NVIDIA open-sources TensorRT Model Connect, two commands to TensorRT inference

JFPuget · x · 2026-08-19

NVIDIA open-sourced TensorRT Model Connect (Public Preview): supported HuggingFace models can reach end-to-end TensorRT inference in two commands, no intermediate ONNX export, with the bundle runnable via native C++ APIs. The entire project—implementations, perf tuning, tests, integrations, docs—was built with OpenAI Codex agents, with humans directing and reviewing.

TVM creator Tianqi Chen highlighted first-class TVM-FFI support, letting custom DSL and agent-generated kernels run inside TensorRT. The official tutorial shows replacing nodes in a Qwen3-8B graph with a TVM-FFI kernel and swapping a DistilBERT region with a CuTe DSL kernel—no TensorRT C++ required. Two paths exist: a recommended path using family-published versioned recipes, and an advanced path with manual graph inspection and node-ID typing. Both share the same backend; DSOs are bound at pipeline construction and can be swapped for ABI-compatible ones without rebuilding the bundle.

Related event: NVIDIA Open-Sources TensorRT Model Connect for Two-Command HF Deployment(2 posts)→

Original post →

More from Infra

Infra channel →