• Fixed an issue with Triton Server’s Sagemaker Service which could result in a server crash resulting from a race condition, caused by unprotected access to the list of models.

• Extended the set of accelerated PyTorch libraries included with the Triton PyTorch backend.

• Upgraded Triton Client’s Golang dependencies to latest stable versions to resolve known issues with the previous version of the dependencies.

• The OpenAI-compatible frontend has transitioned from beta to a stable release.

• Added echo request parameter for TensorRT-LLM and Python backends to OpenAI-compatible API frontend v1/completions endpoint.

• Enabled OpenAI-compatible API frontend multi-LoRA support for TensorRT-LLM backend.

• Backends can now implement the new TRITONBACKEND_ModelInstanceReady function to report accurate model readiness status.

• Updated the Python backend to accurately report model readiness.

• The error '__init__(): incompatible function arguments may occur when using TensorRT-LLM backend python models. To suppress the error temporarily, set input tensor stream with a boolean value explicitly in the request header.

• Since 25.10, vLLM backend uses V1 engine by default. You might see invalid characters in logprobs output and the bug has been reported to the vLLM team.

• PyTorch backend supports PyTorch 2.0 with the limitation that models must be provided as a serialized model file (aka ‘model.pt’). Please see Triton PyTorch Backend documentation for details.

• vLLM's v0 API and Ray are affected by vulnerabilities. Users should consider their own architecture and mitigation steps which may include but should not be limited to:

  • Do not expose Ray executors and vLLM hosts to a network where any untrusted connections might reach the host.
  • Ensure that only the other vLLM hosts are able to connect to the TCP port used for the XPUB socket. Note that the port used is random.

• Perf Analyzer is no longer part of the “client” released archive and can be installed separately using pip install perf-analyzer.

• When using Valgrind or other leak detection tools on AGX-Thor or DGX-Spark systems, you might see memory leaks attributed to NvRmGpuLibOpen. The root cause has been identified and fixed in CUDA.

• Valgrind or other memory leak detection tools may occasionally report leaks related to DCGM. These reports are intermittent and often disappear on retry. The root cause is under investigation.

• CuPy has issues with the CUDA 13 Device API in multithreaded contexts. Avoid using tritonclient cuda_shared_memory APIs in multithreaded environments until fixed by CuPy.

• TensorRT calibration cache may require size adjustment in some cases, which was observed for the IGX platform.

• The core Python binding may incur an additional D2H and H2D copy if the backend and frontend both specify device memory to be used for response tensors.

• A segmentation fault related to DCGM and NSCQ may be encountered during server shutdown on NVSwitch systems. A possible workaround for this issue is to disable the collection of GPU metrics tritonserver --allow-gpu-metrics false ...

• When using TensorRT models, if auto-complete configuration is disabled and is_non_linear_format_io:true for reformat-free tensors is not provided in the model configuration, the model may not load successfully.

• When using Python models in decoupled mode, users need to ensure that the ResponseSender goes out of scope or is properly cleaned up before unloading the model to guarantee that the unloading process executes correctly.

• Triton Inference Server with vLLM backend currently does not support running vLLM models with tensor parallelism sizes greater than 1 and the default "distributed_executor_backend" setting when using explicit model control mode. In attempt to load a vllm model (tp > 1) in explicit mode, users could potentially see failure at initialize step: could not acquire lock for <_io.BufferedWriter name='<stdout>'> at interpreter shutdown, possibly due to daemon threads. For the default model control mode, after server shutdown, vllm related sub-processes are not killed. Related vllm issue: vllm-project/vllm#6766 . Please specify "distributed_executor_backend":"ray" in the model.json when deploying vllm models with tensor parallelism > 1.

• When loading models with file override, multiple model configuration files are not supported. Users must provide the model configuration by setting parameter "config" : "<JSON>" instead of custom configuration file in the following format: "file:configs/<model-config-name>.pbtxt" : "<base64-encoded-file-content>".

• 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.

