TensorRT 10.16 OSS Release (#4729)

Signed-off-by: Kevin Chen <kevinch@nvidia.com>

Author: kevinch-nv

.clang-format

@@ -19,8 +19,30 @@ AlwaysBreakTemplateDeclarations: true
 BasedOnStyle: None
 BinPackArguments: true
 BinPackParameters: true
+# Almost the same as Allman style, but explicitly disabling BeforeLambdaBody
+# for backwards compatibility with clang-format-10 Allman style.
+# See also https://reviews.llvm.org/D44609
+BreakBeforeBraces: Custom
+BraceWrapping:
+  AfterCaseLabel:  true
+  AfterClass:      true
+  AfterControlStatement: Always
+  AfterEnum:       true
+  AfterFunction:   true
+  AfterNamespace:  true
+  AfterObjCDeclaration: true
+  AfterStruct:     true
+  AfterUnion:      true
+  AfterExternBlock: true
+  BeforeCatch:     true
+  BeforeElse:      true
+  BeforeLambdaBody: false
+  BeforeWhile:     false
+  IndentBraces:    false
+  SplitEmptyFunction: true
+  SplitEmptyRecord: true
+  SplitEmptyNamespace: true
 BreakBeforeBinaryOperators: All
-BreakBeforeBraces: Allman
 BreakBeforeTernaryOperators: true
 BreakConstructorInitializersBeforeComma: true
 ColumnLimit:     120
@@ -61,6 +83,7 @@ PenaltyExcessCharacter: 1000000
 PenaltyReturnTypeOnItsOwnLine: 60
 PointerAlignment: Left
 PointerBindsToType: false
+QualifierAlignment: Right
 ReflowComments:  true
 SortIncludes:    true
 SpaceAfterCStyleCast: true
@@ -77,4 +100,3 @@ Standard:        Cpp11
 StatementMacros: [API_ENTRY_TRY,TRT_TRY]
 TabWidth:        4
 UseTab:          Never
-...

CHANGELOG.md

@@ -1,4 +1,15 @@
 # TensorRT OSS Release Changelog
+## 10.16 GA - 2026-3-24
+
+- General
+  - Default CUDA version updated to CUDA 13.2.
+
+- Samples
+  - Added sampleDistCollective sample to showcase multi-device execution in TensorRT.
+
+- Parsers
+  - Added kADJUST_FOR_DLA flag to adjust parsing behavior for ONNX models to be more amenable for DLA hardware execution.
+  - Added DistCollective operator support for multi-device execution in TensorRT.
 
 ## 10.15 GA - 2026-2-2
 
@@ -21,11 +32,8 @@
   - Improved error reporting for models with multiple subgraphs, such as `Loop` or `Scan` nodes.
 
 - Demo changes
-  - demoDiffusion: 
-          - Stable Diffusion 1.5, 2.0 and 2.1 pipelines have been deprecated and removed.
-          - Added support for Wan2.2-T2V-A14B Text to Video pipeline
-           
-  
+  - demoDiffusion: Stable Diffusion 1.5, 2.0 and 2.1 pipelines have been deprecated and removed.
+  - Added support for Wan2.2-T2V-A14B Text to Video pipeline
 
 ## 10.14 GA - 2025-11-7
 - Sample changes

CMakeLists.txt

@@ -67,7 +67,7 @@ endif()
 set(CMAKE_SKIP_BUILD_RPATH True)
 
 # CUDA targets
-set(DEFAULT_CUDA_VERSION 13.1.0)
+set(DEFAULT_CUDA_VERSION 13.2)
 set_ifndef(CUDA_VERSION ${DEFAULT_CUDA_VERSION})
 message(STATUS "CUDA version set to ${CUDA_VERSION}")
 
@@ -204,7 +204,6 @@ if(BUILD_SAFE_SAMPLES OR TRT_SAFETY_INFERENCE_ONLY)
         target_link_options(TRTSAFE::nvinfer_safe_shared INTERFACE LINKER:--unresolved-symbols=ignore-in-shared-libs)
         target_link_options(TRTSAFE::nvinfer_safe_debug INTERFACE LINKER:--unresolved-symbols=ignore-in-shared-libs)
     endif()
-
     # Enable unified builder safety features when building safety samples or in inference-only mode.
     add_compile_definitions(ENABLE_UNIFIED_BUILDER=1)
 endif()
@@ -252,7 +251,7 @@ if(TRT_SAFETY_INFERENCE_ONLY)
 endif()
 
