Graphsignal Profiler

Graphsignal Profiler is a production-scale inference profiling platform that provides continuous, high-resolution profiling timelines exposing operation durations and resource utilization across AI inference workloads. It is designed for AI engineers and ML teams who need deep visibility into inference performance across models, engines, GPUs, and other accelerators. The profiler runs as a sidecar process alongside inference workloads, started with the `graphsignal-run` CLI or `graphsignal.watch()` from Python. It collects operation-level profiling data, LLM generation traces with per-step timing and token throughput, system-level metrics, and error monitoring for device-level failures. Data is sent to Graphsignal servers for post-processing and visualization at app.graphsignal.com. Key features include low-overhead CUDA kernel attribution and host sync wait detection, making it suitable for production deployments. It integrates with major inference frameworks such as NVIDIA PyTorch, vLLM, SGLang, and TensorRT-LLM. The tool also supports telemetry-driven autonomous optimization loops (autodebug) and integrates with AI agents like Claude Code for debugging. Unlike dev-time CUDA profilers, Graphsignal Profiler is built for production environments with minimal overhead. It focuses on inference profiling rather than training, filling a niche that general-purpose APM tools cannot address.