Triton Inference Server Release 23.10
The Triton Inference Server container image, release 23.10, 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.
For the list of what the container includes, refer to Deep Learning Frameworks Support Matrix.
Driver Requirements
Release 23.10 is based on CUDA 12.2.2, which requires NVIDIA Driver release 535 or later. However, if you are running on a data center GPU (for example, T4 or any other data center GPU), you can use NVIDIA driver release 450.51 (or later R450), 470.57 (or later R470), 510.47 (or later R510), or 525.85 (or later R525), or 535.86 (or later R535).
The CUDA driver's compatibility package only supports particular drivers. Thus, users should upgrade from all R418, R440, R460, and R520 drivers, which are not forward-compatible with CUDA 12.2. For a complete list of supported drivers, see the CUDA Application Compatibility topic. For more information, see CUDA Compatibility and Upgrades.GPU Requirements
Release 23.10 supports CUDA compute capability 6.0 and later. This corresponds to GPUs in the NVIDIA Pascal, NVIDIA Volta™, NVIDIA Turing™, NVIDIA Ampere architecture, NVIDIA Hopper™, and NVIDIA Ada Lovelace architecture families. For a list of GPUs to which this compute capability corresponds, 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 handling client-side request cancellation in Triton server and backends. (server docs, client docs).
- Triton can deploy supported models on the vLLM engine using the new vLLM backend. A new container with vLLM backend is available on NGC for 23.10.
- Triton now supports the TensorRT-LLM backend. This backend uses the Nvidia TensorRT-LLM, which replaces the Fastertransformer backend. A new container with TensorRT-LLM backend is available on NGC for 23.10.
- AddedGenerate extension (beta) which provides better REST APIs for inference on Large Language Models.
- New tutorials with respect to how to run vLLM with the new REST API, how to run Llama2 with TensorRT-LLM backend, and how to run with HuggingFace models in the tutorial repo.
- Support Scalar I/O in ONNXRuntime backend.
- Added support for writing custom backends in python, a.k.a. Python-based backends.
- Refer to the 23.10 column of the Frameworks Support Matrix for container image versions on which the 23.10 inference server container is based.
NVIDIA Triton Inference Server Container Versions
The following table shows what versions of Ubuntu, CUDA, Triton Inference Server, and NVIDIA 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
- For its initial release, the TensorRT-LLM backend provides limited support of Triton extensions and features.
- The TensorRT-LLM backend may core dump on server shutdown. This impacts server teardown only and will not impact inferencing.
- When a model uses a backend which is not found, Triton would reference the missing backend as `backend_name /model.py” in the error message. This is already fixed for future releases.
- When using decoupled models, there is a possibility that response order as sent from the backend may not match with the order in which these responses are received by the streaming gRPC client. Note that this only applies to responses from different requests. Any responses corresponding to the same request will still be received in their expected order, relative to each other.
- The FasterTransformer backend is only officially supported for 22.12, though it can be built for Triton container versions up to 23.07.
- The Java CAPI is known to have intermittent segfaults.
- Some systems which implement malloc() may not release memory back to the operating system right away causing a false memory leak. This can be mitigated by using a different malloc implementation. Tcmalloc and jemalloc are installed in the Triton container and can be used by specifying the library in LD_PRELOAD. We recommend experimenting with both tcmalloc and jemalloc to determine which one works better for your use case.
- Auto-complete may cause an increase in server start time. To avoid a start time increase, users can provide the full model configuration and launch the server with --disable-auto-complete-config.
- Auto-complete does not support PyTorch models due to lack of metadata in the model. It can only verify that the number of inputs and the input names matches what is specified in the model configuration. There is no model metadata about the number of outputs and datatypes. Related PyTorch bug:https://github.com/pytorch/pytorch/issues/38273
- Triton Client PIP wheels for Arm SBSA are not available
from PyPI and pip will install an incorrect Jetson
version of Triton Client library for Arm SBSA.
The correct client wheel file can be pulled directly from the Arm SBSA SDK image and manually installed.
- Traced models in PyTorch seem to create overflows when
int8 tensor values are transformed to int32 on the
GPU.
Refer to https://github.com/pytorch/pytorch/issues/66930 for more information.
- Triton cannot retrieve GPU metrics with MIG-enabled GPU devices (A100 and A30).
- Triton metrics might not work if the host machine is running a separate DCGM agent on bare-metal or in a container.
- When cloud storage (AWS, GCS, AZURE) is used as a model repository and a model has multiple versions, Triton creates an extra local copy of the cloud model’s folder in the temporary directory, which is deleted upon server’s shutdown.