TensorRT-RTX Operators#
In TensorRT-RTX, operators represent distinct flavors of mathematical and programmatic operations. The following sections describe every operator that TensorRT-RTX supports. The minimum workspace required by TensorRT-RTX depends on the operators used by the network.
Workspace Size: The default workspace size is 16 MB. For most networks, this is sufficient. However, complex networks with many operators may require a larger workspace. A suggested minimum build-time setting is 16 MB, but you can increase the workspace pool limit by calling IBuilderConfig::setMemoryPoolLimit(MemoryPoolType::kWORKSPACE, poolSize) if needed. Regardless of the workspace pool limit provided to the builder, TensorRT-RTX will allocate at runtime no more than the workspace it actually requires.
Interface - TensorRT-RTX Operators
See also
- Support Matrix
Supported data types and hardware compatibility per operator.
- C++ API
Programmatic access to operators through the C++ interface.
- Python API
Programmatic access to operators through the Python interface.
- Work with Quantized Types
Operator support for reduced-precision types (INT4, INT8, FP4, FP8, BF16).