<![CDATA[Production LLM traffic is bursty, and capacity decisions couple GPU availability with power and carbon. We convert probabilistic demand forecasts into risk-controlled admission, heterogeneous allocation, calibrated overbooking, and cross-site dispatch. The evaluation uses all 1,429,737 requests in the 61-day BurstGPT trace and SustainDC carbon, weather, workload, configuration, and physical power model. Requests form a complete five-minute grid with a weighted-token target. A chronological split assigns 42 days to fitting, nine to conformal calibration, and ten to testing. Point models include persistence, seasonality, Ridge, histogram gradient boosting, and Extra Trees; interval models include direct quantiles, split, rolling, regime-conditioned, and conformalized quantile regression. Histogram gradient boosting attained 49.68% WAPE versus 115.81% for daily seasonality and reduced mean absolute error by 31,545.90 weighted tokens/bin (95% block-bootstrap interval, 23,152.07–37,746.37). Rolling conformal achieved 90.00% coverage at mean width 86,728.29. With a calibration-selected 0.80 overbooking factor, the proposed policy recorded 5.11% bin-level SLO violations and reduced emissions from 52.51 to 12.57 tCO2e through SustainDC-aware dispatch. The remaining 8.81% workload shortfall quantifies burst risk; GPU type, site, and price remain scenario inputs.]]>
Copyrights © 2026
