flâneur

mohit on X: "Day 2/30 of Inference Engineering - looked into where and why different CUDA Graph creation approaches are used, how this differs across PyTorch, JAX/TensorFlow and CUDA C++, and when stream capture vs explicit graph construction makes more sense - read about CUDA Graph https://t.co/OrNehU2cDA" / X

x.com · 127 words · saved by 1 readers

Day 2/30 of Inference Engineering - looked into where and why different CUDA Graph creation approaches are used, how this differs across PyTorch, JAX/TensorFlow and CUDA C++, and when stream capture vs explicit graph construction makes more sense - read about CUDA Graph

Day 2/30 of Inference Engineering - looked into where and why different CUDA Graph creation approaches are used, how this differs across PyTorch, JAX/TensorFlow and CUDA C++, and when stream capture vs explicit graph construction makes more sense - read about CUDA Graph constraints and limitations, including asynchronous/capture restrictions, static graph topology and parameters, memory requirements, multi-device considerations, and more - went through safe vs unsafe capture and explored some advanced CUDA Graph concepts like graph updates, device-side graph launch, and conditional nodes,…

saved by

related reading