TensorRT 是Nvidia 推出的跨 nv–gpu架构的半开源高性能AI 推理引擎框架/库，提供了cpp/python接口，以及用户自定义插件方法，涵盖了AI 推理引擎技术的主要方面。
TensorRT is a semi-open source high-performance AI inference engine framework/library developed by Nvidia, which spans across nv-gpu architectures.
Provides cpp/python interfaces and user-defined plugin methods, covering the main aspects of AI inference engine technology.
| topic | 主题 | notes |
|---|---|---|
| overview | 概述 | |
| layout | 内存布局 | |
| compute_graph_optimize | 计算图优化 | |
| dynamic_shape | 动态shape | |
| plugin | 插件 | |
| calibration | 标定 | |
| asp | 稀疏 | |
| qat | 量化感知训练 | |
| trtexec | OSS辅助工具 | |
| tool | 辅助脚本 | |
| runtime | 运行时 | |
| inferflow | 模型调度 | |
| mps | MPS | |
| deploy | 基于onnx部署流程， trt 工具使用 | |
| py-tensorrt | python tensorrt封装 | 解析 tensorrt __init__
|
| model_benchmark | 模型性能测试 | |
| cookbook | 食谱 | |
| incubator | 孵化器 | |
| developer_guide | 开发者指导 | |
| triton-inference-server | triton | |
| cuda | cuda编程 | |
| onnxruntime op | onnxrt 自定义op | 辅助图优化，layer输出对齐 |
Reference
https://docs.*nvi**dia.com/deeplearning/tensorrt/archives/
https://developer.nvi**d*ia.com/search?page=1&sort=relevance&term=
https://gith**ub*.com/HeKun-NVIDIA/TensorRT-Developer_Guide_in_Chinese/tree/main
https://docs.**nv*idia.com/deeplearning/tensorrt/migration-guide/index.html
https://developer.n*vi*d*ia.com/zh-cn/blog/nvidia-gpu-fp8-training-inference/
