Memory Planning (Tune)
Plans LoRA/QLoRA fine-tuning configurations by computing optimal rank given a VRAM budget. Compares Full, LoRA, and QLoRA methods across model sizes (1B, 7B, 13B), showing trainable parameters, memory estimates, and speedup.
CLI Equivalent
apr tune
Key Concepts
- VRAM budget planning for LoRA/QLoRA fine-tuning
- Trainable parameter count estimation across model sizes
- Method comparison: Full vs LoRA vs QLoRA memory and speedup
Run
cargo run --example optimize_tune