Профилирование DeepStream-пайплайнов с Nsight Systems

★ 7.6 · content

deepstream-profile-pipeline is a Claude Code skill that profiles DeepStream pipelines using Nsight Systems and NVTX traces, replacing guesswork with two measured numbers — inference plateau batch and hardware ceiling — to derive every other config value. It operates in six stages: Stage 0 pre-applies performance defaults (INT8/FP16 precision, NVMM memory, fakesink, perf-tuned NvDCF tracker) before the pipeline is generated, while Stages 1–5 measure and verify using `nsys profile` and `nsys stats` entirely headless, with no GUI dependency. The skill supports nvinfer and nvinferserver across detection, classification, segmentation, VLM, and embedding pipelines with file, RTSP, or USB camera sources. It requires DeepStream SDK 9.0 in the `nvcr.io/nvidia/deepstream:9.0-triton-multiarch` container, Nsight Systems 2024+, and nvidia-smi on PATH. Use it to maximize FPS, benchmark multi-stream throughput, or identify per-plugin bottlenecks on NVIDIA GPU without any graphical tools.