Is this page helpful?

TensorRT 10.6.0 Release Notes#

These Release Notes apply to x86 Linux and Windows users, ARM-based CPU cores for Server Base System Architecture (SBSA) users on Linux, and JetPack users. This release includes several fixes from the previous TensorRT releases and additional changes.

Announcements#

Quickly Deployable Plugins (QDPs): This release introduces a new Python plugin framework that simplifies TensorRT plugin development, allowing for rapid prototyping and deployment of custom layers.

Deprecated APIs: Several APIs have been deprecated in this release, including fcPlugin(CustomFCPluginDynamic) and INT8 implicit quantization APIs.

Key Features and Enhancements#

Quantization and Precision Enhancements

  • FP8 Multi-Head Attention (MHA) on Ada GPUs: Enabled FP8 MHA fusion on Ada GPUs for MHAs with sequence lengths up to 3072 and head sizes up to 64, improving performance for transformer-based models.

  • Per-Channel Scaling for FP8 GEMMs: Added support for per-channel scaling for FP8 matrix multiplications on Hopper GPUs, enabling more flexible quantization strategies.

Model Support and Optimizations

  • FLUX Model Enhancements: Added FP8 support to FLUX models in TensorRT open source demoDiffusion on Hopper GPUs, and optimized FLUX performance across FP8, FP16, and INT8 precisions.

Developer Tools

  • Quickly Deployable Plugins (QDPs): Introduced a new Python-based plugin framework that provides a simple and intuitive approach to defining TensorRT plugins. This feature enables rapid development and deployment of custom layers without extensive C++ knowledge.

API Enhancements

Compatibility#

Limitations#

  • There is a known issue with using the markDebug API to mark multiple graph input tensors as debug tensors.

  • There are no optimized FP8 Convolutions for Group Convolutions and Depthwise Convolutions. Therefore, INT8 is still recommended for ConvNets containing these convolution ops.

  • The FP8 Convolutions only support input/output channels, which are multiples of 16. Otherwise, TensorRT will fall back to non-FP8 convolutions.

  • The FP8 Convolutions do not support kernel sizes larger than 32 (for example, 7x7 convolutions); FP16 or FP32 fallback kernels will be used with suboptimal performance. Therefore, do not add FP8 Q/DQ ops before Convolutions with large kernel sizes for better performance.

  • The accumulation dtype for the batched GEMMS in the FP8 MHA must be in FP32.

    • This can be achieved by adding Cast (to FP32) ops before the batched GEMM and Cast (to FP16) after the batched GEMM.

    • Alternatively, you can convert your ONNX model using TensorRT Model Optimizer, which adds the Cast ops automatically.

  • There cannot be any pointwise operations between the first batched GEMM and the softmax inside FP8 MHAs (for example, having an attention mask). This will be improved in future TensorRT releases.

  • The FP8 MHA fusions only support head sizes being multiples of 16. If the MHA has a head size that is not a multiple of 16, do not add Q/DQ ops in the MHA to fall back to the FP16 MHA for better performance.

  • On QNX, networks that are segmented into a large number of DLA loadables can fail during inference.

  • The DLA compiler can remove identity transposes but cannot fuse multiple adjacent transpose layers into a single transpose layer (likewise for reshaping). For example, given a TensorRT IShuffleLayer consisting of two non-trivial transposes and an identity reshape in between, the shuffle layer is translated into two consecutive DLA transpose layers unless you manually merge the transposes in the model definition in advance.

  • nvinfer1::UnaryOperation::kROUND or nvinfer1::UnaryOperation::kSIGN operations of IUnaryLayer are not supported in the implicit batch mode.

  • For networks containing normalization layers, particularly if deploying with mixed precision, target the latest ONNX opset containing the corresponding function ops (for example, opset 17 for LayerNormalization or opset 18 GroupNormalization). Numerical accuracy using function ops is superior to the corresponding implementation with primitive ops for normalization layers.

  • Weight streaming mainly supports GEMM-based networks like Transformers for now. Convolution-based networks may have only a few weights that can be streamed.

  • When two convolutions with INT8-QDQ and residual add share the same weight, constant weight fusion does not occur. Make a copy of the shared weight for better performance.

  • When building the nonZeroPlugin sample on Windows, you might need to modify the CUDA version specified in the BuildCustomizations paths in the vcxproj file to match the installed version of CUDA.

Deprecated API Lifetime#

  • APIs deprecated in TensorRT 10.6 will be retained until 11/2025.

