graphsignalgraphsignalOfficialยท1 Agent Skills Included

graphsignal-profiler

GPU inference profiling, tracing, and performance monitoring

Profiles GPU inference workloads like vLLM, SGLang, PyTorch, and TensorRT-LLM from a lightweight sidecar process. Captures kernel-level timelines, token throughput, latency breakdowns, and hardware metrics without slowing production. Eliminates blind guesswork when diagnosing slow models, GPU bottlenecks, and device-level errors. Wraps any launch command with one CLI prefix for instant visibility.
npx skills add graphsignal/graphsignal-profiler --all -g -y

All Skills in This Repository (1)

Pure Emerald Level Indicators

Frequently Asked Questions

FAQPage Schema
How to install Graphsignal Profiler?โ–ผ

Run `npx skills add graphsignal/graphsignal-profiler --all -g -y` in your terminal to install this skill globally.

How to profile a vLLM or SGLang server?โ–ผ

Set your GRAPHSIGNAL_API_KEY environment variable, then prefix your launch command with `graphsignal-run`, for example `graphsignal-run vllm serve <model> --port 8000`.

Does Graphsignal slow down production inference?โ–ผ

No. GPU activity is collected through low-overhead CUPTI APIs, and all analysis and uploading happens in a separate sidecar process.

Can I profile PyTorch scripts without changing code?โ–ผ

Yes. Run your script with `graphsignal-run python my_app.py`, or call `graphsignal.watch()` once at startup before any CUDA work begins.

Does Graphsignal record prompts or completions?โ–ผ

No. It only sends performance metrics and traces to api.graphsignal.com over outbound connections; prompt and completion content is never recorded.

Related Repositories in Software Engineering

View All in Software Engineeringโ†’