• 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. NVIDIA recommends 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 pytorch/pytorch#66930 for more information.

• Triton cannot retrieve GPU metrics with MIG-enabled GPU devices.

• 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.

• Python backend support for Windows is limited and does not currently support the following features:

  • GPU tensors
  • CPU and GPU-related metri...

• Enabled endpoint v1/embeddings for vLLM backend in OpenAI-compatible API server.

• Enabled echo parameter for TensorRT-LLM and Python backends in OpenAI-compatible API server.

• Improved error handling in OpenAI-compatible API server by providing more specific and OpenAI-compliant error codes.

• Upgraded the version of starlette used by OpenAI frontend.

• Since 25.10, vLLM backend uses V1 engine by default. You might see invalid characters in logprobs output and the bug has been reported to the vLLM team.

• PyTorch backend supports PyTorch 2.0 with the limitation that models must be provided as a serialized model file (aka ‘model.pt’). Please see Triton PyTorch Backend documentation for details.

• vLLM's v0 API and Ray are affected by vulnerabilities. Users should consider their own architecture and mitigation steps which may include but should not be limited to:

  • Do not expose Ray executors and vLLM hosts to a network where any untrusted connections might reach the host.
  • Ensure that only the other vLLM hosts are able to connect to the TCP port used for the XPUB socket. Note that the port used is random.

• Perf Analyzer is no longer part of the “client” released archive and can be installed separately using pip install perf-analyzer.

• When using Valgrind or other leak detection tools on AGX-Thor or DGX-Spark systems, you might see memory leaks attributed to NvRmGpuLibOpen. The root cause has been identified and fixed in CUDA.

• Valgrind or other memory leak detection tools may occasionally report leaks related to DCGM. These reports are intermittent and often disappear on retry. The root cause is under investigation.

• CuPy has issues with the CUDA 13 Device API in multithreaded contexts. Avoid using tritonclient cuda_shared_memory APIs in multithreaded environments until fixed by CuPy.

• TensorRT calibration cache may require size adjustment in some cases, which was observed for the IGX platform.

• The core Python binding may incur an additional D2H and H2D copy if the backend and frontend both specify device memory to be used for response tensors.

• A segmentation fault related to DCGM and NSCQ may be encountered during server shutdown on NVSwitch systems. A possible workaround for this issue is to disable the collection of GPU metrics tritonserver --allow-gpu-metrics false ...

• When using TensorRT models, if auto-complete configuration is disabled and is_non_linear_format_io:true for reformat-free tensors is not provided in the model configuration, the model may not load successfully.

• When using Python models in decoupled mode, users need to ensure that the ResponseSender goes out of scope or is properly cleaned up before unloading the model to guarantee that the unloading process executes correctly.

• Triton Inference Server with vLLM backend currently does not support running vLLM models with tensor parallelism sizes greater than 1 and the default "distributed_executor_backend" setting when using explicit model control mode. In attempt to load a vllm model (tp > 1) in explicit mode, users could potentially see failure at initialize step: could not acquire lock for <_io.BufferedWriter name='<stdout>'> at interpreter shutdown, possibly due to daemon threads. For the default model control mode, after server shutdown, vllm related sub-processes are not killed. Related vllm issue: vllm-project/vllm#6766 . Please specify "distributed_executor_backend":"ray" in the model.json when deploying vllm models with tensor parallelism > 1.

• When loading models with file override, multiple model configuration files are not supported. Users must provide the model configuration by setting parameter "config" : "<JSON>" instead of custom configuration file in the following format: "file:configs/<model-config-name>.pbtxt" : "<base64-encoded-file-content>".

• 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.

• 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. NVIDIA recommends 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 pytorch/pytorch#66930 for more information.

• Triton cannot retrieve GPU metrics with MIG-enabled GPU devices.

• 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.

