Skip to content

What's New in Holoscan 4.6: Runtime-Adaptable HoloLink Modules and Stronger Data Validation

Written September 2, 2026

Holoscan SDK 4.6 brings Holoscan Sensor Bridge (HoloLink) 2.7 into the SDK development container with precompiled, runtime-adaptable device modules. The release also strengthens validation at several data and process boundaries, improves failure handling in HoloInfer and Holoviz, and smooths C++17 CUDA source builds.

You can get Holoscan 4.6 as a Docker container (v4.6.0-cuda13, v4.6.0-cuda12-dgpu, or v4.6.0-cuda12-igpu), a Python wheel for CUDA 12 (holoscan, pip install holoscan==4.6.0) or CUDA 13 (holoscan-cu13, pip install holoscan-cu13==4.6.0), Debian packages (4.6.0.0-1), or the source from the Holoscan SDK v4.6.0 release.

Holoscan SDK 4.6 ships precompiled hololink_module binaries in its development container. Applications can use them to adapt to different Holoscan Sensor Bridge devices and HSB-IP versions without recompilation. See the complete Holoscan Sensor Bridge 2.7 release notes and user guide for upstream details.

Stronger Validation at Data and Process Boundaries

Holoscan 4.6 hardens multiple deserialization paths so malformed or incomplete input is rejected before it can mutate application state or lead to invalid memory access.

PoseTree's UCX path now validates frame names and detects invalid or truncated snapshots before applying them. Distributed-application and Holoviz codecs add checks for invalid sizes and counts, incomplete payloads, and failed-value access. Python GXF recording deserialization likewise rejects truncated headers, inconsistent tensor metadata, invalid payload ranges, and malformed timestamps.

The application worker also validates requested port counts and ranges before reserving memory. Impossible port-allocation requests now fail early with a clear validation path instead of progressing into resource allocation.

These changes preserve the existing public APIs while making applications more predictable when they encounter damaged recordings, malformed network payloads, or invalid orchestration requests.

More Resilient Inference and Visualization

HoloInfer fixes a race that could surface during simultaneous multi-context initialization as Current Manager not initialized. The public lifecycle API is unchanged, so existing applications receive the reliability improvement without a migration.

Holoviz now handles Vulkan device-creation failures using the returned VkResult. Rather than continuing with an invalid device and crashing later, it reports the original failure at the point of device creation, making platform and driver issues easier to diagnose.

C++17 Source-Build Polish

For C++ source builds, the bundled fmt integration is adjusted to avoid NVCC diagnostic 2417-D when compiling in C++17 mode. This removes an unnecessary warning from a common downstream build configuration while retaining the SDK's existing C++ interfaces.

Getting Started

Install the CUDA 12 wheel with pip install holoscan==4.6.0, install the CUDA 13 wheel with pip install holoscan-cu13==4.6.0, or pull the appropriate container from NGC. Platform compatibility, setup instructions, and the complete release notes are available in the Holoscan SDK User Guide and the v4.6.0 GitHub Release.

Join the conversation on the NVIDIA Developer Forum or Discord, explore the SDK source on GitHub, and share applications, operators, and other contributions through HoloHub.