[bug] CUDA not Working on GeForce RTX 5060

[bam91]

Describe the bug
When trying to use GPU enabled features(like ImageDetectionPrompt), I geht this Error in my Log and the Step fails:
RuntimeError: CUDA error: no kernel image is available for execution on the device

To Reproduce
Steps to reproduce the behavior:

1. Create Project with ImageDetectionPromt Step
2. Start Project
3. Error

Expected behavior
My Graphiccard has CUDA. This should work.

Log

RuntimeError: CUDA error: no kernel image is available for execution on the device

Desktop (please complete the following and other pertinent information):

• OS: Windows 11
• Qt/PySide version [e.g. 6.8.2]
• Meshroom version: 2025.1.0
• GeForce Driver: 576.88
• Cuda Toolkit 12 installed

[alexgubanow]

same for me with rtx 5070.
mrSegmentation has cuda 12.4, but we need at least 12.8
updating the plugin as per https://github.com/meshroomHub/mrSegmentation/blob/main/INSTALL.md leads to:
ImportError: Failed to load PyTorch C extensions

[t3hchipmunk]

same for me with rtx 5070. mrSegmentation has cuda 12.4, but we need at least 12.8 updating the plugin as per https://github.com/meshroomHub/mrSegmentation/blob/main/INSTALL.md leads to: ImportError: Failed to load PyTorch C extensions

I updated the plugin following your link but made the following change to the torch-requirements.txt and am not having any issues now.

