Graphsignal open-sources a GPU profiler designed for AI agents, not humans reading traces

l0g1cs · reddit · 2026-09-17

A team that has long tuned vLLM/SGLang/llama.cpp setups open-sourced Graphsignal, a sidecar GPU profiler built on the premise that the one doing the tuning is usually an agent that can't read an nsys timeline. Wrapping any command with graphsignal-run (no code changes, CUDA and ROCm) exposes a single JSON at 127.0.0.1:18259/signals with per-kernel/CUDA-graph/memcpy/sync timings, NVML metrics (utilization, VRAM, power, clocks, throttling, XID errors), engine Prometheus metrics, and console tracebacks. Hand the agent the repo's SKILL.md plus a goal ("why is my GPU at 40% during decode") and it runs the loop itself; --cuda-graph-trace node breaks kernels out of graph replay, and a lock-free probe header adds intra-kernel instrumentation. Fully local, no root, uploads nothing without an API key.

Original post →

More from coding & agent

coding & agent channel →