Expand description
Model execution traits (Layer, lowering, kernel-launch argument packing).
Structs§
- Kernel
Instance - A concrete, dtype-resolved kernel produced by a
#[kernel(dtypes = [...])]dispatcher (or akernel_group!). - Kernel
Instance Backward - The backward half of a
KernelInstance, when a paired backward kernel is declared via#[kernel(..., backward = ...)].
Enums§
- Lowering
Mode - Whether a graph is being lowered for inference or training (training lowerings additionally wire up backward kernels).
Traits§
- Executable
Op - An op that has been lowered to a compilable kernel representation.
- Lowering
- Lowers a
Graphinto a DAG ofExecutableOps, ready to compile/execute. - Model
- A compiled, runnable model.
- Runtime
Context - A runtime execution context (device/stream handles, etc), threaded through arg-packing.
- Runtime
Op - Encapsulates the runtime-dispatch behaviour for a compiled op node: how many activation inputs it takes, what parameter buffers it needs, how to pack kernel arguments, and how to compute the launch grid.