-i https://download.pytorch.org/whl/cu128
torch==2.7.0
torchvision==0.22.*
``` #

[jwcrawley]

Same as well. I get the "RuntimeError: CUDA error: no kernel image is available for execution on the device" as well.

Running Xubuntu 24.04
NVIDIA GeForce RTX 5060 Ti
Driver Version: 580.95.05
CUDA Version: 13.0

[yonhdee]

Same as well. running alicevision/meshroom:2025.1.0-av3.3.0-ubuntu22.04-cuda12.1.1 on manjero
RTX5060Ti 16G
NVIDIA-SMI 590.48.01 Driver Version: 590.48.01 CUDA Version: 13.1

[gnbl]

I have this issue with Meshroom-2025.1.0-Linux.tar.gz on Linux Mint 22.2 Cinnamon and a RTX 5060 Ti (nvidia-driver-580-open 580.95.05).
The directory plugins/mrSegmentation/ is a virtual environment, but activating it after changing the bin/python link and trying to pip install torch-reqirements.txt as modified by t3hchipmunk, pip gives me an error about an externally managed environment.
--break-system-packages seems to install to /usr/lib/python3/dist-packages/.

Is it possible to upgrade torch and torchvision in the plugin venv?

[knobby4444]

I'm running Linux Mint 22.2 with 5070Ti 16GB and getting the error. I was able to update mrSegmentation as above, but still getting the issue. Ran these commands:

############
cd /opt/Meshroom-2025.1.0/plugins/mrSegmentation/
sudo apt install python3.12-venv
python3 -m venv venv
source venv/bin/activate
vi torch-requirements.txt

The torch-requirements.txt file was not there for me, created it as below

-i https://download.pytorch.org/whl/cu128
torch==2.7.0
torchvision==0.22.*

pip install -r torch-requirements.txt --upgrade --force-reinstall
##############
Install ran without errors and here is the end of the output:

#################
Installing collected packages: nvidia-cusparselt-cu12, mpmath, typing-extensions, sympy, setuptools, pillow, nvidia-nvtx-cu12, nvidia-nvjitlink-cu12, nvidia-nccl-cu12, nvidia-curand-cu12, nvidia-cufile-cu12, nvidia-cuda-runtime-cu12, nvidia-cuda-nvrtc-cu12, nvidia-cuda-cupti-cu12, nvidia-cublas-cu12, numpy, networkx, MarkupSafe, fsspec, filelock, triton, nvidia-cusparse-cu12, nvidia-cufft-cu12, nvidia-cudnn-cu12, jinja2, nvidia-cusolver-cu12, torch, torchvision
Successfully installed MarkupSafe-2.1.5 filelock-3.20.0 fsspec-2025.12.0 jinja2-3.1.6 mpmath-1.3.0 networkx-3.6.1 numpy-2.3.5 nvidia-cublas-cu12-12.8.3.14 nvidia-cuda-cupti-cu12-12.8.57 nvidia-cuda-nvrtc-cu12-12.8.61 nvidia-cuda-runtime-cu12-12.8.57 nvidia-cudnn-cu12-9.7.1.26 nvidia-cufft-cu12-11.3.3.41 nvidia-cufile-cu12-1.13.0.11 nvidia-curand-cu12-10.3.9.55 nvidia-cusolver-cu12-11.7.2.55 nvidia-cusparse-cu12-12.5.7.53 nvidia-cusparselt-cu12-0.6.3 nvidia-nccl-cu12-2.26.2 nvidia-nvjitlink-cu12-12.8.61 nvidia-nvtx-cu12-12.8.55 pillow-12.0.0 setuptools-70.2.0 sympy-1.14.0 torch-2.7.0+cu128 torchvision-0.22.0+cu128 triton-3.3.0 typing-extensions-4.15.0
################

Run Meshroom and try image detect prompt and get this output:

##########################
[2026-01-19 08:51:41.167969] [0x000078483de49740] [trace] Embedded OCIO configuration file: '/opt/Meshroom-2025.1.0/aliceVision/share/aliceVision/config.ocio' found.
[QtAliceVisionImageIO] Plugin Initialized
[QtAliceVision] Plugin Initialized
Starting Process for 'ImageDetectionPrompt_1'

• commandLine: /opt/Meshroom-2025.1.0/meshroom_compute /home/knobby/Documents/SG-Soldier.mg --node ImageDetectionPrompt_1 --extern --inCurrentEnv --iteration 0
• logFile: /home/knobby/Documents/MeshroomCache/ImageDetectionPrompt/42c98c827ee9a46e3f1c0677cb1fe256a48f1da2/0.log
• command full path: /opt/Meshroom-2025.1.0/meshroom_compute
  ERROR:root:Error on node computation: Error on node "ImageDetectionPrompt_1(0)":
  Log:
  [08:51:55.404][info] Chunk range from 0 to 49
  info:ImageDetectionPrompt_1:Chunk range from 0 to 49
  The new embeddings will be initialized from a multivariate normal distribution that has old embeddings' mean and covariance. As described in this article: https://nlp.stanford.edu/~johnhew/vocab-expansion.html. To disable this, use mean_resizing=False
  [2026-01-19 08:51:59.112940] [0x00007b0e8548eb80] [debug] [IO] Read Image: /home/knobby/Pictures/SgGonzales/Soldier/darktable_exported/DSC04058.jpg
  [2026-01-19 08:51:59.151304] [0x00007b0e8548eb80] [trace] Detected image color space: sRGB
  [2026-01-19 08:51:59.151328] [0x00007b0e8548eb80] [trace] Read image /home/knobby/Pictures/SgGonzales/Soldier/darktable_exported/DSC04058.jpg (encoded in sRGB colorspace).
  INFO:root: - elapsed time: 0:00:05.711590
  Traceback (most recent call last):
  File "/usr/local/lib64/python3.11/site-packages/cx_Freeze/initscripts/startup.py", line 140, in run
  File "/opt/Meshroom/setupInitScriptUnix.py", line 47, in run
  File "bin/meshroom_compute", line 101, in
  File "/opt/Meshroom/meshroom/core/node.py", line 502, in process
