Extensibility

Write custom converters, register TensorRT plugins, override built-in converters, and embed custom CUDA/Triton kernels directly in serialized engines.

See also Handling Unsupported Operators for a decision guide on which approach to use.

• Converters
  • Writing Dynamo Converters
  • Converter Registry Internals
  • The impl/ Building-Block Library
  • Example: Auto-generate a Converter for a Custom Kernel
  • Example: Overloading Converters with Custom Converters
• Lowering Passes
  • Writing Dynamo ATen Lowering Passes
• Plugins
  • Plugin System
  • Example: Auto-generate a Plugin for a Custom Kernel
  • Example: Using Custom Kernels within TensorRT Engines
  • Automatically Generate a TensorRT AOT Plugin
  • Step 1: Define the Triton Kernel
  • Step 2: Register the PyTorch op
  • Step 3: Register the QDP Shape Descriptor
  • Step 4: Register the AOT Implementation
  • Step 5: Generate the Converter
  • Step 6: Compile and Run
  • Example: Custom Kernels with NVRTC in TensorRT AOT Plugins