Импорт detection-модели в DeepStream с TRT-оптимизацией

★ 7.6 · content

deepstream-import-vision-model is a Claude Code skill that automates end-to-end integration of a vision model from HuggingFace or NVIDIA NGC into an NVIDIA DeepStream pipeline: it downloads ONNX or exports SafeTensors, builds a dynamic TRT engine via trtexec at BS=1 and BS=MAX_BS, generates a custom C++ nvinfer bbox parser, and runs single-stream and multi-stream benchmarks. The output follows a strictly defined directory structure containing the engine, parser, configs, and a benchmark report in three formats — Markdown, HTML, and PDF with five charts. Authored by NVIDIA CORPORATION (version 1.2.2, CC-BY-4.0 and Apache-2.0 licenses), the skill is scoped exclusively to object detection; classification, segmentation, and other architectures are rejected at startup based on config.json.