Adding metadata

5 files changed, 148 insertions, 5 deletions

diff --git a/sci-ml/cudnn-frontend/Manifest b/sci-ml/cudnn-frontend/Manifest
index 69ed7adcd37d..11d57b939a7f 100644
--- a/sci-ml/cudnn-frontend/Manifest
+++ b/sci-ml/cudnn-frontend/Manifest
@@ -3,3 +3,4 @@ DIST cudnn-frontend-1.11.0.tar.gz 653191 BLAKE2B 2f784424bb4c850683b8ecf0b24e03c
 DIST cudnn-frontend-1.12.1.tar.gz 1683639 BLAKE2B 1d00e0caea27fc63883b867fa192f1346c67318f4781b0255e0203753a7ead80746291736e6972c58aa8a1870ccac38ea4dbfe551e5da97596a857d53bc826ea SHA512 02dba5b4480099d1e14abb2fc16d64466955988932bbfdabfe58104576582c644cacec07674baf23e51bb3d8311e5b7080b1e3b3c140d9f463767c2f117e750e
 DIST cudnn-frontend-1.15.0.tar.gz 1945364 BLAKE2B 0881cef7635d32145d396fd7476c781acf0d6913ff0704552adbb087276cc36d32d53f56a8d1975f7c8b04bf6d5df5e3bbcfbf0ed7dd8d21e6e082a5e0b1bceb SHA512 75943654652bbde8dc90cf52b58078a8c8250940827cebfe96c36750bc41047d016a94f1201f9e4069cdac56e0804c2f5a534ca86d4dc0f0ba4e65a97792a7a7
 DIST cudnn-frontend-1.16.1.tar.gz 1981715 BLAKE2B b8112bc538dba7be1596c5e13e37b031f2e158420d6c80e39e3fdb3b8b87600158eda4f30ea6a20df6a41fa04976200916810cf0c6025b2c32357a177a020a30 SHA512 f7e0d234bda6a1d284a89974ccd1e0e13e87af774372a41d4fb3af43b01a2a73eddc2deb43d2726f536238366124c64cbd48cbb98576550a0fd635815456b008
+DIST cudnn-frontend-1.28.0.tar.gz 21982353 BLAKE2B 625af8f8aebf0a663b5975d1e594deffd07dccb3aaf5cd247809f9091658936f3adf23ce2d1e40d373541c76d72cbd9e90348dea0e0b2eab83b9be2dc25211e8 SHA512 e699e1922bc96af96e734144b0284d6a7d2505e295dfce96b826d4fa1aab90a0545809f98379be114cbdfb852ced39497b01256ce6b1946b15444b49f84a4c1b
diff --git a/sci-ml/cudnn-frontend/cudnn-frontend-1.28.0.ebuild b/sci-ml/cudnn-frontend/cudnn-frontend-1.28.0.ebuild
new file mode 100644
index 000000000000..ad35836ff226
--- /dev/null
+++ b/sci-ml/cudnn-frontend/cudnn-frontend-1.28.0.ebuild
@@ -0,0 +1,107 @@
+# Copyright 1999-2026 Gentoo Authors
+# Distributed under the terms of the GNU General Public License v2
+
+EAPI=8
+
+inherit cuda cmake edo
+
+DESCRIPTION="A c++ wrapper for the cudnn backend API"
+HOMEPAGE="https://github.com/NVIDIA/cudnn-frontend"
+
+SRC_URI="https://github.com/NVIDIA/cudnn-frontend/archive/refs/tags/v${PV}.tar.gz -> ${P}.tar.gz"
+
+LICENSE="MIT"
+SLOT="0/$(ver_cut 1-2)"
+KEYWORDS="~amd64"
+IUSE="samples test"
+RESTRICT="!test? ( test )"
+
+RDEPEND="
+	>=dev-libs/cudnn-9.0.0:=
+"
+DEPEND="${RDEPEND}
+	dev-cpp/nlohmann_json
+	test? (
+		>=dev-cpp/catch-3
+		>=dev-libs/cudnn-9.15.0
+	)
+"
+
+PATCHES=(
+	"${FILESDIR}/${PN}-1.28.0-fix.patch"
+)
+
+src_prepare() {
+	cmake_src_prepare
+
