WARP: Workload-Aware Residual Profiling for Graph-Equivalent Recommender Inference

Hardware-aware graph optimization can accelerate recommender inference, but choosing among equivalent execution plans is vulnerable to inaccurate cost estimates and noisy microbenchmarks. We instantiate the bandwidth-ratio, roofline-plus-dispatch plan-ranking mechanism of the CAGR framework as the analytical prior of an executable selector and extend it with two-parameter residual calibration by batch size and categorical access context. Six equivalent embedding/interactions plans are validated numerically and evaluated using frozen parameters on synthetic CPU request streams. In an initial 16-context study, the calibrated selector reaches 2.55× geometric-mean speedup against an unfused eager reference and reduces the held-out latency-prediction error of the CAGR-style prior from 57.2% to 12.6%. A stronger fixed TorchScript baseline, however, achieves 2.60× in those same observations. We therefore perform a separate six-round, randomized-order remeasurement on fresh requests: WARP obtains 1.94×, whereas fixed TorchScript obtains 1.95×. Their relative speedup is 0.991× (descriptive stratified bootstrap interval [0.977, 1.000]). Thus calibration improves latency estimation but does not establish superiority over the strongest fixed candidate in this bounded library. The study records numerical checks, adaptation cost, raw timings, and the conditions under which policy ranking can become unstable. Results concern synthetic requests and one CPU, not production CTR accuracy or heterogeneous accelerators.