How large-scale-sustainable-computing-lsc/green-code works
GREEN-CODE trains early-exit language models and RL policies offline, then serves energy-aware code completion online. The graph preserves dataset preparation, fine-tuning, RL training, evaluation, and analysis as distinct stages. Endpoint and model excerpts support the runtime links; several training and evaluation details come from the README, so uncertain implementation-level wiring is omitted.
Main entry points: modeling_llama_rl_with_cache.py Llama early-exit · modeling_opt_ee_rl.py OPT early-exit · evaluator.py Benchmark evaluation · rl RL agent training
Components 14
Offline preparation
- Dataset analysis
src/datasets/dataset_vis.py
Model and policy training
- Fine-tuning
src/train/finetuner.py - Llama early-exit
src/models/llama/modeling_llama_rl_with_cache.py - OPT early-exit
src/models/opt/modeling_opt_ee_rl.py - RL environments
src/models/rl/enviornments - RL agent training
src/models/rl
Online generation
- Generation API
src/endpoint/main.py - Generation adapter
src/endpoint/generators.py
Evaluation and analysis
- Benchmark evaluation
src/eval/evaluator.py - Specialized metrics
src/eval - Result visualization
src/plotter
Other components
- Engineer
- Code datasets
- VS Code client
Connections 19
- Engineer → Code datasets: selects
- Code datasets → Dataset analysis: analyzes
- Code datasets → Fine-tuning: loads and prepares
- Fine-tuning → Llama early-exit: fine-tunes
- Fine-tuning → OPT early-exit: fine-tunes
- RL environments → Llama early-exit: samples states
- RL environments → OPT early-exit: samples states
- RL agent training → RL environments: trains in
- RL agent training → Llama early-exit: provides policy
- RL agent training → OPT early-exit: provides policy
- VS Code client → Generation API: requests completion
- Generation API → Generation adapter: dispatches request
- Generation adapter → Llama early-exit: generates with
- Generation API → Generation adapter: updates threshold
- Code datasets → Benchmark evaluation: supplies samples
- Llama early-exit → Benchmark evaluation: evaluated by
- OPT early-exit → Benchmark evaluation: evaluated by
- Benchmark evaluation → Specialized metrics: supports analysis
- Specialized metrics → Result visualization: visualizes results