 # C++17
-set(CMAKE_CXX_STANDARD 17)
+set(CMAKE_CXX_STANDARD 20)
 set(CMAKE_CXX_STANDARD_REQUIRED ON)
 set(CMAKE_CXX_EXTENSIONS OFF)
 

README.md

@@ -43,14 +43,14 @@ To build the TensorRT-OSS components, you will first need the following software
 
 **TensorRT GA build**
 
-- TensorRT v10.15.1.29
+- TensorRT v10.16.0.72
   - Available from direct download links listed below
 
 **System Packages**
 
 - [CUDA](https://developer.nvidia.com/cuda-toolkit)
   - Recommended versions:
-  - cuda-13.1.0
+  - cuda-13.2.0
   - cuda-12.9.0
 - [CUDNN (optional)](https://developer.nvidia.com/cudnn)
   - cuDNN 8.9
@@ -63,6 +63,7 @@ To build the TensorRT-OSS components, you will first need the following software
 
 **Optional Packages**
 
+- [NCCL](https://developer.nvidia.com/nccl/nccl-download) >= v2.19, < v3.0 — only when building with multi-device support (`-DTRT_BUILD_ENABLE_MULTIDEVICE=ON`) for the `sampleDistCollective` sample.
 - Containerized build
   - [Docker](https://docs.docker.com/install/) >= 19.03
   - [NVIDIA Container Toolkit](https://github.com/NVIDIA/nvidia-docker)
@@ -97,24 +98,24 @@ To build the TensorRT-OSS components, you will first need the following software
 
    Else download and extract the TensorRT GA build from [NVIDIA Developer Zone](https://developer.nvidia.com) with the direct links below:
 
-   - [TensorRT 10.15.1.29 for CUDA 13.1, Linux x86_64](https://developer.nvidia.com/downloads/compute/machine-learning/tensorrt/10.15.1/tars/TensorRT-10.15.1.29.Linux.x86_64-gnu.cuda-13.1.tar.gz)
-   - [TensorRT 10.15.1.29 for CUDA 12.9, Linux x86_64](https://developer.nvidia.com/downloads/compute/machine-learning/tensorrt/10.15.1/tars/TensorRT-10.15.1.29.Linux.x86_64-gnu.cuda-12.9.tar.gz)
-   - [TensorRT 10.15.1.29 for CUDA 13.1, Windows x86_64](https://developer.nvidia.com/downloads/compute/machine-learning/tensorrt/10.15.1/zip/TensorRT-10.15.1.29.Windows.win10.cuda-13.1.zip)
-   - [TensorRT 10.15.1.29 for CUDA 12.9, Windows x86_64](https://developer.nvidia.com/downloads/compute/machine-learning/tensorrt/10.15.1/zip/TensorRT-10.15.1.29.Windows.win10.cuda-12.9.zip)
+   - [TensorRT 10.16.0.72 for CUDA 13.2, Linux x86_64](https://developer.nvidia.com/downloads/compute/machine-learning/tensorrt/10.16.0/tars/TensorRT-10.16.0.72.Linux.x86_64-gnu.cuda-13.2.tar.gz)
+   - [TensorRT 10.16.0.72 for CUDA 12.9, Linux x86_64](https://developer.nvidia.com/downloads/compute/machine-learning/tensorrt/10.16.0/tars/TensorRT-10.16.0.72.Linux.x86_64-gnu.cuda-12.9.tar.gz)
+   - [TensorRT 10.16.0.72 for CUDA 13.2, Windows x86_64](https://developer.nvidia.com/downloads/compute/machine-learning/tensorrt/10.16.0/zip/TensorRT-10.16.0.72.Windows.win10.cuda-13.2.zip)
+   - [TensorRT 10.16.0.72 for CUDA 12.9, Windows x86_64](https://developer.nvidia.com/downloads/compute/machine-learning/tensorrt/10.16.0/zip/TensorRT-10.16.0.72.Windows.win10.cuda-12.9.zip)
 
-   **Example: Ubuntu 22.04 on x86-64 with cuda-13.1**
+   **Example: Ubuntu 22.04 on x86-64 with cuda-13.2**
 
