[Backport release-25.11] cudaPackages.tensorrt: 10.9.0 -> 10.14.1 (for... (953e37fd) · Commits · nix / nixpkgs

pkgs/by-name/ka/katago/package.nix

+11 −0

Original line number	Diff line number	Diff line
		@@ -61,6 +61,17 @@ stdenv'.mkDerivation rec {
		makeWrapper
		];

		# Included from release 1.16.2:
		# https://github.com/lightvector/KataGo/commit/9030f72d152da42c1dd03590aa5116993ea842f6
		# Doesn't apply cleanly as a patch so doing a quick replacement to the same effect.
		prePatch = lib.optionalString (backend == "tensorrt") ''
		nixLog "patching $PWD/cpp/CMakeLists.txt to work around outdated TensorRT version detection"
		substituteInPlace "$PWD/cpp/CMakeLists.txt" \
		--replace-fail \
		'if(TENSORRT_VERSION VERSION_LESS 8.5)' \
		'if(NOT TENSORRT_VERSION STREQUAL ".." AND TENSORRT_VERSION VERSION_LESS 8.5)'
		'';

		buildInputs = [
		libzip
		boost

pkgs/development/cuda-modules/_cuda/manifests/tensorrt/redistrib_10.14.1.json

0 → 100644

+36 −0

Original line number	Diff line number	Diff line
		{
		"release_date": "2025-11-04",
		"release_label": "10.14.1",
		"release_product": "tensorrt",
		"tensorrt": {
		"name": "NVIDIA TensorRT",
		"license": "TensorRT",
		"version": "10.14.1.48",
		"cuda_variant": [
		"12",
		"13"
		],
		"linux-sbsa": {
		"cuda13": {
		"md5": "f31d534daa41f517c8c5db8e0478119f",
		"relative_path": "tensorrt/10.14.1/tars/TensorRT-10.14.1.48.Linux.aarch64-gnu.cuda-13.0.tar.gz",
		"sha256": "3d3e438b49196a9b9ed5c5bc37ec24be9aa74d67d41db0f8fa54b4894aad748a",
		"size": "5554126761"
		}
		},
		"linux-x86_64": {
		"cuda12": {
		"md5": "d21dd1cb62518f81173150f7bda77311",
		"relative_path": "tensorrt/10.14.1/tars/TensorRT-10.14.1.48.Linux.x86_64-gnu.cuda-12.9.tar.gz",
		"sha256": "0daa7d5929c78edfbe86b474064d0f82d2064c475cc6be747c5101f1ccc37105",
		"size": "7905642926"
		},
		"cuda13": {
		"md5": "8ef547b7811197572294494551478214",
		"relative_path": "tensorrt/10.14.1/tars/TensorRT-10.14.1.48.Linux.x86_64-gnu.cuda-13.0.tar.gz",
		"sha256": "c74af67db57f1a0d7e66bb01ab93f1ecda5facac491ca76e680d832f1e035ce6",
		"size": "7004299092"
		}
		}
		}
		}

pkgs/development/cuda-modules/_cuda/manifests/tensorrt/redistrib_10.7.0.json

+7 −0

Original line number	Diff line number	Diff line
		@@ -7,6 +7,7 @@
		"license": "TensorRT",
		"version": "10.7.0.23",
		"cuda_variant": [
		"11",
		"12"
		],
		"linux-aarch64": {
		@@ -26,6 +27,12 @@
		}
		},
		"linux-x86_64": {
		"cuda11": {
		"md5": "6adf7d02aae9105abb6bf3b4c1ce89f4",
		"relative_path": "tensorrt/10.7.0/tars/TensorRT-10.7.0.23.Linux.x86_64-gnu.cuda-11.8.tar.gz",
		"sha256": "958e1c32b48e41d1c48bdc94363450e14f996ca9de0e205ccee65af319eea2c0",
		"size": "4341092411"
		},
		"cuda12": {
		"md5": "925c98fbe9abe82058814159727732a2",
		"relative_path": "tensorrt/10.7.0/tars/TensorRT-10.7.0.23.Linux.x86_64-gnu.cuda-12.6.tar.gz",

