pyrox.dev
1{ lib
2, blas
3, buildPythonPackage
4, setuptools
5, importlib-metadata
6, fetchFromGitHub
7, jaxlib
8, jaxlib-bin
9, lapack
10, matplotlib
11, ml-dtypes
12, numpy
13, opt-einsum
14, pytestCheckHook
15, pytest-xdist
16, pythonOlder
17, scipy
18, stdenv
19}:
20
21let
22 usingMKL = blas.implementation == "mkl" || lapack.implementation == "mkl";
23 # jaxlib is broken on aarch64-* as of 2023-03-05, but the binary wheels work
24 # fine. jaxlib is only used in the checkPhase, so switching backends does not
25 # impact package behavior. Get rid of this once jaxlib is fixed on aarch64-*.
26 jaxlib' = if jaxlib.meta.broken then jaxlib-bin else jaxlib;
27in
28buildPythonPackage rec {
29 pname = "jax";
30 version = "0.4.20";
31 pyproject = true;
32
33 disabled = pythonOlder "3.9";
34
35 src = fetchFromGitHub {
36 owner = "google";
37 repo = "jax";
38 # google/jax contains tags for jax and jaxlib. Only use jax tags!
39 rev = "refs/tags/${pname}-v${version}";
40 hash = "sha256-WLYXUtchOaA6SGnKuVhN9CmV06xMCLQTEuEtL13ttZU=";
41 };
42
43 nativeBuildInputs = [
44 setuptools
45 ];
46
47 # The version is automatically set to ".dev" if this variable is not set.
48 # https://github.com/google/jax/commit/e01f2617b85c5bdffc5ffb60b3d8d8ca9519a1f3
49 JAX_RELEASE = "1";
50
51 # jaxlib is _not_ included in propagatedBuildInputs because there are
52 # different versions of jaxlib depending on the desired target hardware. The
53 # JAX project ships separate wheels for CPU, GPU, and TPU.
54 propagatedBuildInputs = [
55 ml-dtypes
56 numpy
57 opt-einsum
58 scipy
59 ] ++ lib.optional (pythonOlder "3.10") importlib-metadata;
60
61 nativeCheckInputs = [
62 jaxlib'
63 matplotlib
64 pytestCheckHook
65 pytest-xdist
66 ];
67
68 # high parallelism will result in the tests getting stuck
69 dontUsePytestXdist = true;
70
71 # NOTE: Don't run the tests in the expiremental directory as they require flax
72 # which creates a circular dependency. See https://discourse.nixos.org/t/how-to-nix-ify-python-packages-with-circular-dependencies/14648/2.
73 # Not a big deal, this is how the JAX docs suggest running the test suite
74 # anyhow.
75 pytestFlagsArray = [
76 "--numprocesses=4"
77 "-W ignore::DeprecationWarning"
78 "tests/"
79 ];
80
81 disabledTests = [
82 # Exceeds tolerance when the machine is busy
83 "test_custom_linear_solve_aux"
84 # UserWarning: Explicitly requested dtype <class 'numpy.float64'>
85 # requested in astype is not available, and will be truncated to
86 # dtype float32. (With numpy 1.24)
87 "testKde3"
88 "testKde5"
89 "testKde6"
90 ] ++ lib.optionals usingMKL [
91 # See
92 # * https://github.com/google/jax/issues/9705
93 # * https://discourse.nixos.org/t/getting-different-results-for-the-same-build-on-two-equally-configured-machines/17921
94 # * https://github.com/NixOS/nixpkgs/issues/161960
95 "test_custom_linear_solve_cholesky"
96 "test_custom_root_with_aux"
97 "testEigvalsGrad_shape"
98 ] ++ lib.optionals stdenv.isAarch64 [
99 # See https://github.com/google/jax/issues/14793.
100 "test_for_loop_fixpoint_correctly_identifies_loop_varying_residuals_unrolled_for_loop"
101 "testQdwhWithRandomMatrix3"
102 "testScanGrad_jit_scan"
103
104 # See https://github.com/google/jax/issues/17867.
105 "test_array"
106 "test_async"
107 "test_copy0"
108 "test_device_put"
109 "test_make_array_from_callback"
110 "test_make_array_from_single_device_arrays"
111
112 # Fails on some hardware due to some numerical error
113 # See https://github.com/google/jax/issues/18535
114 "testQdwhWithOnRankDeficientInput5"
115 ];
116
117 disabledTestPaths = lib.optionals (stdenv.isDarwin && stdenv.isAarch64) [
118 # RuntimeWarning: invalid value encountered in cast
119 "tests/lax_test.py"
120 ];
121
122 pythonImportsCheck = [ "jax" ];
123
124 meta = with lib; {
125 description = "Differentiate, compile, and transform Numpy code";
126 homepage = "https://github.com/google/jax";
127 license = licenses.asl20;
128 maintainers = with maintainers; [ samuela ];
129 };
130}