• Python backend support for Windows is limited and does not currently support the following features:

  • GPU tensors
  • CPU and GPU-related metrics
  • Custom execution environments
  • The model load/unload APIs

Ubuntu 24.04 builds of the client libraries and examples are included in this release in the attached v2.63.0_ubuntu2404.clients.tar.gz file. The SDK is also available for as an Ubuntu 24.04 based NGC Container. The SDK container includes the client libraries and examples, Performance Analyzer and Model Analyzer. Some components are also available in the tritonclient pip package. See Getting the Client Libraries for more information on each of these options.

A release of Triton for [IGX](https://www.nvidia.com/en-us/edge-computing/p...

• Fixed a server crash issue caused by specially crafted request messages sent to /v2/models/<model_name>/infer.

• Fixed a server crash issue caused by incorrect handling of malformed HTTP requests.

• Triton python package uses outdated dependency starlette package version.

• Since 25.10, vLLM backend uses V1 engine by default. You might see invalid characters in logprobs output and the bug has been reported to the vLLM team.

• Enabling vLLM metrics during inferences causes the engine to crash.

• PyTorch backend supports PyTorch 2.0 with the limitation that models must be provided as a serialized model file (aka ‘model.pt’). Please see Triton PyTorch Backend documentation for details.

• vLLM's v0 API and Ray are affected by vulnerabilities. Users should consider their own architecture and mitigation steps which may include but should not be limited to:

  • Do not expose Ray executors and vLLM hosts to a network where any untrusted connections might reach the host.
  • Ensure that only the other vLLM hosts are able to connect to the TCP port used for the XPUB socket. Note that the port used is random.

• Perf Analyzer is no longer part of the “client” released archive and can be installed separately using pip install perf-analyzer.

• When using Valgrind or other leak detection tools on AGX-Thor or DGX-Spark systems, you might see memory leaks attributed to NvRmGpuLibOpen. The root cause has been identified and fixed in CUDA.

• Valgrind or other memory leak detection tools may occasionally report leaks related to DCGM. These reports are intermittent and often disappear on retry. The root cause is under investigation.

• CuPy has issues with the CUDA 13 Device API in multithreaded contexts. Avoid using tritonclient cuda_shared_memory APIs in multithreaded environments until fixed by CuPy.

• TensorRT calibration cache may require size adjustment in some cases, which was observed for the IGX platform.

• The core Python binding may incur an additional D2H and H2D copy if the backend and frontend both specify device memory to be used for response tensors.

• A segmentation fault related to DCGM and NSCQ may be encountered during server shutdown on NVSwitch systems. A possible workaround for this issue is to disable the collection of GPU metrics tritonserver --allow-gpu-metrics false ...

• When using TensorRT models, if auto-complete configuration is disabled and is_non_linear_format_io:true for reformat-free tensors is not provided in the model configuration, the model may not load successfully.

• When using Python models in decoupled mode, users need to ensure that the ResponseSender goes out of scope or is properly cleaned up before unloading the model to guarantee that the unloading process executes correctly.

• Triton Inference Server with vLLM backend currently does not support running vLLM models with tensor parallelism sizes greater than 1 and the default "distributed_executor_backend" setting when using explicit model control mode. In attempt to load a vllm model (tp > 1) in explicit mode, users could potentially see failure at initialize step: could not acquire lock for <_io.BufferedWriter name='<stdout>'> at interpreter shutdown, possibly due to daemon threads. For the default model control mode, after server shutdown, vllm related sub-processes are not killed. Related vllm issue: vllm-project/vllm#6766 . Please specify "distributed_executor_backend":"ray" in the model.json when deploying vllm models with tensor parallelism > 1.

• When loading models with file override, multiple model configuration files are not supported. Users must provide the model configuration by setting parameter "config" : "<JSON>" instead of custom configuration file in the following format: "file:configs/<model-config-name>.pbtxt" : "<base64-encoded-file-content>".

• 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.

• 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. NVIDIA recommends 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 pytorch/pytorch#66930 for more information.

• Triton cannot retrieve GPU metrics with MIG-enabled GPU devices.

• 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.

