python3Packages.jax: add missing cuda libraries (#375186)

pkgs/development/python-modules/jax-cuda12-pjrt/default.nix

···
   cudaLibPath = lib.makeLibraryPath (
     with cudaPackages;
     [
+      (lib.getLib libcublas) # libcublas.so
+      (lib.getLib cuda_cupti) # libcupti.so
       (lib.getLib cuda_cudart) # libcudart.so
       (lib.getLib cudnn) # libcudnn.so
-      (lib.getLib libcublas) # libcublas.so
-      addDriverRunpath.driverLink # libcuda.so
+      (lib.getLib libcufft) # libcufft.so
+      (lib.getLib libcusolver) # libcusolver.so
+      (lib.getLib libcusparse) # libcusparse.so
+      (lib.getLib nccl) # libnccl.so
+      (lib.getLib libnvjitlink) # libnvJitLink.so
+      (lib.getLib addDriverRunpath.driverLink) # libcuda.so
     ]
   );
 
···
   doCheck = true;
 
   pythonImportsCheck = [ "jax_plugins" ];
+
+  inherit cudaLibPath;
 
   meta = {
     description = "JAX XLA PJRT Plugin for NVIDIA GPUs";

pkgs/development/python-modules/jax-cuda12-plugin/default.nix

···
   jax-cuda12-pjrt,
 }:
 let
+  inherit (jaxlib) version;
   inherit (cudaPackages) cudaVersion;
-  inherit (jaxlib) version;
+  inherit (jax-cuda12-pjrt) cudaLibPath;
 
   getSrcFromPypi =
     {
···
     wheelUnpackHook
   ];
 
+  # jax-cuda12-plugin looks for ptxas at runtime, e.g. with a triton kernel.
+  # Linking into $out is the least bad solution. See
+  # * https://github.com/NixOS/nixpkgs/pull/164176#discussion_r828801621
+  # * https://github.com/NixOS/nixpkgs/pull/288829#discussion_r1493852211
+  # * https://github.com/NixOS/nixpkgs/pull/375186
+  # for more info.
+  postInstall = ''
+    mkdir -p $out/${python.sitePackages}/jax_cuda12_plugin/cuda/bin
+    ln -s ${lib.getExe' cudaPackages.cuda_nvcc "ptxas"} $out/${python.sitePackages}/jax_cuda12_plugin/cuda/bin
+    ln -s ${lib.getExe' cudaPackages.cuda_nvcc "nvlink"} $out/${python.sitePackages}/jax_cuda12_plugin/cuda/bin
+  '';
+
+  # jax-cuda12-plugin contains shared libraries that open other shared libraries via dlopen
+  # and these implicit dependencies are not recognized by ldd or
+  # autoPatchelfHook. That means we need to sneak them into rpath. This step
+  # must be done after autoPatchelfHook and the automatic stripping of
+  # artifacts. autoPatchelfHook runs in postFixup and auto-stripping runs in the
+  # patchPhase.
+  preInstallCheck = ''
+    patchelf --add-rpath "${cudaLibPath}" $out/${python.sitePackages}/jax_cuda12_plugin/*.so
+  '';
+
   dependencies = [ jax-cuda12-pjrt ];
 
   pythonImportsCheck = [ "jax_cuda12_plugin" ];
 
-  # no tests
-  doCheck = false;
+  # FIXME: there are no tests, but we need to run preInstallCheck above
+  doCheck = true;
 
   meta = {
     description = "JAX Plugin for CUDA12";

pkgs/development/python-modules/jax/default.nix

···
 
   meta = {
     description = "Source-built JAX frontend: differentiate, compile, and transform Numpy code";
-    longDescription = ''
-      This is the JAX frontend package, it's meant to be used together with one of the jaxlib implementations,
-      e.g. `python3Packages.jaxlib`, `python3Packages.jaxlib-bin`, or `python3Packages.jaxlibWithCuda`.
-    '';
     homepage = "https://github.com/google/jax";
     license = lib.licenses.asl20;
     maintainers = with lib.maintainers; [ samuela ];

pkgs/development/python-modules/jax/test-cuda.nix

···
   }
   ''
     import jax
+    import jax.numpy as jnp
     from jax import random
+    from jax.experimental import sparse
 
-    assert jax.devices()[0].platform == "gpu"
+    assert jax.devices()[0].platform == "gpu"  # libcuda.so
 
-    rng = random.PRNGKey(0)
+    rng = random.key(0)  # libcudart.so, libcudnn.so
     x = random.normal(rng, (100, 100))
-    x @ x
+    x @ x  # libcublas.so
+    jnp.fft.fft(x)  # libcufft.so
+    jnp.linalg.inv(x)  # libcusolver.so
+    sparse.CSR.fromdense(x) @ x  # libcusparse.so
 
     print("success!")
   ''