Expand description
Loaded-model execution.
Structsยง
- Activation
Cache - Activation buffers retained from a
forward_traincall. - Adamw
Kernel - A pre-compiled
adamw_stepkernel ready for use inLoadedModel::adamw_step. - Compiled
Node - One PTX-compiled graph node: where its compiled kernel lives on disk, its output shape/dtype, and how to dispatch it at runtime.
- Cuda
Graph Model - A CUDA graph compiled from a
LoadedModelfor fixed-batch inference. - Cuda
Model - A compiled model: a DAG of
CompiledNodes, not yet loaded into GPU memory. CallCudaModel::loadto load it and get a runnableLoadedModel. - Loaded
Model - A
CudaModelwith all kernels loaded into GPU memory and parameter buffers allocated, ready to run inference (and, with thetrainingfeature, training steps). - Tensor
Ref - A reference to a device-side tensor: raw device pointer + concrete shape.