v4.5.0

[4.5.0] - 2026-07-22

Changed

• Updated Holoscan SDK and Holohub apps to the 4.5.0 source candidates.
• Updated GXF to 5.7.1-20260713-86d402891 to satisfy the Holoscan SDK 4.5.0 find_package(GXF 5.7.1 CONFIG REQUIRED) constraint.
• Refreshed the Holoscan SDK and Holohub apps OE support patches on their exact 4.5.0 candidate sources without changing the reviewed application selection.

Unchanged

• Retained the 4.4.0 recipe versions for PyTorch, torchvision, CCCL, RMM, ONNX Runtime, UCX, UCXX, MatX, fmt, spdlog, CPM CMake, and RAPIDS CMake because Holoscan SDK 4.5.0 does not change those dependency versions.
• Holoscan Sensor Bridge is not built by this layer because HOLOSCAN_BUILD_HOLOLINK remains disabled in the Holoscan SDK recipe.

v4.4.0

[4.4.0] - 2026-07-07

Changed

• Updated Holoscan SDK and Holohub apps to 4.4.0.
• Refreshed the Holoscan SDK external-dependency patch for the RAPIDS CMake bootstrap update in 4.4.0 while continuing to use the layer-provided RAPIDS CMake package.
• Refreshed the Holohub apps patches for the Python module namespace and install-layout updates in the 4.4.0 source.

Fixed

• Updated Pseudo from 1.9.0 to 1.9.8 after SQA reproduced the unknown directory, fd 4 packaging failure tracked as Yocto #16316. The update incorporates upstream openat and EFAULT path-handling fixes.

Unchanged

• Retained GXF 5.7.0-20260515-6a50c8f1a. Holoscan SDK 4.4.0 uses the same GXF artifact as 4.3.0.
• Retained the 4.3.0 recipe versions for PyTorch, torchvision, CCCL, RMM, ONNX Runtime, UCX, UCXX, MatX, fmt, and spdlog because Holoscan SDK 4.4.0 does not change those dependency versions.
• Holoscan Sensor Bridge 2.6.0 is not built by this layer because HOLOSCAN_BUILD_HOLOLINK remains disabled in the Holoscan SDK recipe.