+	sed -e 's#"cudnn_frontend/thirdparty/nlohmann/json.hpp"#<nlohmann/json.hpp>#' \
+		-i include/cudnn_frontend_utils.h || die
+
+	rm -r include/cudnn_frontend/thirdparty || die
+}
+
+src_configure() {
+	local narch
+	if use amd64; then
+		narch="x86_64"
+	elif use arm64; then
+		narch="sbsa"
+	fi
+
+	local mycmakeargs=(
+		-DCUDNN_FRONTEND_BUILD_PYTHON_BINDINGS="no"
+		-DCUDNN_FRONTEND_BUILD_SAMPLES="$(usex test "$(usex samples)")"
+		-DCUDNN_FRONTEND_BUILD_TESTS="$(usex test)"
+		-DCUDNN_FRONTEND_SKIP_JSON_LIB="no"
+	)
+
+	if use samples || use test; then
+		# allow slotted install
+		: "${CUDNN_PATH:=${ESYSROOT}/opt/cuda}"
+		export CUDNN_PATH
+	fi
+
+	cmake_src_configure
+}
+
+src_test() {
+	cuda_add_sandbox -w
+	addwrite "/proc/self/task"
+
+	# List all tests
+	# "${BUILD_DIR}/bin/tests" --list-tests --verbosity quiet
+
+	local catchargs=()
+
+	local CATCH_SKIP_TESTS=()
+
+	if [[ -v CUDAARCHS && "${CUDAARCHS}" != 89 ]]; then
+		CATCH_SKIP_TESTS+=(
+			 # doesn't work on 86 52
+			'Graph key'
+			'Graph key dynamic shape'
+
+			 # doesn't work on 86
+			'sdpa backward graph serialization'
+		)
+	fi
+
+	[[ -v CATCH_SKIP_TESTS ]] && catchargs+=( "${CATCH_SKIP_TESTS[@]/#/\~}" )
+
+	edo "${BUILD_DIR}/bin/tests" -s "${catchargs[@]}"
+
+	if use samples; then
+		edo "${BUILD_DIR}/bin/samples" -s
+		edo "${BUILD_DIR}/bin/legacy_samples" -s
+	fi
+
+	cmake_src_test
+}
+
+src_install() {
+	cmake_src_install
+
+	if use test; then
+		rm -R "${ED}/usr/bin/tests" || die
+	fi
+}
diff --git a/sci-ml/cudnn-frontend/files/cudnn-frontend-1.28.0-fix.patch b/sci-ml/cudnn-frontend/files/cudnn-frontend-1.28.0-fix.patch
new file mode 100644
index 000000000000..75f08557b01b
--- /dev/null
+++ b/sci-ml/cudnn-frontend/files/cudnn-frontend-1.28.0-fix.patch
@@ -0,0 +1,22 @@
+From: Paul Zander <negril.nx+gentoo@gmail.com>
+Subject: [PATCH] samples: include <memory> and <cudnn_graph.h> explicitly
+
+helpers.h uses std::unique_ptr/std::make_unique and the cudnnHandle_t API
+but relies on both arriving transitively. Include them directly; <stdexcept>
+is unused.
+
+Rebased onto 1.28.0 (upstream added an SPDX header and <vector>).
+--- a/samples/cpp/utils/helpers.h
++++ b/samples/cpp/utils/helpers.h
+@@ -6,9 +6,10 @@
+ #pragma once
+ 
+ #include <vector>
+-#include <stdexcept>
++#include <memory>
+ #include <sstream>
+ 
++#include <cudnn_graph.h>
+ #include <cuda_runtime.h>
+ #include <cuda_fp16.h>
+ 
diff --git a/sci-ml/ggml/ggml-0.23.0.ebuild b/sci-ml/ggml/ggml-0.23.0.ebuild
index f099d0b88d86..6ca110bcc73b 100644
--- a/sci-ml/ggml/ggml-0.23.0.ebuild
+++ b/sci-ml/ggml/ggml-0.23.0.ebuild
@@ -16,6 +16,9 @@ SLOT="0/${PV}"
 KEYWORDS="~amd64 ~arm64"
 
