A useful attribution step is to prove whether a latency spike happens before work reaches the GPU or while the GPU is executing it. Start with synchronized timestamps around four boundaries: request arrival, host allocation or tensor preparation, enqueue/submit, and completion. Keep the GPU timeline from the profiler separate from application wall time; otherwise a slow host allocator, page move, or Python-side serialization point can be mislabeled as a model-policy stall.

For a reproducible run, pin the process to one NUMA node, record the visible CUDA devices and driver/runtime versions, and log allocator settings such as expandable segments, memory pools, caching behavior, and any reserved-memory fraction. Capture baseline, single-task, and contended cases. On each run, collect host-side allocation counters plus GPU activity, memory-copy activity, context switches, and PCIe/NVLink counters. The key comparison is whether delayed requests show long host intervals before the first GPU kernel, gaps between kernels with no device work, or long actual kernel/runtime duration.

Classify the result explicitly. If the GPU timeline is empty while the application waits, investigate host allocation, memory pinning, NUMA crossing, lock contention, or dataloader overhead. If kernels are present but copies dominate, treat it as a transfer-path issue. If kernels themselves lengthen under identical inputs and occupancy, then look at SM saturation, memory bandwidth, clock limits, or neighboring workloads. Only after these system causes are ruled out should differences be attributed to agent policy or model behavior.