  • APIs deprecated in TensorRT 10.5 will be retained until 10/2025.

  • APIs deprecated in TensorRT 10.4 will be retained until 9/2025.

  • APIs deprecated in TensorRT 10.3 will be retained until 8/2025.

  • APIs deprecated in TensorRT 10.2 will be retained until 7/2025.

  • APIs deprecated in TensorRT 10.1 will be retained until 5/2025.

  • APIs deprecated in TensorRT 10.0 will be retained until 3/2025.

Deprecated and Removed Features#

  • For a complete list of deprecated C++ APIs, refer to the C++ API Deprecated List.

  • Deprecated fcPlugin(CustomFCPluginDynamic). It is superseded by TensorRT’s IMatrixMultiplyLayer.

  • Deprecated INT8 implicit quantization and calibrator APIs including dynamicRangeIsSet, CalibrationAlgoType, IInt8Calibrator, IInt8EntropyCalibrator, IInt8EntropyCalibrator2, IInt8MinMaxCalibrator, IInt8Calibrator, setInt8Calibrator, getInt8Calibrator, setCalibrationProfile, getCalibrationProfile, setDynamicRange, getDynamicRangeMin, getDynamicRangeMax, and getTensorsWithDynamicRange. They may not give the optimal performance and accuracy. As a workaround, use INT8 explicit quantization instead.

Fixed Issues#

  • The supported I/O formats for the getTensorFormatDesc API were incorrect. This means that the comments for the TensorFormat enum fields might also have been incorrect, and that getFormatDescStr might have indicated Unknown format for a supported combination, or falsely indicated that a combination was valid. This issue has been fixed.

Known Issues#

Functional

  • If IOneHotLayer is used to compute a shape tensor, the builder will fail with a message that a IOneHotLayer cannot be used to compute a shape tensor.

  • There is a known accuracy issue running certain networks on NVIDIA HGX H20.

  • Inputs to the IRecurrenceLayer must always have the same shape. This means that ONNX models with loops whose recurrence inputs change shapes will be rejected.

  • If TensorRT 8.6 or 9.x was installed using the Python Package Index (PyPI), you cannot upgrade TensorRT to 10.x using PyPI. You must first uninstall TensorRT using pip uninstall tensorrt tensorrt-libs tensorrt-bindings, then reinstall TensorRT using pip install tensorrt. This will remove the previous TensorRT version and install the latest TensorRT 10.x. This step is required because the suffix -cuXX was added to the Python package names, which prevents the upgrade from working properly.

  • CUDA compute sanitizer may report racecheck hazards for some legacy kernels. However, related kernels do not have functional issues at runtime.

  • The compute sanitizer initcheck tool may flag false positive Uninitialized __global__ memory read errors when running TensorRT applications on NVIDIA Hopper GPUs. These errors can be safely ignored and will be fixed in an upcoming CUDA release.

  • Multi-Head Attention fusion might not happen and affect performance if the number of heads is small.

  • An occurrence of use-after-free in NVRTC has been fixed in CUDA 12.1. When using NVRTC from CUDA 12.0 together with the TensorRT static library, you may encounter a crash in certain scenarios. Linking the NVRTC and PTXJIT compiler from CUDA 12.1 or newer will resolve this issue.

  • There are known issues reported by the Valgrind memory leak check tool when detecting potential memory leaks from TensorRT applications. The recommendation to suppress the issues is to provide a Valgrind suppression file with the following contents when running the Valgrind memory leak check tool. Add the option --keep-debuginfo=yes to the Valgrind command line to suppress these errors.

    {
    Memory leak errors with dlopen.
    Memcheck:Leak
    match-leak-kinds: definite
    ...
    fun:*dlopen*
    ...
}
{
    Memory leak errors with nvrtc
    Memcheck:Leak
    match-leak-kinds: definite
    fun:malloc
    obj:*libnvrtc.so*
    ...
}

  • SM 7.5 and earlier devices may not have INT8 implementations for all layers with Q/DQ nodes. In this case, you will encounter a could not find any implementation error while building your engine. To resolve this, remove the Q/DQ nodes, which quantize the failing layers.

  • Installing the cuda-compat-11-4 package may interfere with CUDA-enhanced compatibility and cause TensorRT to fail even when the driver is r465. The workaround is to remove the cuda-compat-11-4 package or upgrade the driver to r470.

  • For some networks, using a batch size of 4096 may cause accuracy degradation on DLA.