• Python backend support for Windows is limited and does not currently support the following features:

  • GPU tensors
  • CPU and GPU-related metrics
  • Custom execution environments
  • The model load/unload APIs

Ubuntu 24.04 builds of the client libraries and examples are included in this release in the attached v2.62.0_ubuntu2404.clients.tar.gz file. The SDK is also available for as an Ubuntu 24.04 based NGC Container. The SDK container includes the client libraries and examples, Performance Analyzer and Model Analyzer. Some components are also available in the tritonclient pip package. See Getting the Client Libraries for more information on each of these options.

[!NOTE]
There is no Windows release for 25.08, the latest release is 25.01.

A release of Triton for IGX is provided in the attached tar file: tritonserver2.62.0-igpu.tar.

• This release supports CUDA 12.9, TensorRT 10.11.0.33, Onnx Runtime 1.23.1, PyTorch 2.9.0a0+145a3a7, Python 3.12 and as well as ensembles.
• ONNX Runtime backend does not support the OpenVINO and TensorRT execution providers. The CUDA execution provider is in Beta.
• System shared memory is supported on Jetson. CUDA shared memory is not supported.
• GPU metrics, GCS storage, S3 storage and Azure storage are not supported.

The tar file contains the Triton server executable and shared libraries and also the C++ and Python client libraries and examples. For more information on how to install and use Triton on JetPack refer to jetson.md.

The wheel for the Python client library is present in the tar file and can be installed by running the following command:

python3 -m pip install --upgrade clients/python/tritonclient-2.62.0-py3-none-manylinux2014_aarch64.whl[all]

• Static key authentication for OpenAI Frontend APIs

• Prevented models outside Triton’s repository being loaded from OpenAI Frontend.

• vLLM's v0 API and Ray are affected by vulnerabilities. Users should consider their own architecture and mitigation steps which may include but should not be limited to:

  • Do not expose Ray executors and vLLM hosts to a network where any untrusted connections might reach the host.
  • Ensure that only the other vLLM hosts are able to connect to the TCP port used for the XPUB socket. Note that the port used is random.

• Perf Analyzer is no longer part of the “client” released archive and can be installed separately using pip install perf-analyzer.

• When using Valgrind or other leak detection tools on AGX-Thor or DGX-Spark systems, you might see memory leaks attributed to NvRmGpuLibOpen. The root cause has been identified and fixed in CUDA.

• Valgrind or other memory leak detection tools may occasionally report leaks related to DCGM. These reports are intermittent and often disappear on retry. The root cause is under investigation.

• CuPy has issues with the CUDA 13 Device API in multithreaded contexts. Avoid using tritonclient cuda_shared_memory APIs in multithreaded environments until fixed by CuPy.

• TensorRT calibration cache may require size adjustment in some cases, which was observed for the IGX platform.

• The core Python binding may incur an additional D2H and H2D copy if the backend and frontend both specify device memory to be used for response tensors.

• A segmentation fault related to DCGM and NSCQ may be encountered during server shutdown on NVSwitch systems. A possible workaround for this issue is to disable the collection of GPU metrics tritonserver --allow-gpu-metrics false ...

• When using TensorRT models, if auto-complete configuration is disabled and is_non_linear_format_io:true for reformat-free tensors is not provided in the model configuration, the model may not load successfully.

• When using Python models in decoupled mode, users need to ensure that the ResponseSender goes out of scope or is properly cleaned up before unloading the model to guarantee that the unloading process executes correctly.

• Triton Inference Server with vLLM backend currently does not support running vLLM models with tensor parallelism sizes greater than 1 and the default "distributed_executor_backend" setting when using explicit model control mode. In attempt to load a vllm model (tp > 1) in explicit mode, users could potentially see failure at initialize step: could not acquire lock for <_io.BufferedWriter name='<stdout>'> at interpreter shutdown, possibly due to daemon threads. For the default model control mode, after server shutdown, vllm related sub-processes are not killed. Related vllm issue: vllm-project/vllm#6766 . Please specify "distributed_executor_backend":"ray" in the model.json when deploying vllm models with tensor parallelism > 1.

