mirror of
https://github.com/NixOS/nixpkgs.git
synced 2025-01-15 09:23:37 +00:00
78 lines
3.3 KiB
Nix
78 lines
3.3 KiB
Nix
# The following version combinations are supported:
|
|
# * cuDNN 7.4.2, cudatoolkit 10.0
|
|
# * cuDNN 7.6.5, cudatoolkit 10.2
|
|
# * cuDNN 8.1.1, cudatoolkit 11.0-11.2
|
|
# * cuDNN 8.3.0, cudatoolkit 11.0-11.5
|
|
{ callPackage
|
|
, cudatoolkit_10_0
|
|
, cudatoolkit_10_2
|
|
, cudatoolkit_11_0
|
|
, cudatoolkit_11_1
|
|
, cudatoolkit_11_2
|
|
, cudatoolkit_11_3
|
|
, cudatoolkit_11_4
|
|
, cudatoolkit_11_5
|
|
}:
|
|
|
|
let
|
|
generic = args: callPackage (import ./generic.nix (removeAttrs args [ "cudatoolkit" ])) {
|
|
inherit (args) cudatoolkit;
|
|
};
|
|
in
|
|
rec {
|
|
# cuDNN 7.x
|
|
# Still used by libtensorflow-bin. It should be upgraded at some point.
|
|
cudnn_7_4_cudatoolkit_10_0 = generic rec {
|
|
version = "7.4.2";
|
|
cudatoolkit = cudatoolkit_10_0;
|
|
srcName = "cudnn-${cudatoolkit.majorVersion}-linux-x64-v7.4.2.24.tgz";
|
|
sha256 = "18ys0apiz9afid2s6lvy9qbyi8g66aimb2a7ikl1f3dm09mciprf";
|
|
};
|
|
|
|
# The `cudnn` alias still points to this in all-packages.nix. It should be
|
|
# upgraded at some point.
|
|
cudnn_7_6_cudatoolkit_10_2 = generic rec {
|
|
version = "7.6.5";
|
|
cudatoolkit = cudatoolkit_10_2;
|
|
srcName = "cudnn-${cudatoolkit.majorVersion}-linux-x64-v7.6.5.32.tgz";
|
|
sha256 = "084c13vzjdkb5s1996yilybg6dgav1lscjr1xdcgvlmfrbr6f0k0";
|
|
};
|
|
|
|
cudnn_7_6_cudatoolkit_10 = cudnn_7_6_cudatoolkit_10_2;
|
|
|
|
# cuDNN 8.x
|
|
# cuDNN 8.1 is still used by tensorflow at the time of writing (2022-02-17).
|
|
# See https://github.com/NixOS/nixpkgs/pull/158218 for more info.
|
|
cudnn_8_1_cudatoolkit_11_0 = generic rec {
|
|
version = "8.1.1";
|
|
cudatoolkit = cudatoolkit_11_0;
|
|
# 8.1.0 is compatible with CUDA 11.0-11.2:
|
|
# https://docs.nvidia.com/deeplearning/cudnn/archives/cudnn-811/support-matrix/index.html
|
|
srcName = "cudnn-11.2-linux-x64-v8.1.1.33.tgz";
|
|
hash = "sha256-mKh4TpKGLyABjSDCgbMNSgzZUfk2lPZDPM9K6cUCumo=";
|
|
};
|
|
cudnn_8_1_cudatoolkit_11_1 = cudnn_8_1_cudatoolkit_11_0.override { cudatoolkit = cudatoolkit_11_1; };
|
|
cudnn_8_1_cudatoolkit_11_2 = cudnn_8_1_cudatoolkit_11_0.override { cudatoolkit = cudatoolkit_11_2; };
|
|
cudnn_8_1_cudatoolkit_11 = cudnn_8_1_cudatoolkit_11_2;
|
|
|
|
# cuDNN 8.3 is necessary for the latest jaxlib, esp. jaxlib-bin. See
|
|
# https://github.com/google/jax/discussions/9455 for more info.
|
|
cudnn_8_3_cudatoolkit_11_0 = generic rec {
|
|
# 8.3.0 is the last version to respect the folder structure that generic.nix
|
|
# expects. Later versions have files in a subdirectory `local_installers`.
|
|
# See eg https://developer.download.nvidia.com/compute/redist/cudnn/v8.3.1/.
|
|
version = "8.3.0";
|
|
cudatoolkit = cudatoolkit_11_0;
|
|
# 8.3.0 is compatible with CUDA 11.0-11.5:
|
|
# https://docs.nvidia.com/deeplearning/cudnn/archives/cudnn-830/support-matrix/index.html
|
|
srcName = "cudnn-11.5-linux-x64-v8.3.0.98.tgz";
|
|
hash = "sha256-RMb1rVyxL7dPoMmh58qvTwTXVa3xGi5bbJ5BfaN2srI=";
|
|
};
|
|
cudnn_8_3_cudatoolkit_11_1 = cudnn_8_3_cudatoolkit_11_0.override { cudatoolkit = cudatoolkit_11_1; };
|
|
cudnn_8_3_cudatoolkit_11_2 = cudnn_8_3_cudatoolkit_11_0.override { cudatoolkit = cudatoolkit_11_2; };
|
|
cudnn_8_3_cudatoolkit_11_3 = cudnn_8_3_cudatoolkit_11_0.override { cudatoolkit = cudatoolkit_11_3; };
|
|
cudnn_8_3_cudatoolkit_11_4 = cudnn_8_3_cudatoolkit_11_0.override { cudatoolkit = cudatoolkit_11_4; };
|
|
cudnn_8_3_cudatoolkit_11_5 = cudnn_8_3_cudatoolkit_11_0.override { cudatoolkit = cudatoolkit_11_5; };
|
|
cudnn_8_3_cudatoolkit_11 = cudnn_8_3_cudatoolkit_11_5;
|
|
}
|