Open feolcn opened 3 months ago
We never encountered Seg Fault in our tests, I believe it could related to cuda version mismatch between torch and xformers, do you have the same (ie 11.8) in both?
As an additional check, could you please try running UniDepth with ConvNext backbone, namely change line 39 to "UniDepth_ConvNextL" .
Hi there @lpiccinelli-eth ,
I encounter the same issue as @feolcn.
With ConvNextL
backbone, this is the missing key pixel_encoder.mask_token
.
I can confirm both xformers
and torch
are with cuda 11.8.
The missing keys are just warnings you can totally ignore.
mask_token
and register_tokens
are just dummy nn.Parameter to comply with different backbones.
@lpiccinelli-eth: Any insights on segfault error? :smile:
I built conda
environment from scratch using the provided instructions.
I have a few questions to understand your setting:
I have a few questions to understand your setting:
- Which GPU are you using? one Rtx 3090 ubuntu 22.04
- Are you running locally or remotely? both locally and remotely
- Is it going OOM? (RAM since the model is first loaded on CPU) No OOM issue has occurred Thank you for your relpy
We tested on the same hardware, RTX 3090 with Ubuntu 22.04 and we had no seg fault.
Is it possible that you have conflicts in your environment (CUDA or torch / xformers mismatch)? Did you create a new one and follow the instructions in the README?
I have followed instructions regarding Conda and having same Segmentation fault (core dumped) error.
Can you please share your env and the system information (OS, python, cuda...)?
Now it works. Worked only when I downloaded both pytorch 2.2.0
and xformers
using conda
commands not pip
and commenting things related to torch
and nvidia
inside requirements.txt
.
By the way, thanks for your work.
When it was not working, did you install anything with conda, or only with pip? I suppose some inconsistencies of the conda env could have been a problem.
when it was not working, all dependencies and libraries were downloaded using pip
command just as README. However, this was inside conda
environment.
I am facing the same issue with using conda. I tried to use python venv but am facing this problem
ModuleNotFoundError: No module named 'unidepth'
. I see unidepth being installed via pip so I am unsure of why this is happening. Any help is appreciated!
We corrected the requirements.txt as @maliksyria suggested and we tested the installation copy-pasting what is written in the README with both conda (and installing unidepth via pip) and with python venv on 3 different cards (A6000, RTX3090, and RTX4090) and 2 different clusters and we never encountered Seg Fault or any other problems.
It would be of great help to the community if you could post your env, system (OS and GPU), and the specific commands you used when you face any errors, thanks!
We corrected the requirements.txt as @maliksyria suggested and we tested the installation copy-pasting what is written in the README with both conda (and installing unidepth via pip) and with python venv on 3 different cards (A6000, RTX3090, and RTX4090) and 2 different clusters and we never encountered Seg Fault or any other problems.
It would be of great help to the community if you could post your env, system (OS and GPU), and the specific commands you used when you face any errors, thanks!
Thank you for your work and response. I re-downloaded your project and set up a new environment using conda , following each step in the readme file. However, I encountered the same error. My setup is CUDA 11.8, Ubuntu 22.04, and NVIDIA 3090. Following maliksyria's suggestion, I used pip uninstall torch and then conda install torch, which made it work properly. I'm not sure of the reason, but I am very grateful.
I am also facing the same situation as @Davidyao99 . Have you solved it now? Can you help me
python ./scripts/demo.py
Torch version: 2.2.0+cu121
Downloading: "https://github.com/lpiccinelli-eth/unidepth/zipball/main" to /home/feol/.cache/torch/hub/main.zip
Instantiate: dinov2_vitl14
UniDepthV1_ViTL14 is loaded with:
Segmentation fault (core dumped)