TensorRT cu13

10.16.1.11 verified Mon Apr 27 auth: no python

NVIDIA TensorRT is a high-performance deep learning inference library. The cu13 variant targets CUDA 13.x. Current version 10.16.1.11, release cadence approximately monthly.

pip install tensorrt-cu13

Common errors

error ModuleNotFoundError: No module named 'tensorrt' ↓

cause The tensorrt package is not installed or installed under a different name (e.g., tensorrt-cu13 is required).

fix

Run 'pip install tensorrt-cu13'.

error ImportError: libcudart.so.xx: cannot open shared object file: No such file or directory ↓

cause CUDA runtime library not found or incompatible with the TensorRT variant.

fix

Install the matching CUDA version. For tensorrt-cu13, ensure CUDA 13.x is installed and LD_LIBRARY_PATH set correctly.

Warnings

breaking From TensorRT 10.13, CUDA 11.x support dropped. This cu13 variant requires CUDA 13.x; mixing with older CUDA will fail at import or runtime. ↓

fix Ensure your environment has CUDA 13.2+ (for TRT 10.16) or matching CUDA version for the TensorRT release.

breaking In TensorRT 10.14, pycuda was removed in favor of cuda-python. Code using pycuda will break. ↓

fix Replace pycuda imports with cuda-python equivalents, e.g., 'import pycuda.driver as cuda' -> 'import cuda'.

deprecated Plugin version 1 of several plugins (e.g., cropAndResizeDynamic, modulatedDeformConv) deprecated starting 10.12, will be removed in future versions. ↓

fix Migrate to IPluginV3-based version 2 plugins.

gotcha The package 'tensorrt-cu13' is CUDA-specific. Installing on a system without compatible CUDA runtime may cause import errors like 'Could not find cudart64_13.dll'. ↓

fix Install the correct TensorRT variant matching your CUDA version (e.g., tensorrt-cu12, tensorrt-cu11).

Install

pip install tensorrt-cu13==10.16.1.11

Imports

tensorrt
wrong
```
import tensorrt as trt
```
correct
```
import tensorrt
```
Using 'as trt' is a common convention, but not required.
Builder
```
import tensorrt; tensorrt.Builder
```

Quickstart

Minimal TensorRT setup showing logger, builder, and network creation.

import tensorrt as trt
import os

def get_engine():
    logger = trt.Logger(trt.Logger.WARNING)
    builder = trt.Builder(logger)
    network = builder.create_network(1 << int(trt.NetworkDefinitionCreationFlag.EXPLICIT_BATCH))
    print(f"TensorRT version: {trt.__version__}")
    
if __name__ == '__main__':
    get_engine()