pkgs/development/cuda-modules/_cuda/manifests/tensorrt/redistrib_10.9.0.json

deleted100644 → 0

+0 −29

Original line number	Diff line number	Diff line
		{
		"release_date": "2025-03-11",
		"release_label": "10.9.0",
		"release_product": "tensorrt",
		"tensorrt": {
		"name": "NVIDIA TensorRT",
		"license": "TensorRT",
		"version": "10.9.0.34",
		"cuda_variant": [
		"12"
		],
		"linux-sbsa": {
		"cuda12": {
		"md5": "e56a9f9d7327c65d9b95996d3008ed44",
		"relative_path": "tensorrt/10.9.0/tars/TensorRT-10.9.0.34.Linux.aarch64-gnu.cuda-12.8.tar.gz",
		"sha256": "b81ec2a067f67f082c13caec2dcef54385b42a9de6a4ecae6f318aa2d41964f2",
		"size": "4123115318"
		}
		},
		"linux-x86_64": {
		"cuda12": {
		"md5": "ee49e3e6e00b21274907956216b6769f",
		"relative_path": "tensorrt/10.9.0/tars/TensorRT-10.9.0.34.Linux.x86_64-gnu.cuda-12.8.tar.gz",
		"sha256": "33be0e61e3bf177bbbcabb4892bf013f0c8ac71d2be73f2803848a382cb14272",
		"size": "6917032417"
		}
		}
		}
		}

pkgs/development/cuda-modules/packages/tensorrt.nix

+73 −28

Original line number	Diff line number	Diff line
		@@ -3,24 +3,33 @@
		backendStdenv,
		buildRedist,
		cuda_cudart,
		cudnn,
		cuda_nvrtc,
		cudaAtLeast,
		cudaMajorMinorVersion,
		lib,
		libcudla, # only for Jetson
		patchelf,
		}:
		let
		inherit (_cuda.lib) majorMinorPatch;
		inherit (backendStdenv) hasJetsonCudaCapability;
		inherit (lib.attrsets) getLib;
		inherit (backendStdenv) cudaCapabilities hostRedistSystem;
		inherit (lib.lists) optionals;
		inherit (lib.strings) concatStringsSep;
		in
		buildRedist (
		finalAttrs:
		let
		majorMinorPatchVersion = majorMinorPatch finalAttrs.version;
		majorVersion = lib.versions.major finalAttrs.version;

		tensorrtAtLeast = lib.versionAtLeast finalAttrs.version;
		tensorrtOlder = lib.versionOlder finalAttrs.version;

		# Create variables and use logical OR to allow short-circuiting.
		tensorrtAtLeast105 = tensorrtAtLeast "10.5.0";
		tensorrtAtLeast100 = tensorrtAtLeast105 \|\| tensorrtAtLeast "10.0.0";

		allCCNewerThan75 = lib.all (lib.flip lib.versionAtLeast "7.5") cudaCapabilities;
		allCCNewerThan70 = allCCNewerThan75 \|\| lib.all (lib.flip lib.versionAtLeast "7.0") cudaCapabilities;

		cudaCapabilitiesJSON = builtins.toJSON cudaCapabilities;
		in
		{
		redistName = "tensorrt";
		@@ -32,7 +41,12 @@ buildRedist (
		"dev"
		"include"
		"lib"
		]
		# From 10.14.1, TensorRT samples are distributed through the TensorRT GitHub repository.
		++ optionals (tensorrtOlder "10.14.1") [
		"samples"
		]
		++ [
		"static"
		# "stubs" removed in postInstall
		];
		@@ -42,8 +56,6 @@ buildRedist (
		nativeBuildInputs = [ patchelf ];

		buildInputs = [
		(getLib cudnn)
		(getLib cuda_nvrtc)
		cuda_cudart
		]
		++ optionals libcudla.meta.available [ libcudla ];
		@@ -62,38 +74,44 @@ buildRedist (
		''
		for dir in bin lib; do
		[[ -L "$dir" ]] \|\| continue
		nixLog "replacing symlink $NIX_BUILD_TOP/$sourceRoot/$dir with $NIX_BUILD_TOP/$sourceRoot/targets/${targetString}/$dir"
		rm --verbose "$NIX_BUILD_TOP/$sourceRoot/$dir"
		mv --verbose --no-clobber "$NIX_BUILD_TOP/$sourceRoot/targets/${targetString}/$dir" "$NIX_BUILD_TOP/$sourceRoot/$dir"
		nixLog "replacing symlink $PWD/$dir with $PWD/targets/${targetString}/$dir"
		rm --verbose "$PWD/$dir"
		mv --verbose --no-clobber "$PWD/targets/${targetString}/$dir" "$PWD/$dir"
		done
		unset -v dir
		''
		# Remove symlinks if they exist
		+ ''
		for dir in include samples; do
		if [[ -L "$NIX_BUILD_TOP/$sourceRoot/targets/${targetString}/$dir" ]]; then
		nixLog "removing symlink $NIX_BUILD_TOP/$sourceRoot/targets/${targetString}/$dir"
		rm --verbose "$NIX_BUILD_TOP/$sourceRoot/targets/${targetString}/$dir"
		if [[ -L "$PWD/targets/${targetString}/$dir" ]]; then
		nixLog "removing symlink $PWD/targets/${targetString}/$dir"
		rm --verbose "$PWD/targets/${targetString}/$dir"
		fi
		done
		unset -v dir

