NVIDIA · kevinch-nv · Aug 4, 2026 · Aug 4, 2026
@@ -1,4 +1,15 @@
 # TensorRT OSS Release Changelog
+## 11.2 GA - 2026-8-4
+- Samples
+ - Added a new python sample sample_plugin_v2_to_v3_migration to showcase how to migrate from IPluginV2 to IPluginV3.
+
+- Plugins
+ - Added a new FFTPlugin, a cuFFT-backed plugin for complex-to-complex, real-to-complex, and complex-to-real transforms, to support the ONNX DFT operator.
+
+- Parsers
+ - Added IRefitterObserver class for better refitting of ONNX models.
+ - Added support for the DFT operator and 5D GridSample operators.
+
 ## 11.1 GA - 2026-6-24
 - General
   - Default CUDA version updated to 13.3.

@@ -139,12 +139,15 @@ option(BUILD_PLUGINS "Build TensorRT plugin" ON)
 option(BUILD_PARSERS "Build TensorRT parsers" ON)
 option(BUILD_SAMPLES "Build TensorRT samples" ON)
 option(BUILD_SAFE_SAMPLES "Build TensorRT safety samples" OFF)
+option(BUILD_PYTHON "Build TensorRT python bindings" OFF)
 option(TRT_SAFETY_INFERENCE_ONLY "Build only the safety inference components (no safety builders)" OFF)
 option(TRT_BUILD_TESTING "Build gtests for TensorRT components" OFF)
 
 # Must be set before add_subdirectory(plugin); the option() below is gated on BUILD_SAMPLES.
 set(TRT_BUILD_WINML OFF)
 
+set(TRT_PRODUCT_IS_RTX OFF CACHE INTERNAL "") # TRT-OSS doesn't support TensorRT-RTX
+
 ############################################################################################
 # Early dependency discovery
 # These must be found before they are used in target definitions
@@ -349,7 +352,16 @@ else()
 endif()
 
 find_library_create_target(nvinfer ${nvinfer_lib_name} SHARED "${TRT_LIB_DIR}")
-set_property(TARGET nvinfer PROPERTY INTERFACE_INCLUDE_DIRECTORIES ${TRT_INCLUDE_DIR})
+# Include the OSS repo headers before the package headers so that impl/ headers
+# checked into the repo (e.g. impl/NvInferPythonPlugin.h) take precedence over
+# any same-named impl/ files in the downloaded package, which may differ.
+set_property(TARGET nvinfer PROPERTY INTERFACE_INCLUDE_DIRECTORIES
+    "${CMAKE_CURRENT_SOURCE_DIR}/include"
+    "${TRT_INCLUDE_DIR}")
+
+if(CMAKE_CROSSCOMPILING)
+    target_link_options(nvinfer INTERFACE "LINKER:--unresolved-symbols=ignore-in-shared-libs")
+endif()
 
 # tensorrt is aliased downstream; CMake forbids aliasing an alias.
 add_library(tensorrt INTERFACE IMPORTED)
@@ -378,6 +390,7 @@ set(HINT_PATHS "${TRT_OUT_DIR}" "${TRT_LIB_DIR}")
 if(NOT TARGET trt_global_definitions)
     add_library(trt_global_definitions INTERFACE)
     target_include_directories(trt_global_definitions INTERFACE ${CUDAToolkit_INCLUDE_DIRS})
+    target_compile_definitions(trt_global_definitions INTERFACE TRT_BUILD_ONNX_PARSER=1)
 endif()
 
 if(BUILD_PLUGINS)
@@ -399,6 +412,58 @@ endif()
 add_library(tensorrt_headers INTERFACE)
 target_include_directories(tensorrt_headers INTERFACE ${TRT_INCLUDE_DIR})
 