  File "/opt/Meshroom-2025.1.0/plugins/mrSegmentation/meshroom/imageSegmentation/ImageDetectionPrompt.py", line 201, in processChunk
  bboxes, conf, tags = processor.process(image = img,
  ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/opt/Meshroom-2025.1.0/plugins/mrSegmentation/segmentationRDS/segmentation.py", line 378, in process
  recoOK, tags = self.recognize(wordlist, image, verbose)
  ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/opt/Meshroom-2025.1.0/plugins/mrSegmentation/segmentationRDS/segmentation.py", line 325, in recognize
  tags = self.get_tags(image)
  ^^^^^^^^^^^^^^^^^^^^
  File "/opt/Meshroom-2025.1.0/plugins/mrSegmentation/segmentationRDS/segmentation.py", line 315, in get_tags
  res = inference(ram_image.to(self.DEVICE), self.ram)
  ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/opt/Meshroom-2025.1.0/plugins/mrSegmentation/lib64/python3.11/site-packages/ram/inference.py", line 36, in inference_ram
  tags, tags_chinese = model.generate_tag(image)
  ^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/opt/Meshroom-2025.1.0/plugins/mrSegmentation/lib64/python3.11/site-packages/ram/models/ram_plus.py", line 284, in generate_tag
  image_embeds = self.image_proj(self.visual_encoder(image))
  ^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/opt/Meshroom-2025.1.0/plugins/mrSegmentation/lib64/python3.11/site-packages/torch/nn/modules/module.py", line 1736, in _wrapped_call_impl
  return self._call_impl(*args, **kwargs)
  ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/opt/Meshroom-2025.1.0/plugins/mrSegmentation/lib64/python3.11/site-packages/torch/nn/modules/module.py", line 1747, in _call_impl
  return forward_call(*args, **kwargs)
  ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/opt/Meshroom-2025.1.0/plugins/mrSegmentation/lib64/python3.11/site-packages/ram/models/swin_transformer.py", line 563, in forward
  x = self.patch_embed(x)
  ^^^^^^^^^^^^^^^^^^^
  File "/opt/Meshroom-2025.1.0/plugins/mrSegmentation/lib64/python3.11/site-packages/torch/nn/modules/module.py", line 1736, in _wrapped_call_impl
  return self._call_impl(*args, **kwargs)
  ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/opt/Meshroom-2025.1.0/plugins/mrSegmentation/lib64/python3.11/site-packages/torch/nn/modules/module.py", line 1747, in _call_impl
  return forward_call(args, **kwargs)
  ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/opt/Meshroom-2025.1.0/plugins/mrSegmentation/lib64/python3.11/site-packages/ram/models/swin_transformer.py", line 448, in forward
  x = self.proj(x).flatten(2).transpose(1, 2) # B PhPw C
  ^^^^^^^^^^^^
  File "/opt/Meshroom-2025.1.0/plugins/mrSegmentation/lib64/python3.11/site-packages/torch/nn/modules/module.py", line 1736, in _wrapped_call_impl
  return self._call_impl(*args, **kwargs)
  ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/opt/Meshroom-2025.1.0/plugins/mrSegmentation/lib64/python3.11/site-packages/torch/nn/modules/module.py", line 1747, in _call_impl
  return forward_call(*args, **kwargs)
  ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/opt/Meshroom-2025.1.0/plugins/mrSegmentation/lib64/python3.11/site-packages/torch/nn/modules/conv.py", line 554, in forward
  return self._conv_forward(input, self.weight, self.bias)
  ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/opt/Meshroom-2025.1.0/plugins/mrSegmentation/lib64/python3.11/site-packages/torch/nn/modules/conv.py", line 549, in _conv_forward
  return F.conv2d(
  ^^^^^^^^^
  RuntimeError: CUDA error: no kernel image is available for execution on the device
  CUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect.
  For debugging consider passing CUDA_LAUNCH_BLOCKING=1
  Compile with TORCH_USE_CUDA_DSA to enable device-side assertions.
  ################################

I'm running mint versions of nvidia drivers, not on cuda 12.8:

nvcc --version
nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2023 NVIDIA Corporation
Built on Fri_Jan__6_16:45:21_PST_2023
Cuda compilation tools, release 12.0, V12.0.140
Build cuda_12.0.r12.0/compiler.32267302_0

NVIDIA Driver Version: 580.95.05

[knobby4444]

I tried installed Debian 13.1 with nvidia drivers installed like this: https://docs.nvidia.com/datacenter/tesla/driver-installation-guide/debian.html and have cuda 13.1. Unfortunately it appears 13.1 is not yet supported by pytorch. I see 13.0 but not yet 13.1 here: https://download.pytorch.org/whl/. Have not yet found a way to downgrade cuda to 13.0.

[hryamzik]

Looks like all 50xx cards are affected, I get the same on 5080 and windows 11, looking forward to see a new build with a fresh CUDA version.

[1e100]