		if [[ -d "$NIX_BUILD_TOP/$sourceRoot/targets" ]]; then
		if [[ -d "$PWD/targets" ]]; then
		nixLog "removing targets directory"
		rm --recursive --verbose "$NIX_BUILD_TOP/$sourceRoot/targets" \|\| {
		nixErrorLog "could not delete $NIX_BUILD_TOP/$sourceRoot/targets: $(ls -laR "$NIX_BUILD_TOP/$sourceRoot/targets")"
		rm --recursive --verbose "$PWD/targets" \|\| {
		nixErrorLog "could not delete $PWD/targets: $(ls -laR "$PWD/targets")"
		exit 1
		}
		fi
		'';

		autoPatchelfIgnoreMissingDeps = optionals hasJetsonCudaCapability [
		autoPatchelfIgnoreMissingDeps =
		optionals (hostRedistSystem == "linux-aarch64") [
		"libnvdla_compiler.so"
		]
		++ optionals (tensorrtAtLeast "10.13.3") [
		"libcuda.so.1"
		];

		# Create a symlink for the Onnx header files in include/onnx
		# NOTE(@connorbaker): This is shared with the tensorrt-oss package, with the `out` output swapped with `include`.
		# When updating one, check if the other should be updated.
		# TODO(@connorbaker): It seems like recent versions of TensorRT have separate libs for separate capabilities;
		# we should remove libraries older than those necessary to support requested capabilities.
		postInstall = ''
		mkdir "''${!outputInclude:?}/include/onnx"
		pushd "''${!outputInclude:?}/include" >/dev/null
		@@ -102,6 +120,7 @@ buildRedist (
		popd >/dev/null
		''
		# Move the python directory, which contains header files, to the include output.
		# NOTE: Python wheels should be built from source using the TensorRT GitHub repo.
		+ ''
		nixLog "moving python directory to include output"
		moveToOutput python "''${!outputInclude:?}"
		@@ -115,10 +134,11 @@ buildRedist (
		''
		# Remove the Windows library used for cross-compilation if it exists.
		+ ''
		if [[ -e "''${!outputLib:?}/lib/libnvinfer_builder_resource_win.so.${majorMinorPatchVersion}" ]]; then
		nixLog "removing Windows library"
		rm --verbose "''${!outputLib:?}/lib/libnvinfer_builder_resource_win.so.${majorMinorPatchVersion}"
		fi
		nixLog "removing any Windows libraries"
		for winLib in "''${!outputLib:?}/lib/"_win; do
		rm --verbose "$winLib"
		done
		unset -v winLib
		''
		# Remove the stub libraries.
		+ ''
		@@ -137,10 +157,35 @@ buildRedist (
		--add-needed libnvinfer_plugin.so.${majorVersion}
		'';

		passthru = {
		# The CUDNN used with TensorRT.
		inherit cudnn;
		};
		# NOTE: Like cuDNN, NVIDIA offers forward compatibility within a major releases of CUDA.
		platformAssertions = [
		{
		message =
		"tensorrt releases since 10.0.0 (found ${finalAttrs.version})"
		+ " support CUDA compute capabilities 7.0 and newer (found ${cudaCapabilitiesJSON})";
		assertion = tensorrtAtLeast100 -> allCCNewerThan70;
		}
		{
		message =
		"tensorrt releases since 10.0.0 (found ${finalAttrs.version})"
		+ " support only CUDA compute capability 8.7 (Jetson Orin) for pre-Thor Jetson devices"
		+ " (found ${cudaCapabilitiesJSON})";
		assertion =
		tensorrtAtLeast100 && hostRedistSystem == "linux-aarch64" -> cudaCapabilities == [ "8.7" ];
		}
		{
		message =
		"tensorrt releases since 10.0.0 (found ${finalAttrs.version})"
		+ " support CUDA 12.4 and newer for pre-Thor Jetson devices (found ${cudaMajorMinorVersion})";
		assertion = tensorrtAtLeast100 && hostRedistSystem == "linux-aarch64" -> cudaAtLeast "12.4";
		}
		{
		message =
		"tensorrt releases since 10.5.0 (found ${finalAttrs.version})"
		+ " support CUDA compute capabilities 7.5 and newer (found ${cudaCapabilitiesJSON})";
		assertion = tensorrtAtLeast105 -> allCCNewerThan75;
		}
		];

		meta = {
		description = "SDK that facilitates high-performance machine learning inference";