Triton Inference Server Release 21.07
The Triton Inference Server container image, release 21.07, is available on NGC and is open source on GitHub.
Contents of the Triton Inference Server container
The Triton Inference Server Docker image contains the inference server executable and related shared libraries in /opt/tritonserver.
- Ubuntu 20.04 including Python 3.8
- NVIDIA CUDA 11.4.0
- cuBLAS 11.5.2.43
- NVIDIA cuDNN 8.2.2.6
- NVIDIA NCCL 2.10.3 (optimized for NVLink™)
- rdma-core 32.1
- OpenMPI 4.1.1rc1
- OpenUCX 1.10.1
- GDRCopy 2.2
- NVIDIA HPC-X 2.8.2rc3
- TensorRT 8.0.1.6
Driver Requirements
Release 21.07 is based on NVIDIA CUDA 11.4.0, which requires NVIDIA Driver release 470 or later. However, if you are running on Data Center GPUs (formerly Tesla), for example, T4, you may use NVIDIA driver release 418.40 (or later R418), 440.33 (or later R440), 450.51 (or later R450), or 460.27 (or later R460). The CUDA driver's compatibility package only supports particular drivers. For a complete list of supported drivers, see the CUDA Application Compatibility topic. For more information, see CUDA Compatibility and Upgrades and NVIDIA CUDA and Drivers Support.
GPU Requirements
Release 21.07 supports CUDA compute capability 6.0 and higher. This corresponds to GPUs in the NVIDIA Pascal, Volta, Turing, and Ampere Architecture GPU families. Specifically, for a list of GPUs that this compute capability corresponds to, see CUDA GPUs. For additional support details, see Deep Learning Frameworks Support Matrix.
Key Features and Enhancements
This Inference Server release includes the following key features and enhancements.
- Added support for CPU in RAPIDS FIL Backend.
- Inference requests using the C API are now allowed to provide multiple copies of an input tensor in different memories. Triton will choose the most performant copy to use depending on where the inference request is executed.
- For ONNX models using TensorRT acceleration, the tensorrt_accelerator option in the model configuration can now specify precision and workspace size.https://github.com/triton-inference-server/server/blob/main/docs/user_guide/optimization.md#framework-specific-optimization.
- Model Analyzer added an offline mode, which prioritizes throughput over latency for offline inferencing scenarios. A new set of reports and graphs are created to better analyze the offline use case.
- Refer to the 21.07 column of the Frameworks Support Matrix for container image versions that the 21.07 inference server container is based on.
- Ubuntu 20.04 with June 2021 updates.
NVIDIA Triton Inference Server Container Versions
The following table shows what versions of Ubuntu, CUDA, Triton Inference Server, and TensorRT are supported in each of the NVIDIA containers for Triton Inference Server. For older container versions, refer to the Frameworks Support Matrix.
Known Issues
- Triton’s TensorRT support depends on the input-consumed feature of TensorRT. In some rare cases using TensorRT 8.0 and earlier versions, the input-consumed event fires earlier than expected, causing Triton to overwrite input tensors while they are still in use and leading to corrupt input data being used for inference. This situation occurs when the inputs feed directly into a TensorRT layer that is optimized into a ForeignNode in the builder log. If you encounter accuracy issues with your TensorRT model, you can work around the issue by enabling the output_copy_stream option in your model’s configuration (https://github.com/triton-inference-server/common/blob/main/protobuf/model_config.proto#L816).
- The 21.07 release includes libsystemd and libudev versions that have a known vulnerability that was discovered late in our QA process. See CVE-2021-33910 for details. This will be fixed in the next release.
- ONNX Runtime TRT support was removed due to incompatibility with TensorRT 8.0.
- There is a known issue in TensorRT 8.0 regarding accuracy for a certain case of int8 inferencing on A40 and similar GPUs. The version of TF-TRT in TF2 21.07 includes a feature that works around this issue, but TF1 21.07 does not include that feature and therefore Triton users may experience the accuracy drop for a small subset of model/data type/batch size combinations on A40 when TF-TRT is used through the TF1 backend. This will be fixed in the next version of TensorRT.
- Running a PyTorch TorchScript model using the PyTorch backend, where multiple instances of a model are configured can lead to a slowdown in model execution due to the following PyTorch issue: https://github.com/pytorch/pytorch/issues/27902.
- Compared with the 21.02 and earlier releases, there are backwards incompatible changes in the example Python client shared-memory support library when that library is used for tensors of type BYTES. The utils.serialize_byte_tensor() and utils.deserialize_byte_tensor() functions now return np.object_ numpy arrays where previously they returned np.bytes_ numpy arrays. Code depending on np.bytes_ must be updated. This change was necessary because the np.bytes_ type removes all trailing zeros from each array element and so binary sequences ending in zero(s) could not be represented with the old behavior. Correct usage of the Python client shared-memory support library is shown inhttps://github.com/triton-inference-server/server/blob/r21.03/src/clients/python/examples/simple_http_shm_string_client.py.