    ```bash
    cd ~/Downloads
-   tar -xvzf TensorRT-10.15.1.29.Linux.x86_64-gnu.cuda-13.1.tar.gz
-   export TRT_LIBPATH=`pwd`/TensorRT-10.15.1.29/lib
+   tar -xvzf TensorRT-10.16.0.72.Linux.x86_64-gnu.cuda-13.2.tar.gz
+   export TRT_LIBPATH=`pwd`/TensorRT-10.16.0.72/lib
    ```
 
    **Example: Windows on x86-64 with cuda-12.9**
 
    ```powershell
-   Expand-Archive -Path TensorRT-10.15.1.29.Windows.win10.cuda-12.9.zip
-   $env:TRT_LIBPATH="$pwd\TensorRT-10.15.1.29\lib"
+   Expand-Archive -Path TensorRT-10.16.0.72.Windows.win10.cuda-12.9.zip
+   $env:TRT_LIBPATH="$pwd\TensorRT-10.16.0.72\lib"
    ```
 
 ## Setting Up The Build Environment
@@ -123,34 +124,34 @@ For Linux platforms, we recommend that you generate a docker container for build
 
 1. #### Generate the TensorRT-OSS build container.
 
-   **Example: Ubuntu 24.04 on x86-64 with cuda-13.1 (default)**
+   **Example: Ubuntu 24.04 on x86-64 with cuda-13.2 (default)**
 
    ```bash
-   ./docker/build.sh --file docker/ubuntu-24.04.Dockerfile --tag tensorrt-ubuntu24.04-cuda13.1
+   ./docker/build.sh --file docker/ubuntu-24.04.Dockerfile --tag tensorrt-ubuntu24.04-cuda13.2
    ```
 
-   **Example: Rockylinux8 on x86-64 with cuda-13.1**
+   **Example: Rockylinux8 on x86-64 with cuda-13.2**
 
    ```bash
-   ./docker/build.sh --file docker/rockylinux8.Dockerfile --tag tensorrt-rockylinux8-cuda13.1
+   ./docker/build.sh --file docker/rockylinux8.Dockerfile --tag tensorrt-rockylinux8-cuda13.2
    ```
 
-   **Example: Ubuntu 24.04 cross-compile for Jetson (aarch64) with cuda-13.1 (JetPack SDK)**
+   **Example: Ubuntu 24.04 cross-compile for Jetson (aarch64) with cuda-13.2 (JetPack SDK)**
 
    ```bash
-   ./docker/build.sh --file docker/ubuntu-cross-aarch64.Dockerfile --tag tensorrt-jetpack-cuda13.1
+   ./docker/build.sh --file docker/ubuntu-cross-aarch64.Dockerfile --tag tensorrt-jetpack-cuda13.2
    ```
 
-   **Example: Ubuntu 24.04 on aarch64 with cuda-13.1**
+   **Example: Ubuntu 24.04 on aarch64 with cuda-13.2**
 
    ```bash
-   ./docker/build.sh --file docker/ubuntu-24.04-aarch64.Dockerfile --tag tensorrt-aarch64-ubuntu24.04-cuda13.1
+   ./docker/build.sh --file docker/ubuntu-24.04-aarch64.Dockerfile --tag tensorrt-aarch64-ubuntu24.04-cuda13.2
    ```
 
 2. #### Launch the TensorRT-OSS build container.
    **Example: Ubuntu 24.04 build container**
    ```bash
-   ./docker/launch.sh --tag tensorrt-ubuntu24.04-cuda13.1 --gpus all
+   ./docker/launch.sh --tag tensorrt-ubuntu24.04-cuda13.2 --gpus all
    ```
    > NOTE:
    > <br> 1. Use the `--tag` corresponding to build container generated in Step 1.
@@ -163,7 +164,7 @@ For Linux platforms, we recommend that you generate a docker container for build
 
 - Generate Makefiles and build
 
-  **Example: Linux (x86-64) build with default cuda-13.1**
+  **Example: Linux (x86-64) build with default cuda-13.2**
 
   ```bash
   cd $TRT_OSSPATH
@@ -172,7 +173,7 @@ For Linux platforms, we recommend that you generate a docker container for build
   make -j$(nproc)
   ```
 
-  **Example: Linux (aarch64) build with default cuda-13.1**
+  **Example: Linux (aarch64) build with default cuda-13.2**
 
