Diagnose Before You Compress: Prediction-Independent Bottleneck Witness Refinement for LLM Serving Traces

Liming Liu, Chao Hu, Mingfei Lu +7

Production LLM serving generates millions of diverse requests, making full-trace replay across serving configurations increasingly expensive. Existing trace reduction methods mainl…