Changelog

v7.0.0-rc1 (2026-05-29)

First release candidate of the v7 rewrite — a ground-up reimplementation into a reusable, installable C++ library. This is a clean break from v6 — the API is not source-compatible. See docs/upgrading_from_v6.md.

Added

• No-throw API: every fallible call returns Status or Result<T> (no exceptions); TRTCPP_TRY sugar for propagation.
• Name-keyed IO (unordered_map<string, TensorView>) at the inference boundary; non-templated Tensor/TensorView with a runtime DType (replaces v6's Engine<T> and nested vectors).
• Caller-owned CUDA streams (Stream, including Stream::wrap for an external handle); explicit host/device transfers (Tensor::toHost/to/copyFrom) — never an implicit D2H copy.
• Safe engine cache: EngineBuilder::buildOrLoad/buildAndLoad keyed by ONNX content hash + build options + TensorRT version + GPU UUID, JSON sidecar, atomic write, stale-cache detection.
• Dynamic shapes via per-input min/opt/max optimization profiles and a -1-aware Shape.
• EnginePool for concurrent multi-stream inference (one optimization profile per context).
• Quantization that is version-aware and never a silent no-op (Precision::kFp16/kInt8Qdq/ kFp8/…); legacy INT8 calibration is available only when built against TensorRT < 11.
• Optional fused preprocessing sublibrary (tensorrt_cpp_api::preproc): a single CUDA kernel for letterbox-resize → BGR↔RGB → per-channel normalize → HWC→NCHW → cast.
• Optional OpenCV interop header (zero-copy views over cv::Mat / cv::cuda::GpuMat), strictly opt-in behind -DTRT_CPP_API_WITH_OPENCV.
• Optional Python bindings (trtcpp, pybind11 + scikit-build-core) with zero-copy __cuda_array_interface__/DLPack interop, caller streams, and GIL release during inference.
• CMake install/export: find_package(tensorrt_cpp_api) with a relocatable package config and a bundled FindTensorRT module (apt or tarball).
• Reference examples (classification, detection, segmentation, zero-copy Python) and GitHub Actions CI (build + CPU tests, sanitizers, Python wheel, lint).

Changed

• Minimum TensorRT is 10.0; the code is written to the TensorRT 11 surface and version-gates the removed-in-11 features (legacy calibrators, weak typing, IPluginV2).
• C++20 (was C++17); namespace trtcpp; include root tensorrt_cpp_api/; CMake target tensorrt_cpp_api::tensorrt_cpp_api.

Removed

• The v6 API in its entirety — no source-compatibility shim (see the upgrade guide):
  • the monolithic templated Engine<T> and its .inl-in-header implementation;
  • the single Options struct, replaced by BuildOptions + EngineOptions;
  • OpenCV cv::cuda::GpuMat in the inference signatures and the triply-nested std::vector inputs/outputs, replaced by name-keyed TensorView/Tensor;
  • bool/exception error handling, replaced by Status/Result<T>;
  • per-call stream creation, replaced by caller-owned Stream;
  • the run_inference_benchmark CLI executable and the v6 include/ header layout.

---

v6.0

• Implementation now requires TensorRT >= 10.0.

v5.0

• Engine became a class template parameterized on the model's output data type (float, __half, int8_t, int32_t, bool, uint8_t).
• Added loading a TensorRT engine file directly (without compiling from ONNX).
• Added a command-line parser.

v4.1

• Support for fixed batch size > 1.

v4.0

• Added INT8 precision support.

v3.0

• Updated to the TensorRT 8.6 API (IExecutionContext::enqueueV3()).
• Benchmark executable renamed to run_inference_benchmark; takes the ONNX path as an argument.
• Auto-detect supported batch sizes; stop limiting workspace memory.

v2.2

• Serialize the model name as part of the engine file.

v2.1

• Support for models with multiple inputs.

v2.0

• Requires OpenCV with CUDA.
• Support for models with more than one output and for non-batchable models; more error checking.