  • For broadcasting elementwise layers running on DLA with GPU fallback enabled with one NxCxHxW input and one Nx1x1x1 input, there is a known accuracy issue if at least one of the inputs is consumed in kDLA_LINEAR format. It is recommended to explicitly set the input formats of such elementwise layers to different tensor formats.

  • Exclusive padding with kAVERAGE pooling is not supported.

  • Asynchronous CUDA calls are not supported in the user-defined processDebugTensor function for the debug tensor feature due to a bug in Windows 10.

  • inplace_add mini-sample of the quickly_deployable_plugins Python sample may produce incorrect outputs on Windows. This will be fixed in a future release.

  • Plugin CustomQKVToContextPluginDynamic, version 4 (bertQKVToContextPlugin, version 4) raises an internal error if either the batch or sequence dimensions differ at runtime from the ones used to serialize the engine. As a workaround, use the deprecated version 1 of the same plugin (which is identical in I/O and attributes). If using demoBERT, the engine builder scripts will use the bertQKVToContextPlugin version 1 by default.

Performance

  • A performance regression is expected for TensorRT 10.x with respect to TensorRT 8.6 for networks with operations that involve data-dependent shapes, such as non-max suppression or non-zero operations. The amount of regression is roughly proportional to the number of such layers in the network.

  • Up to 22% context memory size regression for HiFi-GAN networks in INT8 precision compared to TensorRT 10.5 on Ampere GPUs.

  • Up to 10% performance regression for BERT networks exported from TensorFlow2 in FP16 precision compared to TensorRT 10.4 for BS1 and Seq128 on A16 GPUs.

  • Up to 16% regression in context memory usage for StableDifussion XL VAE network in FP8 precision on H100 GPUs compared to TensorRT 10.3 due to a necessary functional fix.

  • Up to 15% regressing in context memory usage for networks containing InstanceNorm and Activation ops compared to TensorRT 10.0.

  • Up to 15% CPU memory usage regression for mbart-cnn/mamba-370m in FP16 precision and OOTB mode on NVIDIA Ada Lovelace GPUs compared to TensorRT 10.2.

  • Up to 6% performance regression for BERT/Megatron networks in FP16 precision compared to TensorRT 10.2 for BS1 and Seq128 on H100 GPUs.

  • Up to 6% performance regression for Bidirectional LSTM in FP16 precision on H100 GPUs compared to TensorRT 10.2.

  • Up to 25% performance regression when running TensorRT-LLM without the attention plugin. The current recommendation is to always enable the attention plugin when using TensorRT-LLM.

  • There are known performance gaps between engines built with REFIT enabled and engines built with REFIT disabled.

  • Up to 60 MB engine size fluctuations for the BERT-Large INT8-QDQ model on Orin due to unstable tactic selection among tactics.

  • Up to 16% performance regression for BasicUNet, DynUNet, and HighResNet in INT8 precision compared to TensorRT 9.3.

  • Up to 40-second increase in engine building for BART networks on NVIDIA Hopper GPUs.

  • Up to 20-second increase in engine building for some large language models (LLMs) on NVIDIA Ampere GPUs.

  • Up to 2.5x build time increase compared to TensorRT 9.0 for certain Bert-like models due to additional tactics available for evaluation.

  • Up to 13% performance drop for the CortanaASR model on NVIDIA Ampere GPUs compared to TensorRT 8.5.

  • Up to 18% performance drop for the ShuffleNet model on A30/A40 compared to TensorRT 8.5.1.

  • Convolution on a tensor with an implicitly data-dependent shape may run significantly slower than on other tensors of the same size. Refer to the Glossary for the definition of implicitly data-dependent shapes.

  • Up to 5% performance drop for networks using sparsity in FP16 precision.

  • Up to 6% performance regression compared to TensorRT 8.5 on OpenRoadNet in FP16 precision on NVIDIA A10 GPUs.

  • Up to 70% performance regression compared to TensorRT 8.6 on BERT networks in INT8 precision with FP16 disabled on L4 GPUs. Enable FP16 and disable INT8 in the builder config to work around this.

  • In explicitly quantized networks, a group convolution with a Q/DQ pair before but no Q/DQ pair after is expected to run with INT8-IN-FP32-OUT mixed precision. However, NVIDIA Hopper may fall back to FP32-IN-FP32-OUT if the input channel count is small.

  • The kREFIT and kREFIT_IDENTICAL have performance regressions compared with non-refit engines where convolution layers are present within a branch or loop, and the precision is FP16/INT8. This issue will be addressed in future releases.