How vizuaraai/capstone-gpu-economics-engine works
GPU Economics Engine predicts inference throughput and token cost from GPU/model specifications, measures the same serving points with vLLM, then compares results and recommends the cheapest GPU meeting latency targets. The graph preserves the model’s roofline, MoE, and FLOPs-only branches alongside the measurement and validation stages. Relationships for the harness flow rely on README documentation; unsampled harness details are not inferred.
Main entry points: specs.py Typed specs · cli.py CLI commands · predict.py Bandwidth model · records.py Run records
Components 22
Prediction model
- Typed specs
src/gpu_econ/specs.py - GPU specs
data/gpus.yaml - Model specs
data/models.yaml - GPU prices
data/prices.yaml - Roofline analysis
src/gpu_econ/roofline.py - MoE dispersion
src/gpu_econ/moe.py - Bandwidth model
src/gpu_econ/predict.py - FLOPs baseline
src/gpu_econ/naive.py - SLO recommender
src/gpu_econ/slo.py
Measurement harness
- vLLM server
harness/serve.py - Serving benchmark
harness/bench.py - Measurement sweep
harness/sweep.py - Record normalizer
harness/measure.py
Records and validation
- Run records
src/gpu_econ/records.py - Predicted records
results/predicted/.gitkeep - Measured records
results/measured/.gitkeep - Record comparison
validate/compare.py - Validation report
validate/report.py
User interface
- CLI commands
src/gpu_econ/cli.py
Other components
- Engineer
- vLLM tooling
- GPU hardware
Connections 34
- Engineer → CLI commands: runs commands
README.md - GPU specs → Typed specs: loaded as specs
src/gpu_econ/specs.py - Model specs → Typed specs: loaded as specs
src/gpu_econ/specs.py - GPU prices → Typed specs: loaded as specs
src/gpu_econ/specs.py - Typed specs → Roofline analysis: provides GPU specs
src/gpu_econ/roofline.py - Typed specs → MoE dispersion: provides model specs
src/gpu_econ/moe.py - Typed specs → Bandwidth model: provides inputs
src/gpu_econ/predict.py - MoE dispersion → Bandwidth model: models dispersion
src/gpu_econ/predict.py - Typed specs → FLOPs baseline: provides inputs
src/gpu_econ/naive.py - Bandwidth model → FLOPs baseline: computes comparison
src/gpu_econ/naive.py - Bandwidth model → SLO recommender: supplies decode estimates
src/gpu_econ/slo.py - Typed specs → SLO recommender: provides candidates
src/gpu_econ/slo.py - CLI commands → Bandwidth model: dispatches prediction
src/gpu_econ/cli.py - CLI commands → FLOPs baseline: dispatches baseline
src/gpu_econ/cli.py - CLI commands → SLO recommender: dispatches recommendation
src/gpu_econ/cli.py - Bandwidth model → Run records: creates records
README.md - CLI commands → Run records: saves predictions
README.md - Run records → Predicted records: writes records
src/gpu_econ/records.py - Engineer → vLLM server: starts measurement
README.md - vLLM server → vLLM tooling: launches server
README.md - vLLM server → GPU hardware: runs on hardware
README.md - Engineer → Serving benchmark: runs benchmark
README.md - Serving benchmark → vLLM tooling: drives benchmark
README.md - vLLM server → Serving benchmark: serves workload
README.md - Serving benchmark → Record normalizer: produces benchmark JSON
README.md - Record normalizer → Run records: normalizes results
harness/measure.py - Record normalizer → Measured records: writes records
harness/measure.py - Engineer → Measurement sweep: runs sweep
README.md - CLI commands → Record comparison: dispatches validation
src/gpu_econ/cli.py - Predicted records → Record comparison: joins predictions
README.md - Measured records → Record comparison: joins measurements
README.md - Run records → Record comparison: loads records
validate/compare.py - Record comparison → Validation report: provides comparison
validate/report.py - CLI commands → Validation report: writes report
src/gpu_econ/cli.py