Profiling is an evidence-gathering phase, not a substitute for the neutral benchmark.
- Repeat the cold C4 configuration and confirm the ITL P95 outlier.
- Capture Nsight Systems traces for representative C1 and C4 cold/warm runs.
- Correlate request timestamps with scheduler batches, GPU kernels, KV-cache allocation, and queueing.
- Compare prefill/decode overlap and kernel launch behavior across engines.
- Record profiler version, capture command, image digest, driver, and CUDA versions beside the trace.
Do not turn an observed correlation into a root-cause claim until the trace supports it.