+# Python bindings
+if(BUILD_PYTHON)
+    include(FetchContent)
+    include(Platforms)
+
+    # nvonnxparser is always defined (built or imported), regardless of BUILD_PARSERS.
+    set(TRT_BUILD_ONNX_PARSER ON)
+    set(TRT_BUILD_PLUGINS ${BUILD_PLUGINS})
+
+    if(NOT DEFINED TRT_BUILD_PLATFORM)
+        set(TRT_BUILD_PLATFORM ${CMAKE_SYSTEM_PROCESSOR})
+    endif()
+
+    # Default to bindings for just the interpreter's Python version; callers
+    # can still expand by passing -DTRT_BUILD_PYTHON_PY_VERSIONS=...
+    find_package(Python3 COMPONENTS Interpreter REQUIRED)
+    set(TRT_BUILD_PYTHON_PY_VERSIONS "${Python3_VERSION_MAJOR}.${Python3_VERSION_MINOR}"
+        CACHE STRING "The list of python versions to build TensorRT bindings for.")
+
+    # The lean / dispatch bindings link the prebuilt runtime libs. Import them as
+    # targets the same way nvinfer is imported above so they are always in-graph.
+    if(MSVC)
+        set(nvinfer_lean_lib_name "nvinfer_lean_${TRT_SOVERSION}${TRT_LIB_SUFFIX}")
+        set(nvinfer_dispatch_lib_name "nvinfer_dispatch_${TRT_SOVERSION}${TRT_LIB_SUFFIX}")
+    else()
+        set(nvinfer_lean_lib_name "nvinfer_lean")
+        set(nvinfer_dispatch_lib_name "nvinfer_dispatch")
+    endif()
+    find_library_create_target(tensorrt_lean_runtime ${nvinfer_lean_lib_name} SHARED "${TRT_LIB_DIR}")
+    set_property(TARGET tensorrt_lean_runtime PROPERTY INTERFACE_INCLUDE_DIRECTORIES
+        "${CMAKE_CURRENT_SOURCE_DIR}/include"
+        "${TRT_INCLUDE_DIR}")
+    find_library_create_target(tensorrt_dispatch_runtime ${nvinfer_dispatch_lib_name} SHARED "${TRT_LIB_DIR}")
+    set_property(TARGET tensorrt_dispatch_runtime PROPERTY INTERFACE_INCLUDE_DIRECTORIES
+        "${CMAKE_CURRENT_SOURCE_DIR}/include"
+        "${TRT_INCLUDE_DIR}")
+
+    # OSS uses the system Python headers (the Debian/Ubuntu python<ver>-dev layout).
+    set(TRT_BUILD_PYTHON_EXTERNALS_PATH "/usr/include" CACHE PATH
+        "Path to the parent folder of the versioned python headers/libs.")
+
+    # Variables consumed by python/packaging when building the bindings wheel.
+    # (TensorRT_VERSION / TensorRT_SOURCE_DIR are provided by project(TensorRT VERSION ...).)
+    string(TOLOWER ${CMAKE_SYSTEM_PROCESSOR} TRT_LOWERCASE_CMAKE_SYSTEM_PROCESSOR)
+    set(TENSORRT_BASE_NAME "nvinfer")
+    set(TRT_CUDA_VERSION "${CUDAToolkit_VERSION_MAJOR}.${CUDAToolkit_VERSION_MINOR}")
+    set(TensorRT_PACKAGE_VERSION "${TRT_VERSION}.${TRT_BUILD}")
+    set(TRT_BUILD_PYTHON_STANDALONE_WHEELS OFF)
+
+    add_subdirectory(python)
+endif()
+
 # Samples
 if(BUILD_SAMPLES OR BUILD_SAFE_SAMPLES)
     set(TRT_BUILD_ENABLE_NEW_SAMPLES_FLOW ON)

@@ -48,7 +48,7 @@ To build the TensorRT-OSS components, you will first need the following software
 
 **TensorRT GA build**
 
-- TensorRT v11.1.0.106
+- TensorRT v11.2.1.2
   - Available from direct download links listed below
 