v4.3.0

[4.3.0] - 2026-05-25

Changed

• Updated Holoscan SDK and Holohub apps to 4.3.0.
• Updated GXF to 5.7.0-20260515-6a50c8f1a (HSDK 4.3.0 raises find_package(GXF 5.7 CONFIG REQUIRED)).
• Updated fmt to 12.1.0 (required by HSDK 4.3.0; PREFERRED_VERSION_fmt in conf/holoscan-common.conf updated to match).
• Updated spdlog to 1.17.0 (required by HSDK 4.3.0; PREFERRED_VERSION_spdlog updated to match). Dropped CVE-2025-6140.patch — the upstream commit 10320184df1eb4 (size-underflow guard on truncate in pattern_formatter-inl.h) is integrated in 1.17.0.
• Updated cccl to 3.2.0 (HSDK 4.3.0 requests CCCL_REQUESTED_VERSION 3.2.0; also satisfies RMM 26.02.00's CCCL 3.1+ floor — RMM PR 2180). Retires the 4.2.0 cycle's CCCL-version-strip hunk inside holoscan-sdk/0002-Use-external-dependencies-deps.patch.
• Updated rapids-cmake to 26.02.00 (required by HSDK 4.3.0 PyTorch chain; kept in lock-step with RMM 26.02.00).
• Updated rmm to 26.02.00 (HSDK 4.3.0 requests RMM_TAG "26.02.00"). Removed the do_install:append GXF 5.6.0 compat-forwarder shim (rmm/mr/cuda_memory_resource.hpp, rmm/mr/pool_memory_resource.hpp) introduced in 4.2.0 — RMM 26.02.00 restores both headers at the short rmm/mr/ path (RMM PR 2151), and GXF 5.7.0 in 4.3.0 picks them up natively.
• Updated pytorch to 2.11.0. HSDK 4.3.0 pins PYTORCH_IGPU_VERSION=2.11.0 / PYTORCH_DGPU_VERSION=2.11.0 in the Dockerfile, and the 2.5.0 → 2.11.0 jump is also required to compile against the new CCCL 3.2.0 / RMM 26.02.00 (PyTorch 2.5.0's anonymous-namespace at::native::offset_t in aten/src/ATen/native/cuda/SortStable.cu fails the random-access-iterator concept that cub::DispatchSegmentedRadixSort::SortPairs now requires). Replaced the 2.5.0 patch set with the 2.11.0-compatible set, and added -DUSE_SYSTEM_ONNX=ON, -DUSE_SYSTEM_PYBIND11=ON, -DFMT_INSTALL=OFF to EXTRA_OECMAKE. The 2.5.0 recipe and 4 retired 2.5-only patches (0001-Remove-Modules_CUDA_fix.patch, 0003-Fix-RPATH.patch, 0005-Disable-various-warnings.patch, 0006-Fixups-for-cross-building-in-OE-with-CCCL-3.0.patch) were removed. Credit to Ilies CHERGUI for the corresponding pytorch recipe and patch set in meta-tegra-community that this work is derived from.
• Updated torchvision to 0.26.0 (paired with PyTorch 2.11.x). Set COMPATIBLE_MACHINE = "(cuda)", added -DCMAKE_CUDA_IMPLICIT_INCLUDE_DIRECTORIES=${RECIPE_SYSROOT}/usr/include and libcublas to DEPENDS. Pass TORCH_CUDA_ARCH_LIST via EXTRA_OECMAKE (derived from CUDA_ARCHITECTURES) to bypass the select_compute_arch.cmake try_run() probe that is unsupported in OE cross-compile sysroots. Credit to Ilies CHERGUI for the corresponding torchvision_0.26.0.bb recipe in meta-tegra-community that this work mirrors.
• holohub-apps: bumped find_package(rmm 25.10.0 REQUIRED) pin to rmm 26.02.0 across 16 CMakeLists.txt files (applications/, operators/, gxf_extensions/lstm_tensor_rt_inference/) in 0009-Updates-for-OE-cross-builds.patch.

Fixed

• Added MIRRORS:append rule to conf/holoscan-common.conf redirecting https://crates.io/api/v1/crates/<name>/<version>/download to https://static.crates.io/crates/<name>/<name>-<version>.crate. The crates.io API endpoint started returning HTTP 403 to default-UA clients (bitbake's wget, curl, etc.) in mid-2026 due to tightened User-Agent enforcement, breaking do_fetch on every clean build of any cargo-based recipe (notably librsvg / librsvg-native pulled in by meta-openembedded). The Fastly CDN at static.crates.io serves byte-identical .crate tarballs without UA enforcement. The rule must live in MIRRORS rather than PREMIRRORS: meta/classes-global/uninative.bbclass enumerates every http* PREMIRRORS rule at BuildStarted and synthesizes a new entry by embedding the rule's replace field as a literal URL suffix. Because our replace contains regex backreferences (\1, \2), the synthesized rule trips re.error: invalid group reference 1 at config-parse time, bricking the build before do_fetch even runs.
• Added recipe-level PARALLEL_MAKE = "-j 4" cap to onnxruntime_1.24.2.bb to keep do_compile from being OOM-killed on 32 GB-class builders — onnxruntime 1.24.2's flash-attention CUDA kernels compile with nvcc spawning multiple cicc/ptxas processes per TU, peaking well above 4 GiB resident per TU. SQA reported such a do_compile OOM on 2026-05-28. Mirrors the existing recipe-level convention used by pytorch_2.11.0.bb (capped at -j 8 for skyloft's 31 GB + 8 GB swap profile). The recipe-level setting overrides the layer-wide PARALLEL_MAKE in build/conf/local.conf for the capped recipe only; downstream users can further override via BitBake's PARALLEL_MAKE:pn-<recipe> mechanism in local.conf. Per-host tuning guidance is documented in env/README.md under "Troubleshooting / Memory exhausted / BB_NUMBER_THREADS has no effect" as a new sub-section on per-recipe PARALLEL_MAKE caps.

