Hello, I want to build a new model for inference, but the activation after convolution is a new one. I use the Legacy API, but the speed is too slow (slower than Pytorch). So how to use the Graph API to build a model with a new activation?
Related topics
| Topic | Replies | Views | Activity | |
|---|---|---|---|---|
| How use the cudnn graph API for do a convolution | 2 | 67 | February 15, 2025 | |
| cudnnConvolutionBiasActivationForward API Usage | 2 | 1504 | May 1, 2018 | |
| Fail to call cuDNN or cuBLAS routine with explicit Graph APIs in CUDA Graph | 0 | 369 | May 19, 2022 | |
| Constructing CUDA Graphs with Dynamic Parameters | 1 | 420 | August 23, 2022 | |
| How to use graph API for cuDNN to do layer normalization? | 1 | 31 | February 28, 2025 | |
| Pytorch 1.2 cuda 10.0 vs. pytorch 1.9 cuda 11.1 significant slowdown | 3 | 937 | July 8, 2021 | |
| cuDNN update backend graph arguments / support for indirection pointers | 1 | 685 | June 24, 2021 | |
| Conceptual questions about how to use cuDNN v9 | 2 | 126 | January 3, 2025 | |
| Why not directly add an MHA Operation Descriptor? | 1 | 96 | June 30, 2024 | |
| Does cuGraph support Graph Convolutional Networks (GCNs)? | 2 | 491 | April 12, 2023 |