NVIDIA
Engine building failure of TensorRT 10.2.0 (pip install) when building a custom diffusion model on RTX 4090
Description
Fresh install of pip install tensorrt==10.2.0
Following engine build crashes on Ubuntu 22.04.4 LTS:
from polygraphy.backend.trt import EngineFromNetwork
EngineFromNetwork(
network,
config=CreateConfig(fp16=fp16,
tf32=tf32,
int8=int8,
refittable=enable_refit,
profiles=[p],
load_timing_cache=timing_cache,
builder_optimization_level=3,
**extra_build_args
),
save_timing_cache=timing_cache
)()
Error message:
IBuilder::buildSerializedNetwork: Error Code 6: API Usage Error (Unable to load library: libnvinfer_builder_resource_win.so.10.2.0: libnvinfer_builder_resource_win.so.10.2.0: cannot open shared object file: No such file or directory)
Build works fine on 10.1.0 and 10.0.0
Environment
TensorRT Version: 10.2.0
NVIDIA GPU: RTX 4090
NVIDIA Driver Version: 550
CUDA Version: 12.1.r12.1
CUDNN Version: 8.9.7
Operating System: Ubuntu 22.04.4 LTS
Python Version (if applicable):
PyTorch Version (if applicable): 2.3.1
Baremetal or Container (if so, version):
Relevant Files
Model link:
Steps To Reproduce
Commands or scripts:
Have you tried the latest release?:
This is the latest release.
Can this model run on other frameworks? For example run ONNX model with ONNXRuntime (polygraphy run <model.onnx> --onnxrt):
Yes, the above command completes successfully, the ONNX file is correct.
+1. Same issue: tensorrt fails due to non-existent windows library in a linux distro (libnvinfer_builder_resource_win.so.10.2.0: cannot open shared object file: No such file or directory)
Almost exactly the same setup here, same problem. Using it via ComfyUI.
TensorRT Version: 10.2.0 NVIDIA GPU: RTX 4090 CUDA Version: 12.1.105 CUDNN Version: 8.9.2.26 Operating System: Ubuntu 22.04.3 Python Version (if applicable): 3.10 PyTorch Version (if applicable): 2.3.1+cu121
I did a little bit of research on this and determined that the non-Windows library (libnvinfer_builder_resource.so.10.2.0) was already opened by the process, so it's a real mystery to me why it was trying to open the Windows version. The dlopen (or whatever) is happening inside the tensorrt.so compiled code, not anything to do with the Python wrapper around it, so it's hard to debug farther.
I made a symlink from the proper DSO to the Windows filename, but that fixed nothing: The symbols that it then looks for inside are also suffixed with _win.
I asked in the discussion forum for the Comfy nodes... https://github.com/comfyanonymous/ComfyUI_TensorRT/discussions/49 But clearly they have nothing to do with it.
10.1.0 is also working for me on the setup outlined above where 10.2.0 did not.
Downgrading by running this command fixed the issue for me.
pip install tensorrt==10.1.0 tensorrt-cu12==10.1.0 tensorrt-cu12-bindings==10.1.0 tensorrt-cu12-libs==10.1.0 --force-reinstall
通过运行此命令降级为我解决了这个问题。
pip install tensorrt==10.1.0 tensorrt-cu12==10.1.0 tensorrt-cu12-bindings==10.1.0 tensorrt-cu12-libs==10.1.0 --force-reinstall
Solved my problem, thanks.
Resolved in Ultralytics package by pinning tensorrt<=10.2.0, but does not resolve underlying issue unfortunately.
https://github.com/ultralytics/ultralytics/pull/14239
通过运行此命令降级为我解决了这个问题。
pip install tensorrt==10.1.0 tensorrt-cu12==10.1.0 tensorrt-cu12-bindings==10.1.0 tensorrt-cu12-libs==10.1.0 --force-reinstallSolved my problem, thanks.
same!!! thank uuuu
libnvinfer_builder_resource_win.so.10.2.0: libnvinfer_builder_resource_win.so.10.2.0: cannot open shared object file: No such file or directory)
can you find libnvinfer_builder_resource_win.so.10.2.0 ?
The tensorrt Python wheel files only support Python versions 3.8 to 3.12 at this time and will not work with other Python versions. Only the Linux and Windows operating systems and the x86_64 CPU architecture are currently supported. These Python wheel files are expected to work on RHEL 8 or newer, Ubuntu 20.04 or newer, and Windows 10 or newer.
libnvinfer_builder_resource_win.so.10.2.0: libnvinfer_builder_resource_win.so.10.2.0: cannot open shared object file: No such file or directory)
can you find
libnvinfer_builder_resource_win.so.10.2.0?The tensorrt Python wheel files only support Python versions 3.8 to 3.12 at this time and will not work with other Python versions. Only the Linux and Windows operating systems and the x86_64 CPU architecture are currently supported. These Python wheel files are expected to work on RHEL 8 or newer, Ubuntu 20.04 or newer, and Windows 10 or newer.
No, those _win files dont exist on ubuntu.
Closing this ticket; please re-open if this is still an issue on TensorRT 10.8.
