Description
Graphsignal is a GPU profiler for production inference. It watches how models run on real serving infrastructure and reports what actually happened, rather than what a local benchmark suggests, so you can see problems such as KV cache pressure, preemptions and slow time to first token under live traffic. Its profiler supports NVIDIA and AMD GPUs and common serving engines including vLLM, SGLang and TensorRT, and it works with PyTorch. Its Signals API lets an AI agent read production data, such as latency percentiles, engine queue lengths, kernel timings and deploy tags, and suggest changes to flags or code. The site describes an agent-driven loop in which the agent finds what to fix, changes the configuration and ships the result, while a replay of the production request log can be used to test changes before they reach users. Install the profiler with a single command and an API key, then open the Docs or the production loop guide for setup. It is aimed at teams that run large language models in production and want to tune them against real traffic.…
Highlights
- Profiles GPUs under real production traffic
- Supports NVIDIA and AMD GPUs
- Works with vLLM, SGLang, TensorRT and PyTorch
- Signals API for agents and scripts
- Replays production requests to test changes
Ratings & Reviews
No ratings yet
Be the first to rate Graphsignal.
People also visit
Time To First Token
A 10-week, 30-minutes-a-day roadmap for LLM inference serving and optimization
Graphsignal FAQ
What is Graphsignal?
A production GPU profiler for AI inference that reports what models actually did on live traffic.
Which hardware and engines does it support?
NVIDIA and AMD GPUs, with vLLM, SGLang, TensorRT and PyTorch named on the site.
How do I install it?
Install the profiler from the site, using your API key. The Docs cover setup.
Is the code public?
The company publishes code on GitHub under graphsignal.
What are the best Graphsignal alternatives?
Popular alternatives to Graphsignal include Inference.net, Sipp and LLMKube. See all Graphsignal alternatives








