The current PyTorch install supports CUDA capabilities sm_37 sm_50 sm_60 sm_61 sm_70 sm_75 compute_37

1

I’m trying to install some packages in Conda and run a python program. But when I try to run the program it throws this error:

UserWarning:
NVIDIA GeForce RTX 3070 with CUDA capability sm_86 is not compatible with the current PyTorch installation.
The current PyTorch install supports CUDA capabilities sm_37 sm_50 sm_60 sm_61 sm_70 sm_75 compute_37.
If you want to use the NVIDIA GeForce RTX 3070 GPU with PyTorch, please check the instructions at https://pytorch.org/get-started/locally/

  warnings.warn(incompatible_device_warn.format(device_name, capability, " ".join(arch_list), device_name))

I’m installing all the packages using these commands:

conda create -y -n styleflow python==3.6.7
conda activate styleflow
conda install -y git
git clone https://github.com/Tuba202/StyleFlow-Made-Easy.git StyleFlow
cd StyleFlow/
conda install -y pytorch==1.2.0 torchvision==0.4.0 cudatoolkit=10.2 -c pytorch
pip install torch==1.8.1+cu102 torchvision==0.9.1+cu102 -f https://download.pytorch.org/whl/torch_stable.html
conda install -y tensorflow==1.14 tensorflow-gpu==1.14
pip install torchdiffeq==0.1.0 tensorflow==1.14 tensorflow-gpu==1.14 scikit-image scikit-learn requests qdarkstyle qdarkgraystyle pyqt5 opencv-python

I’m trying to use this repo StyleFlow.

I tried to find solutions but nothing seems to work. Installing PyTorch and Tensorflow is the hardest thing I’ve ever tried.

2

In your install command you are explicitly specifying an old PyTorch release with the CUDA 10.2 runtime, which is too old for your Ampere GPU. Use any install command posted on our website and it would work since the default CUDA runtime for the current PyTorch release is 11.7.
I also don’t understand why you are trying to install multiple (old) PyTorch binaries from conda and pip in the same environment.

Installing PyTorch and Tensorflow is the hardest thing I’ve ever tried.

It is as easy as pip install torch and I’m unsure how you came up with your install command using old releases.

3

Hi, I meet the same problem. Because i want use fcos3d module built on mmdetection3d framework, when i use latest pytorch version, error reporting " THC/THC.h: No such file or directory". And i found this “remove THC.h by zyddnys · Pull Request #37 · CoinCheung/pytorch-loss · GitHub” tells the file has been removed after torch1.11. And i have been struggling for a long time.

4

Yes, the TH(C) namespace is deprecated and this header file was removed.
I see this issue raised the same error, but it also seems that mmdet3d/ops/ball_query/src/ball_query.cpp was removed and I didn’t see any includes for TH(C) anymore. Which file is raising the error in your build?

5

i have solved this problem by reinstall all env dependencies, thanks.

6

Pyt1
Pyt11109×426 47.7 KB

I face the same problem. As your explanation, we should consider the current cuda arc whille installing PyTorch and that is understandable. But the project developed with old PyTorch refuse to built on the new environment. The figure above is the sample I tried to execute. It works with CPU.

8

Could you explain this issue in more detail, please?

11

The following two errors happened with the higher cuda version (11.1)
/media/…/models/dla/DCNv2/src/cuda/dcn_v2_cuda.cu(126): error: identifier “THCudaBlas_SgemmBatched” is undefined

/media/…/models/dla/DCNv2/src/cuda/dcn_v2_cuda.cu(273): error: identifier “THCudaBlas_Sgemm” is undefined

RuntimeError: Error compiling objects for extension

12

The THC namespace was migrated to ATen, e.g. in this PR, so you would need to replace all legacy calls with their corresponding ATen ones. I’m not familiar with the repository you are using and thus also don’t know how much work it would be.

1 Like