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

  /home/ict-526/anaconda3/envs/neucon/lib/python3.7/site-packages/torch/cuda/__init__.py:125: UserWarning:
  NVIDIA GeForce RTX 3080 Ti 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 3080 Ti GPU with PyTorch, please check the instructions at https://pytorch.org/get-started/locally/

cuda version 확인 (11.6)

username@computername:~/ $ nvcc -V
nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2022 NVIDIA Corporation
Built on Thu_Feb_10_18:23:41_PST_2022
Cuda compilation tools, release 11.6, V11.6.112
Build cuda_11.6.r11.6/compiler.30978841_0

 

gpu와 잘 붙었는지 확인.(python 코드이므로, idle에서 실행)

import torch

torch.cuda.is_available()
>>> True

torch.cuda.current_device()
>>> 0

torch.cuda.device(0)
>>> <torch.cuda.device at 0x7efce0b03be0>

torch.cuda.device_count()
>>> 1

torch.cuda.get_device_name(0)
>>> 'GeForce GTX 950M'

내가 생각한 해결책: neural reconstruction은 environment.yaml 파일을 제공한다. 내용은 다음과 같다.

name: neucon
channels:
  # You can use the TUNA mirror to speed up the installation if you are in mainland China.
  # - https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud/pytorch
  - pytorch
  - defaults
  - conda-forge
dependencies:
  - python=3.7.9
  - pytorch=1.6.0
  - torchvision=0.7.0
  - cudatoolkit=10.2
  - ipython
  - tqdm
  - numba
  - sparsehash  # dependency for torchsparse
  - pip
  - pip:
    - -r requirements.txt
    - git+https://github.com/mit-han-lab/torchsparse.git

3080, 3080ti, 3090등은 cuda버전이 11이상이어야 하고, 이에 맞는 torch 버전은 최신 1.9 버전, 이에 맞는 torchsparse는 1.4이므로, 다음과 같이 바꿔본다.

name: neucon
channels:
  # You can use the TUNA mirror to speed up the installation if you are in mainland China.
  # - https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud/pytorch
  - pytorch
  - defaults
  - conda-forge
dependencies:
  - python=3.9.3
  - pytorch=1.7.0
  - torchvision=0.7.0(??????????)
  - cudatoolkit=11.6
  - ipython
  - tqdm
  - numba
  - sparsehash  # dependency for torchsparse
  - pip
  - pip:
    - -r requirements.txt
    - git+https://github.com/mit-han-lab/torchsparse.git