Sakhanda Wire
NVDA $153.73 -1.54% MSFT $506.30 -2.26% GOOGL $280.09 -1.28% META $619.28 -1.14% AMZN $256.62 -0.44%
← Back to the news

Tutorial explores cuDNN Graph API fusion and plan reuse

A practical tutorial examines NVIDIA’s cuDNN Frontend Graph API for describing and executing deep-learning computations below the framework level. Operations are represented as graphs, while tensors are defined through their dimensions and strides. The workflow uses cuDNN to select an execution engine, then shows how developers can inspect and control that choice.

Each example follows a five-stage build process: validation, operation-graph construction, execution-plan creation, support checking, and plan building. The resulting graph is executed with a variant pack containing tensor pointers, and outputs are compared with PyTorch references on a single Colab GPU.

The examples begin with a fused convolution, bias operation, and ReLU. They then cover autotuning across engine configurations, including timing candidate plans and selecting a faster configuration for a frequently used shape. Later sections demonstrate FP8-style epilogues, scaled dot-product attention, plan serialization and reuse, dynamic shapes, and CUDA graph capture. The tutorial also addresses workspace allocation, tensor layouts, and benchmarking so that fusion correctness and execution cost can be examined together.

Written by Sakhanda Wire AI from reporting by MarkTechPost. Facts, quotes and full context are in the original article below.

Originally published by MarkTechPost on

Read the original on MarkTechPost ↗

The reporting and the image belong to MarkTechPost and are credited and linked above; the text on this page is Sakhanda Wire's own summary.

← Back to the news

More stories

All the latest news