   ```bash
   cd $TRT_OSSPATH
@@ -181,7 +182,7 @@ For Linux platforms, we recommend that you generate a docker container for build
   make -j$(nproc)
   ```
 
-  **Example: Native build on Jetson Thor (aarch64) with cuda-13.1**
+  **Example: Native build on Jetson Thor (aarch64) with cuda-13.2**
 
   ```bash
   cd $TRT_OSSPATH
@@ -192,7 +193,7 @@ For Linux platforms, we recommend that you generate a docker container for build
 
   > NOTE: C compiler must be explicitly specified via CC= for native aarch64 builds of protobuf.
 
-  **Example: Ubuntu 24.04 Cross-Compile for Jetson Thor (aarch64) with cuda-13.1 (JetPack)**
+  **Example: Ubuntu 24.04 Cross-Compile for Jetson Thor (aarch64) with cuda-13.2 (JetPack)**
 
   ```bash
   cd $TRT_OSSPATH
@@ -201,7 +202,7 @@ For Linux platforms, we recommend that you generate a docker container for build
   make -j$(nproc)
   ```
 
-  **Example: Ubuntu 24.04 Cross-Compile for DriveOS (aarch64) with cuda-13.1**
+  **Example: Ubuntu 24.04 Cross-Compile for DriveOS (aarch64) with cuda-13.2**
 
   ```bash
   cd $TRT_OSSPATH
@@ -210,7 +211,7 @@ For Linux platforms, we recommend that you generate a docker container for build
   make -j$(nproc)
   ```
 
-  **Example: Native builds on Windows (x86) with cuda-13.1**
+  **Example: Native builds on Windows (x86) with cuda-13.2**
 
   ```bash
   cd $TRT_OSSPATH
@@ -220,7 +221,7 @@ For Linux platforms, we recommend that you generate a docker container for build
   msbuild TensorRT.sln /property:Configuration=Release -m:$env:NUMBER_OF_PROCESSORS
   ```
 
-  > NOTE: The default CUDA version used by CMake is 13.1. To override this, for example to 12.9, append `-DCUDA_VERSION=12.9` to the cmake command.
+  > NOTE: The default CUDA version used by CMake is 13.2. To override this, for example to 12.9, append `-DCUDA_VERSION=12.9` to the cmake command.
 
 - Required CMake build arguments are:
   - `TRT_LIB_DIR`: Path to the TensorRT installation directory containing libraries.
@@ -238,6 +239,7 @@ For Linux platforms, we recommend that you generate a docker container for build
   - `TRT_SAFETY_INFERENCE_ONLY`: Specify if only build the safety inference components, for example [`ON`] | `OFF`. If turned ON, all other components will be turned OFF except `BUILD_SAFE_SAMPLES`.
   - `GPU_ARCHS`: GPU (SM) architectures to target. By default we generate CUDA code for all major SMs. Specific SM versions can be specified here as a quoted space-separated list to reduce compilation time and binary size. Table of compute capabilities of NVIDIA GPUs can be found [here](https://developer.nvidia.com/cuda-gpus). Examples: - NVidia A100: `-DGPU_ARCHS="80"` - RTX 50 series: `-DGPU_ARCHS="120"` - Multiple SMs: `-DGPU_ARCHS="80 120"`
   - `TRT_PLATFORM_ID`: Bare-metal build (unlike containerized cross-compilation). Currently supported options: `x86_64` (default).
+  - `TRT_BUILD_ENABLE_MULTIDEVICE`: Enable the multi-device sample (`sampleDistCollective`). Use `-DTRT_BUILD_ENABLE_MULTIDEVICE=ON` to build it; requires [NCCL](https://developer.nvidia.com/nccl/nccl-download) >= v2.19, < v3.0.
 
 ## Building TensorRT DriveOS Samples
 
@@ -313,7 +315,7 @@ For Linux platforms, we recommend that you generate a docker container for build
   ```bash
   cd $TRT_OSSPATH
   mkdir -p build && cd build
-  export CUDA_VERSION=13.1
+  export CUDA_VERSION=13.2
   export CUDA=cuda-$CUDA_VERSION
   export CUDA_ROOT=/usr/local/cuda-safe-$CUDA_VERSION
   export QNX_BASE=/drive/toolchains/qnx_toolchain  # Set to your QNX toolchain installation path

VERSION

@@ -1 +1 @@
-10.15.1.29
+10.16.0.72