Is this page helpful?

Refitting Engines#

Refit updates constant weights in a serialized engine without a full rebuild. Weightless engines strip weights at build time and refit them on first launch (from custom buffers or ONNX). Enable kREFIT at build time, then use IRefitter or IParserRefitter at runtime.

Workflow diagram for refit and weightless engines from AOT build with kREFIT through deserialize, IRefitter weight updates, refitCudaEngine, and inference with existing execution contexts

Table 5 Refit-related deployment patterns#
Pattern	When to use	Related topics
Standard refittable engine (`kREFIT`)	Update weights in place after deployment without re-running the full builder	CPU-Only AOT and TensorRT-RTX Engines: weightless workflow uses `kREFIT` with `kSTRIP_PLAN`
Weightless engine (`kSTRIP_PLAN` + `kREFIT`)	Minimize installer size; ship a small engine and refit weights on first run	CPU-Only AOT and TensorRT-RTX Engines: Refitting a Weightless Engine
ONNX-driven refit (`IParserRefitter`)	Weight-stripped engine built from ONNX; refit weights from the same ONNX file	CPU-Only AOT and TensorRT-RTX Engines: Refitting a Weightless Engine Directly with ONNX Models

Build flags, refit workflows, serialization after refit, and IParserRefitter examples are documented in CPU-Only AOT and TensorRT-RTX Engines.