This article will focus on minimizing GPU memory footprint — for both optimization and inference workloads. Throughput and latency usually get all the attention, but reducing memory consumption without making architecture sacrifices is often just as valuable.
PyTorch in 2023 is a complex beast, with many great performance features hidden away. Simple top-N lists are weak content, so I’ve empirically tested the most important PyTorch tuning techniques and settings in all combinations. I’ve benchmarked inference across a handful of different model architectures and sizes, different versions of PyTorch and even different Docker containers.
This article is a high-level introduction to an efficient worfklow for optimizing runtime performance of machine learning systems running on the GPU. Using traces from Nsight Systems to show real production scenarios, I introduce a set of common utilization patterns and outline effective approaches to improve performance.
This article is a deep dive into the techniques needed to get SSD300 object detection throughput to 2530 FPS. We will rewrite Pytorch model code, perform ONNX graph surgery, optimize a TensorRT plugin and finally we’ll quantize the model to an 8-bit representation. We will also examine divergence from the accuracy of the full-precision model.
TorchScript is one of the most important parts of the Pytorch ecosystem, allowing portable, efficient and nearly seamless deployment. With just a few lines of torch.jit code and some simple model changes you can export an asset that runs anywhere libtorch does. It’s an important toolset to master if you want to run your models outside the lab at high efficiency. This article is a collection of topics going beyond the basics of your first export.
In this article we take performance of the SSD300 model even further, leaving Python behind and moving towards true production deployment technologies: TorchScript, TensorRT and DeepStream. We also identify and understand several limitations in Nvidia’s DeepStream framework, and then remove them by modifying how the nvinfer element works.
Making code run fast on GPUs requires a very different approach to making code run fast on CPUs because the hardware architecture is fundamentally different. Machine learning engineers of all kinds should care about squeezing performance from their models and hardware — not just for production purposes, but also for research and training. In research as in development, a fast iteration loop leads to faster improvement. This article is a practical deep dive into making a specific deep learning model (Nvidia’s SSD300) run fast on a powerful GPU server, but the general principles apply to all GPU programming.
Taking machine learning models into production for video analytics doesn’t have to be hard. A pipeline with reasonable efficiency can be created very quickly just by plugging together the right libraries. In this post we’ll create a video pipeline with a focus on flexibility and simplicity using two main libraries: Gstreamer and Pytorch.