error messages and logs