Installation

The easiest way to use the TensorRT RTX EP is via the Python wheel, which bundles the EP plugin and all TensorRT RTX runtime libraries — no separate TensorRT RTX installation required.

To build the EP from source instead, see the build guide.

Requirements

NVIDIA RTX GPU (Ampere / RTX 30xx or later)
NVIDIA GPU driver compatible with your installed CUDA version
Python 3.8+
pip install "onnxruntime>=1.24"

Install

Note: If you have onnxruntime-gpu installed, uninstall it first — it ships a different version of the onnxruntime module:
pip uninstall onnxruntime-gpu

CUDA 13 (default):

pip install "onnxruntime>=1.24"
pip install onnxruntime-ep-nv-tensorrt-rtx

CUDA 12:

pip install "onnxruntime>=1.24"
pip install onnxruntime-ep-nv-tensorrt-rtx-cu12

The default meta wheel (onnxruntime-ep-nv-tensorrt-rtx) installs the cu13 variant. CUDA 12 users must install the cu12 variant explicitly.

Quick start

import onnxruntime as ort
import onnxruntime_ep_nv_tensorrt_rtx as trt_ep

# Register the EP plugin
ort.register_execution_provider_library(trt_ep.get_ep_name(), trt_ep.get_library_path())

# Discover available TensorRT RTX devices
devices = [d for d in ort.get_ep_devices() if d.ep_name == trt_ep.get_ep_name()]
if not devices:
    raise RuntimeError("No TensorRT RTX EP devices found")

# Create a session with the EP
so = ort.SessionOptions()
so.add_provider_for_devices(devices, {})
sess = ort.InferenceSession("model.onnx", sess_opts=so)

The onnxruntime_ep_nv_tensorrt_rtx package exposes three helper functions:

Function	Returns
`get_ep_name()`	`"nv_tensorrt_rtx"` — the EP provider name used for registration and device filtering
`get_library_path()`	Absolute path to the bundled EP plugin DLL / SO
`get_ep_names()`	`["nv_tensorrt_rtx"]` — list form of the above

For the full Python and C++ usage patterns including provider options, see the integration guide.