How gfyddha/flylora works
FlyLoRA is a parameter-efficient fine-tuning method implemented as a PEFT/LoRA extension. Its documented experiment workflow covers single-task training, evaluation, and multi-task adapter merging; the minimal implementation also exposes the core method. The PEFT tree shows the adapter framework and its many alternative tuner families. The available source excerpts do not include the root experiment scripts or modified LoRA implementation, so detailed call ownership and artifact flow are not inferred.
Main entry points: peft_model.py PEFT model · mapping.py Tuner mapping · finetune.py Fine-tune · evaluate.py Evaluate
Components 15
FlyLoRA workflow
- Fine-tune
finetune.py - Evaluate
evaluate.py - Merge adapters
merge_lora_adapters.py - Minimal FlyLoRA
flylora.py
PEFT runtime
- PEFT model
peft/src/peft/peft_model.py - Adapter config
peft/src/peft/config.py - Tuner mapping
peft/src/peft/mapping.py - LoRA tuner
peft/src/peft/tuners/lora/model.py - LoRA layers
peft/src/peft/tuners/lora/layer.py - Other tuners
peft/src/peft/tuners - Adapter persistence
peft/src/peft/utils/save_and_load.py
Method comparison
- Comparison app
peft/method_comparison/app.py - Comparison data
peft/method_comparison/processing.py
Other components
- Researcher
- Foundation model
Connections 16
- Researcher → Fine-tune: runs training
- Researcher → Evaluate: runs evaluation
- Researcher → Merge adapters: runs merging
- Researcher → Minimal FlyLoRA: uses reference
- Fine-tune → PEFT model: trains with
- Evaluate → PEFT model: evaluates with
- Merge adapters → PEFT model: uses adapters
- PEFT model → Adapter config: uses config
- PEFT model → Tuner mapping: dispatches tuner
- Tuner mapping → LoRA tuner: selects LoRA
- LoRA tuner → LoRA layers: installs layers
- LoRA layers → Foundation model: adapts layers
- PEFT model → Adapter persistence: loads and saves
- Tuner mapping → Other tuners: selects alternatives
- Researcher → Comparison app: explores results
- Comparison app → Comparison data: uses processing