cudaPackages: build redists from manifests and add CUDA 13

doc/redirects.json

@@ -58,9 +58,8 @@
   ],
   "cuda-updating-redistributables": [
     "index.html#cuda-updating-redistributables",
-    "index.html#updating-cuda-redistributables"
-  ],
-  "cuda-updating-cutensor": [
+    "index.html#updating-cuda-redistributables",
+    "index.html#updating-the-cuda-toolkit",
     "index.html#cuda-updating-cutensor",
     "index.html#updating-cutensor"
   ],
@@ -72,10 +71,6 @@
     "index.html#cuda-updating-the-cuda-package-set",
     "index.html#updating-the-cuda-package-set"
   ],
-  "cuda-updating-the-cuda-toolkit": [
-    "index.html#cuda-updating-the-cuda-toolkit",
-    "index.html#updating-the-cuda-toolkit"
-  ],
   "cuda-user-guide": [
     "index.html#cuda-user-guide"
   ],

pkgs/by-name/ac/actiona/package.nix

@@ -44,7 +44,10 @@ stdenv.mkDerivation (finalAttrs: {
   buildInputs = [
     bluez
     libnotify
-    opencv
+    # NOTE: Specifically not using lib.getOutput here because it would select the out output of opencv, which changes
+    # semantics since make-derivation uses lib.getDev on the dependency arrays, which won't touch derivations with
+    # specified outputs.
+    (opencv.cxxdev or opencv)
     qt6.qtbase
     qt6.qtmultimedia
     qt6.qttools

pkgs/by-name/bl/blender/package.nix

@@ -434,9 +434,7 @@ stdenv'.mkDerivation (finalAttrs: {
     # They comment two licenses: GPLv2 and Blender License, but they
     # say: "We've decided to cancel the BL offering for an indefinite period."
     # OptiX, enabled with cudaSupport, is non-free.
-    license =
-      with lib.licenses;
-      [ gpl2Plus ] ++ lib.optional cudaSupport (unfree // { shortName = "NVidia OptiX EULA"; });
+    license = with lib.licenses; [ gpl2Plus ] ++ lib.optional cudaSupport nvidiaCudaRedist;
 
     platforms = [
       "aarch64-linux"

pkgs/by-name/co/colmap/package.nix

@@ -55,7 +55,7 @@ let
   ]
   ++ lib.optionals cudaSupport [
     cudatoolkit
-    cudaPackages.cuda_cudart.static
+    (lib.getOutput "static" cudaPackages.cuda_cudart)
   ]
   ++ lib.optional stdenv'.cc.isClang llvmPackages.openmp;
 
@@ -129,7 +129,7 @@ stdenv'.mkDerivation {
     mainProgram = "colmap";
     homepage = "https://colmap.github.io/index.html";
     license = licenses.bsd3;
-    platforms = platforms.unix;
+    platforms = if cudaSupport then platforms.linux else platforms.unix;
     maintainers = with maintainers; [
       lebastr
       usertam

pkgs/by-name/dc/dcgm/package.nix

@@ -60,7 +60,7 @@ let
     in
     [
       (lib.cmakeFeature "CUDA${version}_INCLUDE_DIR" "${headers}")
-      (lib.cmakeFeature "CUDA${version}_LIBS" "${cudaPackages.cuda_cudart.stubs}/lib/stubs/libcuda.so")
+      (lib.cmakeFeature "CUDA${version}_LIBS" "${lib.getOutput "stubs" cudaPackages.cuda_cudart}/lib/stubs/libcuda.so")
       (lib.cmakeFeature "CUDA${version}_STATIC_LIBS" "${lib.getLib cudaPackages.cuda_cudart}/lib/libcudart.so")
       (lib.cmakeFeature "CUDA${version}_STATIC_CUBLAS_LIBS" (
         lib.concatStringsSep ";" [

