Replace fixed chunked-prefill settings in an LLM serving engine with a feedback controller that decreases the number of prompt tokens processed per scheduling quantum as GPU saturation or long-context load increases. The controller targets a high-quantile bound on the absolute GPU-power ramp while preserving the existing peak-power ceiling and measuring the resulting latency-throughput tradeoff.
Replace a raw minibatch gradient with an unbiased control-variate estimator that subtracts predictable components of per-example gradients and adds back their exactly or cheaply known population mean. Select the control-variate directions using leading eigenvectors of an online covariance operator, rather than using arbitrary scalar baselines. This should reduce gradient variance at fixed batch size and permit fewer examples per optimization step.
Represent every pairwise training interaction as a directed edge and edge-color the interaction graph. Use one color class, or a small set of colors, as a microbatch: because a color class is a matching, no data example participates in two pairwise terms within that microbatch. This removes within-batch gradient collisions and provides a graph-dependent estimate of how many dependent strata are being combined.