在Anaconda中使用Pytork GPU时，是否不需要安装CUDA？

我发现在Anaconda中安装Pytorch 0.4 GPU版本后，不需要在本地安装CUDA即可调用GPU加速。在运行代码时，GPU内核的使用率可以达到90%以上

编辑：我在Windows10中使用过它。不知道它是否在Linux下工作。

@talonmies

谢谢你的网址。pytorch在Windows中似乎不需要cuda，因为它的依赖项是cffi、mkl、numpy和python

我在Anaconda提示符中输入了conda search-c pytorch pytorch=0.4.0-info命令，它说

Loading channels: done
pytorch 0.4.0 py35_cuda80_cudnn7he774522_1
------------------------------------------
file name   : pytorch-0.4.0-py35_cuda80_cudnn7he774522_1.tar.bz2
name        : pytorch
version     : 0.4.0
build string: py35_cuda80_cudnn7he774522_1
build number: 1
size        : 528.5 MB
arch        : x86_64
constrains  : ()
platform    : Platform.win
license     : BSD 3-Clause
subdir      : win-64
url         : https://conda.anaconda.org/pytorch/win-64/pytorch-0.4.0-py35_cuda80_cudnn7he774522_1.tar.bz2
md5         : 7db3971bb054079d7c7ff84b6286c58e
dependencies:
  - cffi
  - mkl >=2018
  - numpy >=1.11
  - python >=3.5,<3.6.0a0


pytorch 0.4.0 py35_cuda90_cudnn7he774522_1
------------------------------------------
file name   : pytorch-0.4.0-py35_cuda90_cudnn7he774522_1.tar.bz2
name        : pytorch
version     : 0.4.0
build string: py35_cuda90_cudnn7he774522_1
build number: 1
size        : 578.5 MB
arch        : x86_64
constrains  : ()
platform    : Platform.win
license     : BSD 3-Clause
subdir      : win-64
url         : https://conda.anaconda.org/pytorch/win-64/pytorch-0.4.0-py35_cuda90_cudnn7he774522_1.tar.bz2
md5         : 8200c9841f9cad6f2e605015812aa3f2
dependencies:
  - cffi
  - mkl >=2018
  - numpy >=1.11
  - python >=3.5,<3.6.0a0


pytorch 0.4.0 py35_cuda91_cudnn7he774522_1
------------------------------------------
file name   : pytorch-0.4.0-py35_cuda91_cudnn7he774522_1.tar.bz2
name        : pytorch
version     : 0.4.0
build string: py35_cuda91_cudnn7he774522_1
build number: 1
size        : 546.1 MB
arch        : x86_64
constrains  : ()
platform    : Platform.win
license     : BSD 3-Clause
subdir      : win-64
url         : https://conda.anaconda.org/pytorch/win-64/pytorch-0.4.0-py35_cuda91_cudnn7he774522_1.tar.bz2
md5         : 79d99a825f66b55b1aa6f04d22d68aac
dependencies:
  - cffi
  - mkl >=2018
  - numpy >=1.11
  - python >=3.5,<3.6.0a0


pytorch 0.4.0 py36_cuda80_cudnn7he774522_1
------------------------------------------
file name   : pytorch-0.4.0-py36_cuda80_cudnn7he774522_1.tar.bz2
name        : pytorch
version     : 0.4.0
build string: py36_cuda80_cudnn7he774522_1
build number: 1
size        : 529.2 MB
arch        : x86_64
constrains  : ()
platform    : Platform.win
license     : BSD 3-Clause
subdir      : win-64
url         : https://conda.anaconda.org/pytorch/win-64/pytorch-0.4.0-py36_cuda80_cudnn7he774522_1.tar.bz2
md5         : 27d20c9869fb57ffe0d6d014cf348855
dependencies:
  - cffi
  - mkl >=2018
  - numpy >=1.11
  - python >=3.6,<3.7.0a0


pytorch 0.4.0 py36_cuda90_cudnn7he774522_1
------------------------------------------
file name   : pytorch-0.4.0-py36_cuda90_cudnn7he774522_1.tar.bz2
name        : pytorch
version     : 0.4.0
build string: py36_cuda90_cudnn7he774522_1
build number: 1
size        : 577.6 MB
arch        : x86_64
constrains  : ()
platform    : Platform.win
license     : BSD 3-Clause
subdir      : win-64
url         : https://conda.anaconda.org/pytorch/win-64/pytorch-0.4.0-py36_cuda90_cudnn7he774522_1.tar.bz2
md5         : 138dcca8eeff1d58a8fd9b1febf702f6
dependencies:
  - cffi
  - mkl >=2018
  - numpy >=1.11
  - python >=3.6,<3.7.0a0


pytorch 0.4.0 py36_cuda91_cudnn7he774522_1
------------------------------------------
file name   : pytorch-0.4.0-py36_cuda91_cudnn7he774522_1.tar.bz2
name        : pytorch
version     : 0.4.0
build string: py36_cuda91_cudnn7he774522_1
build number: 1
size        : 546.4 MB
arch        : x86_64
constrains  : ()
platform    : Platform.win
license     : BSD 3-Clause
subdir      : win-64
url         : https://conda.anaconda.org/pytorch/win-64/pytorch-0.4.0-py36_cuda91_cudnn7he774522_1.tar.bz2
md5         : 326265665000de6f7501160b10b089c8
dependencies:
  - cffi
  - mkl >=2018
  - numpy >=1.11
  - python >=3.6,<3.7.0a0

---

仔细看，你可能会看到安装了蟒蛇CUDA@talonmies不，我没找到库达。我在基本环境和我使用的虚拟环境中输入了conda列表，但没有找到cudatoolkit或cudnn。在我的系统和我使用过的每个conda系统上，pytorch对CUDA和cudnn都有内置依赖关系：