system-profile

What problem does it solve? Diagnosing performance bottlenecks in scripts, running processes, and GPU workloads requires choosing the right profiling tools and often writing custom instrumentation, which is time-consuming and error-prone. ## Core Features & Use Cases - Multi-target profiling: Profile Python scripts, running processes by PID, specific functions, or entire frameworks like vLLM serving end-to-end. - Tool and instrumentation hybrid: Combines external tools (cProfile, py-spy, nvidia-smi, nsys, torch.profiler, NCCL debug) with custom code instrumentation for CPU-GPU transfer, memory allocation, and NCCL collective measurement. - Structured reporting: Produces bottleneck tables across CPU, memory, interconnect, and GPU compute dimensions, plus a mandatory changelog of all instrumentation edits for easy review and cleanup. - Use Case: Ask to profile a training script and receive hotspot analysis, GPU utilization metrics, communication-to-computation ratios, and ranked optimization recommendations. ## Quick Start Ask the agent to profile train.py and summarize the CPU, memory, and GPU bottlenecks with recommendations.