Live campaign signal

Distributed performance readout with throughput, communication, and idle-gap decomposition.

This page tracks the most recent campaign runs and exposes where time is spent in the execution loop. The objective is simple: keep compute dominant while minimizing communication and idle fractions.

Loading distributed run summary...

Key Visuals

All visuals below are sourced from published artifacts in this repository and refreshed by the site publish pipeline.

Latest run summary chart — Latest run summary SVG

Latest distributed throughput chart — Distributed throughput ranking (recent runs)

Latest distributed pipeline breakdown chart — Latest run stage breakdown: compute, comm, idle, and H2D

Compiler Adapter Baseline

Adapter-level benchmark summaries are shown for CPU and GPU paths to anchor distributed campaign interpretation.

CPU

Adapter	Mode	Mean (ms)	P50 (ms)	P95 (ms)	Throughput

GPU

Adapter	Mode	Mean (ms)	P50 (ms)	P95 (ms)	Throughput

Recent Distributed Runs

Run ID	World	Samples/s	Tokens/s	Comm (ms)	Idle (ms)

Raw files: manifest.json | latest-summary.json | distributed-latest.json