This repository accompanies the paper describing BLIS, a collection of trace-only methods for estimating step-level execution coefficients for vLLM-style inference engines.

The core idea is to infer per-step execution costs using only production traces (phase start/end times and token counts), without engine instrumentation or per-step timing.

paper/    # LaTeX source for the paper
python/   # Reference Python implementation of the baseline estimator

• See paper/README.md for building the paper.
• See python/README.md to run the baseline estimator.

• The Python code implements the baseline time-integrated NNLS estimator described in the paper.
• Estimated coefficients parameterize a simulator; no trace timing is embedded in the simulator itself.

This repo is intended for research, reproducibility, and artifact evaluation.