This seems extra flaky because it did actually work for me on a 5090 a month ago, but does not work now even if I provide CUDA_HOME to point to cuda-12. Previously however, I was using it for things that do not require PyTorch - just plain photogrammetry. Now I'm trying to use it for turntable photogrammetry.

[MarcosRRM]

Having the same problem on Windows 11 25H2 with a RTX 5090.

[brosequist]

I opened a draft PR in the bundled segmentation plugin that targets this RTX 50-series failure: meshroomHub/mrSegmentation#50

The issue appears to be that the plugin's optional GPU requirements currently install PyTorch CUDA 12.4 wheels, which can expose CUDA but do not include kernels for RTX 50-series / Blackwell devices. The PR updates the optional GPU requirements to PyTorch CUDA 12.8 and adds a CUDA architecture guard so unsupported PyTorch builds fall back with a clear message instead of failing later with no kernel image is available for execution on the device.

[seebeck88]

I had the same problem, I'm not an expert so the terminology I use can be no the right one but I'll explain what I think is the issue, and how I solved it. I also have RTX50 series

Mesroom comes with its own Python 3.11 and PyTorch 2.5.1 that version of PyTorch needs CUDA 12.4. The problem is RTX 50s only works with versions of CUDA 12.8 and higher, the latest CUDA version that Meshroom can use is 12.9, so basically you can use only 12.8 and 12.9 CUDA tool kit versions, but if you install CUDA 12.8 or 12.9 + PyTorch 2.8.0 Meshroom won't be able to reach PyTorch because PyTorch is just installed for the Python on your system not for the Python Meshroom comes with, so you need to also install the PyTorch library for the Python Meshroom use.

Install Pytorch was kind of a challenge for me because I'm not familiar with how Python works and I has to do some quick research to learn the basics and install PyTorch. So if is your case to there are a lot of tutorials out there about how to do it. this video help me a lot https://www.youtube.com/watch?v=ibS3My-nbhc in the video explains how to install CUDA, PyTorch and cuDNN, I didn't installed that last one since I didn't think it was a Meshroom requirement. In the part of the video that explains how to choose PyThor don’t use the chart since it won't show the 12.8 or 12.9 version, below the chart there is an option that says "Previous versions of PyTorch" click there then look for the version 12.8 or 12.8 for for windows and Linux. and use that command instead of the command they use in the video.

The fix

-First, I recommend you install the same version of Python that comes with Meshroom that is 3.11, Meshroom use its own Python

-Second, you need to match Pytorch and CUDA version, Meshroom only support until CUDA 12.9 soy download and install the CUDA took kit 12 version you want and then you need to install the PyTorch version that match that version of CUDA

Example for CUDA 12.9
"pip install torch==2.8.0 torchvision==0.23.0 torchaudio==2.8.0 --index-url https://download.pytorch.org/whl/cu129"

-Third, Meshroom comes with is own Python so when I was trying to solve the problem I test a lot of versions of CUDA and Pytorch but even before running the test successfully and be sure everything was good, it didn't solve the problem, in that point I figured out the problem was Meshroom was not able to reach PyTorch, so I inspect the folders on Meshroom looking for clues, and I found a Torch library on this address:

"Meshroom-2025.1.0\plugins\mrSegmentation\venv\Lib\site-packages"

and also found a folder with the same name on the Python installed on my system with this address:

"Python311\Lib\site-packages"

So I copied the "site-packages" from the Python installed on my system and paste and merge it with the "site-packages" inside Meshroom.

That was how I solve it.

As I said before I'm not an expert of any kind but my intuition told me "if I can't install Torch directly in the Python Meshroom use, maybe I can put it there manually" and it work.

I also learn that the folder "Meshroom-2025.1.0>plugins>mrSegmentation>venv" contains a Python environment or wherever is the right terminology, though you can't run this Python.exe, because in the venv folder there is a file name Pyvenv this file is creates when Meshroom is packed an it is linked to the python folder that was the machine it was packed in so if you try to run with cmd Python won't finde the way home so I put the right address on that file and I was able to run a successful test with the Python that comes with Meshroom. I could've provably installed PyTorch if I could've learned how to run the Python Meshroon use before using the copy paste folder method but as it worked I didn't touch anything after that.

It might not be the most orthodox way to do but it worked.