pkgs/by-name/gp/gpu-burn/package.nix

@@ -6,7 +6,6 @@
   lib,
 }:
 let
-  inherit (lib.attrsets) getBin;
   inherit (lib.lists) last map optionals;
   inherit (lib.trivial) boolToString;
   inherit (config) cudaSupport;
@@ -36,7 +35,11 @@ backendStdenv.mkDerivation {
     substituteInPlace gpu_burn-drv.cpp \
       --replace-fail \
         '#define COMPARE_KERNEL "compare.ptx"' \
-        "#define COMPARE_KERNEL \"$out/share/compare.ptx\""
+        '#define COMPARE_KERNEL "${placeholder "out"}/share/compare.ptx"'
+    substituteInPlace Makefile \
+      --replace-fail \
+        '${''''${CUDAPATH}/bin/nvcc''}' \
+        '${lib.getExe cuda_nvcc}'
   '';
 
   nativeBuildInputs = [
@@ -52,7 +55,8 @@ backendStdenv.mkDerivation {
   ];
 
   makeFlags = [
-    "CUDAPATH=${getBin cuda_nvcc}"
+    # NOTE: CUDAPATH assumes cuda_cudart is a single output containing all of lib, dev, and stubs.
+    "CUDAPATH=${cuda_cudart}"
     "COMPUTE=${last (map dropDots cudaCapabilities)}"
     "IS_JETSON=${boolToString isJetsonBuild}"
   ];

pkgs/by-name/ko/koboldcpp/package.nix

@@ -57,6 +57,14 @@ effectiveStdenv.mkDerivation (finalAttrs: {
     python3Packages.wrapPython
   ];
 
+  postPatch = ''
+    nixLog "patching $PWD/Makefile to remove explicit linking against CUDA driver"
+    substituteInPlace "$PWD/Makefile" \
+      --replace-fail \
+        'CUBLASLD_FLAGS = -lcuda ' \
+        'CUBLASLD_FLAGS = '
+  '';
+
   pythonInputs = builtins.attrValues { inherit (python3Packages) tkinter customtkinter packaging; };
 
   buildInputs = [

pkgs/by-name/on/onnxruntime/package.nix

@@ -23,7 +23,7 @@
   darwinMinVersionHook,
   pythonSupport ? true,
   cudaSupport ? config.cudaSupport,
-  ncclSupport ? config.cudaSupport,
+  ncclSupport ? cudaSupport && cudaPackages.nccl.meta.available,
   withFullProtobuf ? false,
   cudaPackages ? { },
 }@inputs:
@@ -154,12 +154,7 @@ effectiveStdenv.mkDerivation rec {
       cudnn # cudnn.h
       cuda_cudart
     ]
-    ++ lib.optionals (cudaSupport && ncclSupport) (
-      with cudaPackages;
-      [
-        nccl
-      ]
-    )
+    ++ lib.optionals ncclSupport [ nccl ]
   )
   ++ lib.optionals effectiveStdenv.hostPlatform.isDarwin [
     (darwinMinVersionHook "13.3")
@@ -270,7 +265,7 @@ effectiveStdenv.mkDerivation rec {
   '';
 
   passthru = {
-    inherit cudaSupport cudaPackages; # for the python module
+    inherit cudaSupport cudaPackages ncclSupport; # for the python module
     inherit protobuf;
     tests = lib.optionalAttrs pythonSupport {
       python = python3Packages.onnxruntime;

pkgs/by-name/op/openmpi/package.nix

@@ -123,8 +123,9 @@ stdenv.mkDerivation (finalAttrs: {
     # TODO: add UCX support, which is recommended to use with cuda for the most robust OpenMPI build
     # https://github.com/openucx/ucx
     # https://www.open-mpi.org/faq/?category=buildcuda
-    (lib.withFeatureAs cudaSupport "cuda" (lib.getDev cudaPackages.cuda_cudart))
-    (lib.withFeatureAs cudaSupport "cuda-libdir" "${cudaPackages.cuda_cudart.stubs}/lib")
+    # NOTE: Open MPI requires the header files specifically, which are in the `include` output.
+    (lib.withFeatureAs cudaSupport "cuda" (lib.getOutput "include" cudaPackages.cuda_cudart))
+    (lib.withFeatureAs cudaSupport "cuda-libdir" "${lib.getLib cudaPackages.cuda_cudart}/lib")
     (lib.enableFeature cudaSupport "dlopen")
     (lib.withFeatureAs fabricSupport "psm2" (lib.getDev libpsm2))
     (lib.withFeatureAs fabricSupport "ofi" (lib.getDev libfabric))