Deferred

The following bumps from the Holoscan SDK 4.3.0 release notes were intentionally not taken this cycle:

• NCCL 2.27.5-1 -> 2.29: release notes tie this bump to PyTorch 2.11 compatibility. In meta-tegra-holoscan, pytorch_2.11.0.bb is built with -DUSE_NCCL=OFF, and the only remaining in-layer nccl consumer is python3-cupy_13.6.0.bb, which builds fine against 2.27.
• NSight Systems 2023.3.3.42-1 -> 2025.3.1: the standalone nsight-systems-cli debian package line stopped at 2024.2.3 on the NVIDIA CUDA repo; 2025.x only ships as the full nsight-systems-<ver> deb (different packaging). Will be revisited in a follow-up that switches the recipe accordingly.

v4.2.0

[4.2.0] - 2026-04-23

Changed

• Updated Holoscan SDK and Holohub apps to 4.2.0.
• Updated GXF to 5.6.0-20260429-73f41cf00.
• Updated ONNX to 1.20.1 (required by ONNX Runtime 1.24.2 for the new OpSchema::SetNodeDeterminism, InferenceContext::hasOutput, and TensorProto_DataType_{FLOAT4E2M1,INT2,UINT2} symbols).
• Updated ONNX Runtime to 1.24.2 (cuDNN frontend to v1.12.0).
• Updated UCX to 1.20.0 and UCXX to 0.48.00.
• Updated abseil-cpp to 20250814.0 (required by ONNX Runtime 1.24.2 for the new absl::hashtable_control_bytes symbol).
• Added libeigen 3.4.1 recipe (meta-openembedded scarthgap only ships 3.4.0; ONNX Runtime 1.24.2 needs the templated min<Eigen::PropagateNaN> / max<Eigen::PropagateNaN> overloads added between 3.4.0 and 3.4.1) and pinned PREFERRED_VERSION_libeigen = "3.4.1" in conf/holoscan-common.conf.
• holoscan-sdk: raised CLI11 floor from 2.3.2 to 2.5.0 in 0002-Use-external-dependencies-deps.patch to match SDK 4.2.0's CLI11_VERSION 2.5.0.
• holoscan-sdk: in 0002-Use-external-dependencies-deps.patch, also drop the CCCL 3.2 version pin from the find_dependency(CCCL ...) call inside the holoscan_install_hook_code_string heredoc at top-level CMakeLists.txt. That heredoc is what rapids_export(... FINAL_CODE_BLOCK ...) bakes verbatim into the installed holoscan-config.cmake, so without this drop, every downstream consumer (e.g. holohub-apps) failed do_configure with Could not find a configuration file for package "CCCL" that is compatible with requested version "3.2" even though the SDK itself built fine against external CCCL 3.0.3.

Fixed

• onnxruntime: added libeigen to DEPENDS so the new Eigen3::Eigen link from the CUDA / TensorRT EP modules in ORT 1.24.2 resolves against the meta-tegra Eigen3 sysroot.
• abseil-cpp: dropped obsolete 0002-Remove-maes-option-from-cross-compilation.patch and 0003-Remove-neon-option-from-cross-compilation.patch (the ABSL_RANDOM_HWAES_*_FLAGS lists they targeted were removed from absl/copts/GENERATED_AbseilCopts.cmake upstream in 20250814.0). conf/holoscan-common.conf PREFERRED_VERSION_abseil-cpp bumped to match.

Deferred

The following bumps from the Holoscan SDK 4.2.0 release notes were intentionally not taken this cycle:

