We invite the community to replicate and challenge these results.
- Replicate the surgery — Use the scripts in
code/surgery/on the same or different MoE models - Measure on your hardware — Deploy to your own platform and measure power
- Calculate EPI — Use epi-bench
- Submit results — PR to
data/community/with your hardware description and measurements - Discuss findings — Open a Discussion with observations, disagreements, or alternative analyses
data/community/<github-username>_<hardware>/
├── README.md # Your setup description
├── run_<surgery_id>.json # EPI result per configuration
├── power_trace_<id>.csv # Power traces
└── benchmark_<id>.json # Benchmark scores
- Surgery scripts for alternative MoE models
- Alternative pruning strategies (not just router frequency)
- Visualization improvements
- Statistical analysis extensions