pkgs/by-name/sl/slurm/package.nix

@@ -34,7 +34,7 @@
   # enable internal X11 support via libssh2
   enableX11 ? true,
   enableNVML ? config.cudaSupport,
-  nvml,
+  cudaPackages,
 }:
 
 stdenv.mkDerivation (finalAttrs: {
@@ -110,8 +110,8 @@ stdenv.mkDerivation (finalAttrs: {
   ++ lib.optionals enableNVML [
     (runCommand "collect-nvml" { } ''
       mkdir $out
-      ln -s ${lib.getDev nvml}/include $out/include
-      ln -s ${lib.getLib nvml}/lib/stubs $out/lib
+      ln -s ${lib.getOutput "include" cudaPackages.cuda_nvml_dev}/include $out/include
+      ln -s ${lib.getOutput "stubs" cudaPackages.cuda_nvml_dev}/lib/stubs $out/lib
     '')
   ];
 

pkgs/by-name/uc/ucc/package.nix

@@ -15,7 +15,7 @@ inputs@{
   enableSse42 ? stdenv.hostPlatform.sse4_2Support,
 }:
 let
-  inherit (lib.attrsets) getLib;
+  inherit (lib.attrsets) getOutput;
   inherit (lib.lists) optionals;
   inherit (lib.strings) concatStringsSep;
 
@@ -88,8 +88,10 @@ effectiveStdenv.mkDerivation (finalAttrs: {
   # referring to an array!
   env.LDFLAGS = toString (
     optionals enableCuda [
+      # Fake libcuda.so (the real one is deployed impurely)
+      "-L${getOutput "stubs" cuda_cudart}/lib/stubs"
       # Fake libnvidia-ml.so (the real one is deployed impurely)
-      "-L${getLib cuda_nvml_dev}/lib/stubs"
+      "-L${getOutput "stubs" cuda_nvml_dev}/lib/stubs"
     ]
   );
 

pkgs/by-name/uc/ucx/package.nix

@@ -34,6 +34,13 @@ let
   };
 in
 stdenv.mkDerivation (finalAttrs: {
+  __structuredAttrs = true;
+  # TODO(@connorbaker):
+  # When strictDeps is enabled, `cuda_nvcc` is required as the argument to `--with-cuda` in `configureFlags` or else
+  # configurePhase fails with `checking for cuda_runtime.h... no`.
+  # This is odd, especially given `cuda_runtime.h` is provided by `cuda_cudart.dev`, which is already in `buildInputs`.
+  strictDeps = true;
+
   pname = "ucx";
   version = "1.19.0";
 
@@ -75,10 +82,17 @@ stdenv.mkDerivation (finalAttrs: {
   ]
   ++ lib.optionals enableRocm rocmList;
 
-  LDFLAGS = lib.optionals enableCuda [
-    # Fake libnvidia-ml.so (the real one is deployed impurely)
-    "-L${lib.getLib cudaPackages.cuda_nvml_dev}/lib/stubs"
-  ];
+  # NOTE: With `__structuredAttrs` enabled, `LDFLAGS` must be set under `env` so it is assured to be a string;
+  # otherwise, we might have forgotten to convert it to a string and Nix would make LDFLAGS a shell variable
+  # referring to an array!
+  env.LDFLAGS = toString (
+    lib.optionals enableCuda [
+      # Fake libcuda.so (the real one is deployed impurely)
+      "-L${lib.getOutput "stubs" cudaPackages.cuda_cudart}/lib/stubs"
+      # Fake libnvidia-ml.so (the real one is deployed impurely)
+      "-L${lib.getOutput "stubs" cudaPackages.cuda_nvml_dev}/lib/stubs"
+    ]
+  );
 
   configureFlags = [
     "--with-rdmacm=${lib.getDev rdma-core}"

pkgs/by-name/xg/xgboost/package.nix

@@ -16,7 +16,7 @@
   rPackages,
 }@inputs:
 
-assert ncclSupport -> (cudaSupport && !cudaPackages.nccl.meta.unsupported);
+assert ncclSupport -> (cudaSupport && cudaPackages.nccl.meta.available);
 # Disable regular tests when building the R package
 # because 1) the R package runs its own tests and
 # 2) the R package creates a different binary shared

pkgs/development/cuda-modules/README.md

@@ -10,39 +10,48 @@ package set by [cuda-packages.nix](../../top-level/cuda-packages.nix).
 
