graphsignal/graphsignal-python

Captures continuous operation-level profiling timelines, LLM generation traces with per-step timing and token metrics, and system-level resource utilization across inference workloads. Integrates with PyTorch, vLLM, and SGLang to automatically instrument inference engines while supporting manual tracing APIs. Uses low-overhead profiling and sends collected metrics to a cloud platform for performance analysis and AI-assisted debugging.

204 stars.