Commit 2f4ac6f8 authored by Samuel Ainsworth's avatar Samuel Ainsworth
Browse files

python3Packages.jax: add CUDA tests in passthru

Organize CUDA-enabled tests for jax and jaxlib into passthru scripts to facilitate testing.
parent 9e8cdafe
Loading
Loading
Loading
Loading
+18 −0
Original line number Diff line number Diff line
{ lib
, blas
, buildPythonPackage
, callPackage
, setuptools
, importlib-metadata
, fetchFromGitHub
@@ -126,6 +127,23 @@ buildPythonPackage rec {

  pythonImportsCheck = [ "jax" ];

  # Test CUDA-enabled jax and jaxlib. Running CUDA-enabled tests is not
  # currently feasible within the nix build environment so we have to maintain
  # this script separately. See https://github.com/NixOS/nixpkgs/pull/256230
  # for a possible remedy to this situation.
  #
  # Run these tests with eg
  #
  #   NIXPKGS_ALLOW_UNFREE=1 nixglhost -- nix run --impure .#python3Packages.jax.passthru.tests.test_cuda_jaxlibBin
  passthru.tests = {
    test_cuda_jaxlibSource = callPackage ./test-cuda.nix {
      jaxlib = jaxlib.override { cudaSupport = true; };
    };
    test_cuda_jaxlibBin = callPackage ./test-cuda.nix {
      jaxlib = jaxlib-bin.override { cudaSupport = true; };
    };
  };

  meta = with lib; {
    description = "Differentiate, compile, and transform Numpy code";
    homepage = "https://github.com/google/jax";
+17 −0
Original line number Diff line number Diff line
{ jax
, jaxlib
, pkgs
}:

pkgs.writers.writePython3Bin "jax-test-cuda" { libraries = [ jax jaxlib ]; } ''
  import jax
  from jax import random

  assert jax.devices()[0].platform == "gpu"

  rng = random.PRNGKey(0)
  x = random.normal(rng, (100, 100))
  x @ x

  print("success!")
''
+1 −10
Original line number Diff line number Diff line
@@ -2,16 +2,7 @@
# backend will require some additional work. Those wheels are located here:
# https://storage.googleapis.com/jax-releases/libtpu_releases.html.

# For future reference, the easiest way to test the GPU backend is to run
#   NIX_PATH=.. nix-shell -p python3 python3Packages.jax "python3Packages.jaxlib-bin.override { cudaSupport = true; }"
#   export XLA_FLAGS=--xla_gpu_force_compilation_parallelism=1
#   python -c "from jax.lib import xla_bridge; assert xla_bridge.get_backend().platform == 'gpu'"
#   python -c "from jax import random; random.PRNGKey(0)"
#   python -c "from jax import random; x = random.normal(random.PRNGKey(0), (100, 100)); x @ x"
# There's no convenient way to test the GPU backend in the derivation since the
# nix build environment blocks access to the GPU. See also:
#   * https://github.com/google/jax/issues/971#issuecomment-508216439
#   * https://github.com/google/jax/issues/5723#issuecomment-913038780
# See `python3Packages.jax.passthru` for CUDA tests.

{ absl-py
, autoPatchelfHook