 **System Packages**
@@ -103,24 +103,24 @@ 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 11.1.0.106 for CUDA 13.3, Linux x86_64](https://developer.nvidia.com/downloads/compute/machine-learning/tensorrt/11.1.0/tars/TensorRT-Enterprise-11.1.0.106-Linux-x86_64-cuda-13.3-Release-external.tar.zst)
-   - [TensorRT 11.1.0.106 for CUDA 12.9, Linux x86_64](https://developer.nvidia.com/downloads/compute/machine-learning/tensorrt/11.1.0/tars/TensorRT-Enterprise-11.1.0.106-Linux-x86_64-cuda-12.9-Release-external.tar.zst)
-   - [TensorRT 11.1.0.106 for CUDA 13.3, Windows x86_64](https://developer.nvidia.com/downloads/compute/machine-learning/tensorrt/11.1.0/zip/TensorRT-Enterprise-11.1.0.106-Windows-amd64-cuda-13.3-Release-external.zip)
-   - [TensorRT 11.1.0.106 for CUDA 12.9, Windows x86_64](https://developer.nvidia.com/downloads/compute/machine-learning/tensorrt/11.1.0/zip/TensorRT-Enterprise-11.1.0.106-Windows-amd64-cuda-12.9-Release-external.zip)
+   - [TensorRT 11.2.1.2 for CUDA 13.3, Linux x86_64](https://developer.nvidia.com/downloads/compute/machine-learning/tensorrt/11.2.1/tars/TensorRT-Enterprise-11.2.1.2-Linux-x86_64-cuda-13.3-Release-external.tar.zst)
+   - [TensorRT 11.2.1.2 for CUDA 12.9, Linux x86_64](https://developer.nvidia.com/downloads/compute/machine-learning/tensorrt/11.2.1/tars/TensorRT-Enterprise-11.2.1.2-Linux-x86_64-cuda-12.9-Release-external.tar.zst)
+   - [TensorRT 11.2.1.2 for CUDA 13.3, Windows x86_64](https://developer.nvidia.com/downloads/compute/machine-learning/tensorrt/11.2.1/zip/TensorRT-Enterprise-11.2.1.2-Windows-amd64-cuda-13.3-Release-external.zip)
+   - [TensorRT 11.2.1.2 for CUDA 12.9, Windows x86_64](https://developer.nvidia.com/downloads/compute/machine-learning/tensorrt/11.2.1/zip/TensorRT-Enterprise-11.2.1.2-Windows-amd64-cuda-12.9-Release-external.zip)
 
    **Example: Ubuntu 22.04 on x86-64 with cuda-13.3**
 
    ```bash
    cd ~/Downloads
-   tar --zstd -xvf TensorRT-Enterprise-11.1.0.106-Linux-x86_64-cuda-13.3-Release-external.tar.zst
-   export TRT_LIBPATH=`pwd`/TensorRT-11.1.0.106/lib
+   tar --zstd -xvf TensorRT-Enterprise-11.2.1.2-Linux-x86_64-cuda-13.3-Release-external.tar.zst
+   export TRT_LIBPATH=`pwd`/TensorRT-11.2.1.2/lib
    ```
 
    **Example: Windows on x86-64 with cuda-12.9**
 
    ```powershell
-   Expand-Archive -Path TensorRT-Enterprise-11.1.0.106-Windows-amd64-cuda-12.9-Release-external.zip
-   $env:TRT_LIBPATH="$pwd\TensorRT-11.1.0.106\lib"
+   Expand-Archive -Path TensorRT-Enterprise-11.2.1.2-Windows-amd64-cuda-12.9-Release-external.zip
+   $env:TRT_LIBPATH="$pwd\TensorRT-11.2.1.2\lib"
    ```
 