 ## Top-level directories
 
-- `cuda`: CUDA redistributables! Provides extension to `cudaPackages` scope.
-- `cudatoolkit`: monolithic CUDA Toolkit run-file installer. Provides extension
-    to `cudaPackages` scope.
-- `cudnn`: NVIDIA cuDNN library.
-- `cutensor`: NVIDIA cuTENSOR library.
-- `fixups`: Each file or directory (excluding `default.nix`) should contain a
-    `callPackage`-able expression to be provided to the `overrideAttrs` attribute
-    of a package produced by the generic manifest builder.
-    These fixups are applied by `pname`, so packages with multiple versions
-    (e.g., `cudnn`, `cudnn_8_9`, etc.) all share a single fixup function
-    (i.e., `fixups/cudnn.nix`).
-- `generic-builders`:
-  - Contains a builder `manifest.nix` which operates on the `Manifest` type
-      defined in `modules/generic/manifests`. Most packages are built using this
-      builder.
-  - Contains a builder `multiplex.nix` which leverages the Manifest builder. In
-      short, the Multiplex builder adds multiple versions of a single package to
-      single instance of the CUDA Packages package set. It is used primarily for
-      packages like `cudnn` and `cutensor`.
-- `modules`: Nixpkgs modules to check the shape and content of CUDA
-    redistributable and feature manifests. These modules additionally use shims
-    provided by some CUDA packages to allow them to re-use the
-    `genericManifestBuilder`, even if they don't have manifest files of their
-    own. `cudnn` and `tensorrt` are examples of packages which provide such
-    shims. These modules are further described in the
-    [Modules](./modules/README.md) documentation.
+- `_cuda`: Fixed-point used to configure, construct, and extend the CUDA package
+    set. This includes NVIDIA manifests.
+- `buildRedist`: Contains the logic to build packages using NVIDIA's manifests.
 - `packages`: Contains packages which exist in every instance of the CUDA
     package set. These packages are built in a `by-name` fashion.
-- `setup-hooks`: Nixpkgs setup hooks for CUDA.
-- `tensorrt`: NVIDIA TensorRT library.
+- `tests`: Contains tests which can be run against the CUDA package set.
+
+Many redistributable packages are in the `packages` directory. Their presence
+ensures that, even if a CUDA package set which no longer includes a given package
+is being constructed, the attribute for that package will still exist (but refer
+to a broken package). This prevents missing attribute errors as the package set
+evolves.
 
 ## Distinguished packages
 
+Some packages are purposefully not in the `packages` directory. These are packages
+which do not make sense for Nixpkgs, require further investigation, or are otherwise
+not straightforward to include. These packages are:
+
+- `cuda`:
+  - `collectx_bringup`: missing `libssl.so.1.1` and `libcrypto.so.1.1`; not sure how
+    to provide them or what the package does.
+  - `cuda_sandbox_dev`: unclear on purpose.
+  - `driver_assistant`: we don't use the drivers from the CUDA releases; irrelevant.
+  - `mft_autocomplete`: unsure of purpose; contains FHS paths.
+  - `mft_oem`: unsure of purpose; contains FHS paths.
+  - `mft`: unsure of purpose; contains FHS paths.
+  - `nvidia_driver`: we don't use the drivers from the CUDA releases; irrelevant.
+  - `nvlsm`: contains FHS paths/NVSwitch and NVLINK software
+  - `libnvidia_nscq`: NVSwitch software
+  - `libnvsdm`: NVSwitch software
+- `cublasmp`:
+  - `libcublasmp`: `nvshmem` isnt' packaged.
+- `cudnn`:
+  - `cudnn_samples`: requires FreeImage, which is abandoned and not packaged.
+
+> [!NOTE]
+>
+> When packaging redistributables, prefer `autoPatchelfIgnoreMissingDeps` to providing
+> paths to stubs with `extraAutoPatchelfLibs`; the stubs are meant to be used for
+> projects where linking against libraries available only at runtime is unavoidable.
+
 ### CUDA Compatibility
 
 [CUDA Compatibility](https://docs.nvidia.com/deploy/cuda-compatibility/),