graphsignal-top: a `top` for inference GPUs in the open-source Graphsignal profiler, kernels ranked by time, live, with readable names

A few weeks ago we posted the open-source Graphsignal profiler here: wrap vllm serve / sglang serve / whatever in graphsignal-run and it serves everything it measures as JSON, mainly so an agent can read it in a loop.

Turns out we kept opening the JSON ourselves to see what the GPU was doing. So now there is graphsignal-top:

graphsignal-run vllm serve <model> --port 8000 # one terminal
graphsignal-top # another

What you get: per-GPU util over time, memory, power, throttling; the engine's running/waiting requests, tok/s and TTFT/TPOT p50/p95; a GPU time split (kernels / graphs / memcpy / sync); and every kernel ranked by time over the last 5 s with a readable name and category, not just the mangled symbol. Tabs for CUDA graph replays, transfers and syncs.

Screenshot is Qwen3-32B on 2x RTX PRO 6000, tp 2, over PCIe. NCCL all-reduce is a quarter of GPU time and half the wall time is host sync waiting on it. You see that in one glance.

Works remotely too: graphsignal-top --host gpu-box against a profiler started with --listen-host, or port-forward over ssh. --once prints a single frame if you just want it in a log.

Repo: https://github.com/graphsignal/graphsignal