 ## Setting Up The Build Environment

@@ -1 +1 @@
-11.1.0.106
+11.2.1.2
@@ -24,6 +24,7 @@ set(CCCL_TAG "v3.4.0-rc0" CACHE STRING "The commit hash to FetchContent_Declare
 # We use this directory to ensure we only fetch a single copy of dependencies, even between builds.
 # $HOME/storage is expected to be mounted from the host for developers.
 set(TRT_THIRD_PARTY_DL_DIR "$ENV{HOME}/storage" CACHE PATH "Directory to download third party dependencies to")
+file(TO_CMAKE_PATH "${TRT_THIRD_PARTY_DL_DIR}" TRT_THIRD_PARTY_DL_DIR)
 
 FetchContent_Declare(
     cccl

@@ -0,0 +1,41 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+include_guard()
+
+# This is set and immediately overwritten intentionally. It's here to document the public repo, and to provide the boilerplate we'd use if we ever ship it publicly:
+set(_pybind11_default_repo "https://github.com/pybind/pybind11.git")
+
+
+set(PYBIND11_REPO ${_pybind11_default_repo} CACHE STRING "The base project URL to FetchContent_Declare for pybind11" )
+set(PYBIND11_TAG "v3.0.1" CACHE STRING "The commit hash to FetchContent_Declare for pybind11")
+
+# We use this directory to ensure we only fetch a single copy of dependencies, even between builds.
+# $HOME/storage is expected to be mounted from the host for developers.
+set(TRT_THIRD_PARTY_DL_DIR "$ENV{HOME}/storage" CACHE PATH "Directory to download third party dependencies to")
+file(TO_CMAKE_PATH "${TRT_THIRD_PARTY_DL_DIR}" TRT_THIRD_PARTY_DL_DIR)
+
+FetchContent_Declare(
+    pybind11
+    PREFIX              "${CMAKE_BINARY_DIR}/third_party/pybind11"
+    GIT_REPOSITORY      ${PYBIND11_REPO}
+    GIT_TAG             ${PYBIND11_TAG}
+    GIT_SHALLOW         TRUE
+    SOURCE_DIR          "${TRT_THIRD_PARTY_DL_DIR}/pybind11/${PYBIND11_TAG}"
+    EXCLUDE_FROM_ALL
+    UPDATE_DISCONNECTED ${TRT_FETCH_CONTENT_UPDATES_DISCONNECTED}
+    OVERRIDE_FIND_PACKAGE    # ONNX is going to try and look for pybind11, so we redirect it to here.
+)
+FetchContent_MakeAvailable(pybind11)
@@ -0,0 +1,24 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# \brief Converts a truthy flag (usually a boolean) to an integer (0 or 1).
+# \param flagName The name of the flag to convert.
+# \return A CMake variable with the same name as the flag, but suffixed with "_INT" containing 1 if the flag was true and 0 otherwise.
+function(flagToInt flagName)
+    if(${${flagName}})
+        set(${flagName}_INT 1 PARENT_SCOPE)
+    else()
+        set(${flagName}_INT 0 PARENT_SCOPE)
+    endif()
+endfunction()
@@ -1,4 +1,4 @@
-# SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-FileCopyrightText: Copyright (c) 2025-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 # SPDX-License-Identifier: Apache-2.0
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
@@ -31,7 +31,7 @@ function(update_windows_output_name target_name major_version minor_version)
             set(tgt_output_name ${target_name})
         endif()
 
-        if(${TRT_BUILD_WINML})
+        if(${TRT_PRODUCT_IS_RTX})
             set(tgt_output_name "${tgt_output_name}_${major_version}_${minor_version}")
         else()
             set(tgt_output_name "${tgt_output_name}_${major_version}")

@@ -7,7 +7,7 @@ This demo application ("demoDiffusion") showcases the acceleration of Stable Dif
 ### Clone the TensorRT OSS repository
 
 ```bash
-git clone git@github.com:NVIDIA/TensorRT.git -b release/11.0 --single-branch
+git clone git@github.com:NVIDIA/TensorRT.git
 cd TensorRT
 ```