 X86_CPU_FLAGS=(
+	amx_bf16
+	amx_int8
+	amx_tile
 	avx
 	avx_vnni
 	avx2
@@ -77,6 +80,9 @@ src_configure() {
 		-DGGML_HIP_MMQ_MFMA=OFF
 
 		# CPU Flags
+		-DGGML_AMX_BF16=$(usex cpu_flags_x86_amx_bf16)
+		-DGGML_AMX_INT8=$(usex cpu_flags_x86_amx_int8)
+		-DGGML_AMX_TILE=$(usex cpu_flags_x86_amx_tile)
 		-DGGML_AVX=$(usex cpu_flags_x86_avx)
 		-DGGML_AVX_VNNI=$(usex cpu_flags_x86_avx_vnni)
 		-DGGML_AVX2=$(usex cpu_flags_x86_avx2)
diff --git a/sci-ml/torchdata/torchdata-0.11.0.ebuild b/sci-ml/torchdata/torchdata-0.11.0-r1.ebuild
index 0249e38158a3..946ea472b90b 100644
--- a/sci-ml/torchdata/torchdata-0.11.0.ebuild
+++ b/sci-ml/torchdata/torchdata-0.11.0-r1.ebuild
@@ -29,25 +29,32 @@ BDEPEND="
 	test? (
 		$(python_gen_cond_dep '
 			dev-python/expecttest[${PYTHON_USEDEP}]
+			dev-python/parameterized[${PYTHON_USEDEP}]
 		')
+		sci-ml/datasets[${PYTHON_SINGLE_USEDEP}]
 	)
 "
 
+EPYTEST_PLUGINS=()
 distutils_enable_tests pytest
 
 src_test() {
 	local EPYTEST_DESELECT=(
-		test/stateful_dataloader/test_state_dict.py::TestMultiEpochSDL_shard0::test_multi_epoch_sdl_2
-		test/stateful_dataloader/test_state_dict.py::TestMultiEpochSDL_shard0::test_multi_epoch_sdl_3
+		test/stateful_dataloader/test_dataloader.py::TestDataLoader::test_get_worker_info
+		test/stateful_dataloader/test_dataloader.py::TestDataLoaderPersistentWorkers::test_get_worker_info
+		test/stateful_dataloader/test_dataloader.py::TestIndividualWorkerQueue::test_ind_worker_queue
+		test/stateful_dataloader/test_dataloader.py::TestSetAffinity::test_set_affinity_in_worker_init
 		test/stateful_dataloader/test_state_dict.py::TestEndOfEpochBehavior_shard0::test_end_of_epoch_behavior_2
 		test/stateful_dataloader/test_state_dict.py::TestEndOfEpochBehavior_shard0::test_end_of_epoch_behavior_3
+		test/stateful_dataloader/test_state_dict.py::TestMultiEpochSDL_shard0::test_multi_epoch_sdl_2
+		test/stateful_dataloader/test_state_dict.py::TestMultiEpochSDL_shard0::test_multi_epoch_sdl_3
 		test/stateful_dataloader/test_state_dict.py::TestMultiEpochState_shard0::test_pw
-		test/stateful_dataloader/test_state_dict.py::TestSingleIterCalled_shard0::test_mp
-		test/stateful_dataloader/test_state_dict.py::TestSingleIterCalled_shard0::test_mp_iter
-		test/stateful_dataloader/test_state_dict.py::TestStateInitializationDataset::test_mp
 		test/stateful_dataloader/test_state_dict.py::TestOutOfOrderWithCheckpointing::test_out_of_order_index_ds
 		test/stateful_dataloader/test_state_dict.py::TestOutOfOrderWithCheckpointing::test_out_of_order_iterable_ds_no_completed_workers
 		test/stateful_dataloader/test_state_dict.py::TestOutOfOrderWithCheckpointing::test_out_of_order_iterable_ds_one_completed_worker
+		test/stateful_dataloader/test_state_dict.py::TestSingleIterCalled_shard0::test_mp
+		test/stateful_dataloader/test_state_dict.py::TestSingleIterCalled_shard0::test_mp_iter
+		test/stateful_dataloader/test_state_dict.py::TestStateInitializationDataset::test_mp
 	)
 	distutils-r1_src_test
 }