Spaces:
Sleeping
Sleeping
Commit History
-
Add provenance fields to /health response
b4a2e6b
-
Persist rate limiter state across crash-restarts, harden state loading
392b662 -
Add infeasibility response shape (top_recommendation/infeasibility)
09d0905 -
Add API request-id tracing, rate limiting, and response provenance
9f8362f
-
Code cleanup
0623ff0 -
Close power-efficiency Pareto gap, add training-data disclosure, harden data lineage, cut Docker image 84%
Pareto recommender now ranks on (throughput, price, watts) with a user-selectable ranking objective (tokens/$, tokens/sec, tokens/watt, lowest cost per million tokens) instead of a fixed cost-efficiency sort. Adds a 3-state training-data-tier disclosure (none/below_floor/sufficient) so under-floor AMD GPUs are no longer reported identically to well-covered ones. Adds sMAPE/median APE alongside MAPE, a data manifest pinning MLPerf/calibration source snapshots, corpus SHA-256 + validation-metrics sidecar fields for model lineage, a roofline-vs-GBDT-vs-hybrid ablation validating the core architecture, and a real top-1 recommendation benchmark. Removes an unused 303 MB transitive CUDA dependency and unused ML libraries, shrinking the production Docker image from 2.88 GB to 452 MB. 55 reliability gates, 466 tests.3c35b12
-
GPU Perf Prophet v1 — cross-vendor GPU performance forecasting & recommendation engine
Roofline-physics + XGBoost efficiency-gap model predicting inference throughput across AMD Instinct and NVIDIA GPUs from MLPerf Inference v4.1-v6.0 data plus self-run AMD Dev Cloud MI300X calibration, served via FastAPI + Streamlit and deployed to Hugging Face Spaces.a4508e8