1

When installing pytorch in conda, cudatoolkit is also installed. My understanding is that the pytorch code is pre-compiled into machine code. If this is true, is the cudatoolkit used when writing/running regular pytorch programs or is it just there in the case you want to write custom CUDA code?

talonmies
  • 70,661
  • 34
  • 192
  • 269
curiousgeorge
  • 593
  • 1
  • 5
  • 14
  • 1
    the thing that conda installs when it installs the cudatoolkit is not actually the full cuda toolkit. It is bits and pieces (such as libraries) that are still required even if you have compiled code. A compiled code can still depend on external libraries, for example. The cudatoolkit installed by conda for this purpose is not sufficient for writing your own custom CUDA code, in my experience. For example, as far as I know, it does not install the `nvcc` compiler-driver. The "cudatoolkit" thing that conda installs as a dependency for the GPU-enabled version of pytorch is definitely necessary. – Robert Crovella Aug 02 '20 at 04:18
  • 1
    https://stackoverflow.com/a/61538568/681865 – talonmies Aug 02 '20 at 06:18

0 Answers0