Configuration Model
A pipeline configuration has two top-level sections:
shared_memory: named stream definitionskernels: processing stages wired between those streams
Kernel definitions
Each kernel consumes one trigger input stream and writes one output stream:
- name: scale_stage
kind: cpu.scale
input: input_frame
output: scaled_frame
parameters:
factor: 2.0
Supported fields:
name: unique kernel namekind: registered kernel kind such ascpu.scaleorgpu.affine_transforminput: trigger input stream nameoutput: output stream nameauxiliary: optional extra stream bindingsoperation: expression string forcustom_operationkernelsparameters: kind-specific parameter mappingread_timeout: trigger read timeout in secondspause_sleep: worker pause polling interval in seconds
Auxiliary inputs
Auxiliary inputs can be declared as a mapping when aliases matter:
auxiliary:
matrix: affine_transform_matrix
bias: affine_offset_vector
This is especially useful for custom_operation kernels and affine-style stages that read several non-trigger streams.
Graph validation rules
The loader and graph layer validate several important invariants before workers start:
shared-memory names must be unique
kernel names must be unique
every referenced stream must exist
a kernel cannot reuse the same stream as both input and output
a kernel cannot reuse the same auxiliary binding multiple times
a stream cannot have more than one producer kernel
CPU and GPU pipelines
The same pipeline topology can often be expressed for either CPU or GPU streams by switching stream storage and kernel kinds together.
CPU example:
storage: cpu
kind: cpu.scale
GPU example:
storage: gpu
gpu_device: cuda:0
kind: gpu.scale
Custom operations
cpu.custom_operation and gpu.custom_operation accept a restricted expression language for fused elementwise workflows.
Supported syntax:
elementwise
+,-,*,/unary
+and-matrix multiplication with
@abs,minimum,maximum,min,max, andclipnumeric constants
names bound through
inputandauxiliary
Where to look next
See CLI guide for validation and run flows.
See runtime guide for health states and snapshots.
See API reference for the configuration classes.