• PyTorch 2.5.0 -> 2.11.0 (and the paired torchvision 0.20.1 -> 0.26.0): PyTorch is pinned to the IGX 1.1 / DLFW 24.08 software stack baseline. The PyTorch stack will be revisited in a later cycle.
• NCCL 2.27.5-1 -> 2.29: release notes tie this bump to PyTorch 2.11 compatibility. In meta-tegra-holoscan, pytorch_2.5.0.bb is built with -DUSE_NCCL=OFF, and the only remaining in-layer nccl consumer is python3-cupy_13.6.0.bb, which builds fine against 2.27.
• CCCL 3.0.3 -> 3.2.0 and RMM 25.10.00 -> 26.02.00: both are pinned to the DLFW 26.03 baseline that PyTorch 2.5.0 builds against. The existing 0002-Use-external-dependencies-deps.patch already replaces the SDK's rapids_cpm_find(CCCL 3.2.0) with an unversioned find_package(CCCL REQUIRED) and pins rapids_find_package(rmm 25.10.0 REQUIRED), neutralizing SDK 4.2.0's newer version requests. Note: release notes flag RMM API breakage at the SDK source level; local bitbake validation is required to confirm SDK 4.2.0 source does not rely on any RMM 26.02-only APIs.
• NSight Systems 2023.3.3.42-1 -> 2025.3.1: the standalone nsight-systems-cli debian package line stopped at 2024.2.3 on the NVIDIA CUDA repo; 2025.x only ships as the full nsight-systems-<ver> deb (different packaging). Will be revisited in a follow-up that switches the recipe accordingly.

v4.1.0

[4.1.0] - 2026-04-09

Changed

• Updated Holoscan SDK and Holohub apps to 4.1.0.
• Updated GXF to 5.5.1-20260318-8b9561654.

v4.0.0

2026-03-13

Changed

• Updated Holoscan SDK and Holohub apps to 4.0.0.
• Updated GXF to 5.4.0-20260213-9fe4995d9.
• Updated MatX to 0.9.4.

Fixed

• Fixed magic-enum recipe so #include <magic_enum.hpp> resolves correctly.
• Fixed holohub-apps build with separate GXF by adding find_package(GXF) before holoscan in apps and operators.

Cleanup

• holoscan-sdk: removing passing of obsolete HOLOSCAN_BUILD_DOCS CMake option (removed in Holoscan SDK 0.6.0).
• holoscan-sdk: explicitly set HOLOSCAN_USE_SCCACHE=OFF (HSDK defaults to OFF; keeps OE build deterministic).

v3.11.0

2026-02-17

Changed

• Updated Holoscan SDK and Holohub apps to 3.11.0.
• Updated GXF to 5.3.0-20260130-b825fab47.
• Updated UCX to 1.19.0.
• Dropped Rivermax.
• Dropped Emergent.

v3.10.0

2026-01-22

Changed

• Updated Holoscan SDK and Holohub apps to 3.10.0.
• Updated GXF to 5.2.0-20251212-8f83bf174.
• Updated UCXX to 0.46.00.
• Dropped the prebuild OE builder docker container image from NGC.
  Users can build the OE builder then start building the holoscan reference image.

Fixed

• Fixed Preempt-RT enablement.
• Fixed OpenRM driver build when Preempt-RT is enabled.

v3.9.0

2025-11-26

Changed

• Updated Holoscan SDK and Holohub apps to 3.9.0.
• Updated GXF to 5.1.0-20251114-0652b7b15.
• Updated yaml-cpp to 0.8.0.
• Updated CCCL to 3.0.3.
• Updated CPM cmake to 0.40.2.
• Updated RAPIDS cmake to 25.10.00.
• Updated fmt to 11.0.2.
• Added spdlog 1.14.1 recipe to satisfy dependency with RAPIDS logger.
• Added RAPIDS logger 0.2.0 to satisfy dependency with RMM.
• Updated RMM to 25.10.00 to satisfy dependency with Holoscan SDK.
• Updated NVTX to 3.3.0 to satisfy dependency with CuPy and Holoscan SDK.
• Updated nccl to 2.27.5 to satisfy dependency with CuPy.
• Updated CuPY to 13.6.0.

Fixed

• Fixed Pytorch build with CCCL 3.x.
• Fixed ONNX Runtime build with CCCL 3.x.
• Fixed MatX build with CPM cmake and RAPIDS cmake.
• Fixed vulkan-tools recipe.
• Fixed docker container support for dGPU mode.

v3.8.0

2025-11-05

Changed

• Updated Holoscan SDK and Holohub apps to 3.8.0.
• Updated MatX to 0.9.3.