• When loading models with file override, multiple model configuration files are not supported. Users must provide the model configuration by setting parameter "config" : "<JSON>" instead of custom configuration file in the following format: "file:configs/<model-config-name>.pbtxt" : "<base64-encoded-file-content>".

• 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.

• 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. NVIDIA recommends 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 pytorch/pytorch#66930 for more information.

• Triton cannot retrieve GPU metrics with MIG-enabled GPU devices.

• 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.

• Python backend support for Windows is limited and does not currently support the following features:

  • GPU tensors
  • CPU and GPU-related metrics
  • Custom execution environments
  • The model load/unload APIs

Ubuntu 24.04 builds of the client libraries and examples are included in this release in the attached v2.61.0_ubuntu2404.clients.tar.gz file. The SDK is also available for as an Ubuntu 24.04 based NGC Container. The SDK container includes the client libraries and examples, Performance Analyzer and Model Analyzer. Some components are also available in the tritonclient pip package. See Getting the Client Libraries for more information on each of these options.

[!NOTE]
There is no Windows release for 25.08, the latest release is 25.01.

A release of Triton for IGX is provided in the attached tar file: tritonserver2.61.0-igpu.tar.

• This release supports CUDA 12.9, TensorRT 10.11.0.33, Onnx Runtime 1.23.0, PyTorch 2.9.0a0+50eac81, Python 3.12 and as well as ensembles.
• ONNX Runtime backend does not support the OpenVINO and TensorRT execution providers. The CUDA execution provider is in Beta.
• System shared memory is supported on Jetson. CUDA shared memory is not supported.
• GPU metrics, GCS storage, S3 storage and Azure storage are not supported.

The tar file contains the Triton server executable and shared libraries and also the C++ and Python client libraries and examples. For more information on how to install and use Triton on JetPack refer to jetson.md.

The wheel for the Python client library is present in the tar file and can be installed by running the following command:

python3 -m pip install --upgrade clients/python/tritonclient-2.61.0-py3-none-manylinux2014_aarch64.whl[all]

A release of Triton for AGX Systems is provided in the attached tar file: tritonserver2.61.0-agx.tar.

• This release supports CUDA 13.0, TensorRT 10.13.3.9, Onnx Runtime 1.23.0, PyTorch [2.9.0a0+50eac81](https://docs.nvidia.com/deeplearning/frameworks/install-pytorch-jetson-platform-release-notes/pytorch-jetson-rel.html...

• Added CUDA 13 support.

• Triton ONNX Runtime Backend build uses microsoft/onnxruntime/commit/1d1712fdaf and may have some limitations on DGX Spark hardware which will be addressed in future versions.

• CuPy has issues with the CUDA 13 Device API in multithreaded contexts. Avoid using tritonclient cuda_shared_memory APIs in multithreaded environments until fixed by CuPy

• CuPy does not support CUDA 13 at the time of writing. Issues may be encountered when using CuPy before it officially supports CUDA 13, see https://github.com/triton-inference-server/server/tree/r25.08/python/openai#pre-requisites for more details

• TensorRT calibration cache may require size adjustment in some cases, which was observed for the IGX platform.

• The core Python binding may incur an additional D2H and H2D copy if the backend and frontend both specify device memory to be used for response tensors.

• A segmentation fault related to DCGM and NSCQ may be encountered during server shutdown on NVSwitch systems. A possible workaround for this issue is to disable the collection of GPU metrics tritonserver --allow-gpu-metrics false ...

• When using TensorRT models, if auto-complete configuration is disabled and is_non_linear_format_io:true for reformat-free tensors is not provided in the model configuration, the model may not load successfully.

• When using Python models in decoupled mode, users need to ensure that the ResponseSender goes out of scope or is properly cleaned up before unloading the model to guarantee that the unloading process executes correctly.

• Restart support was temporarily removed for Python models.

• Triton Inference Server with vLLM backend currently does not support running vLLM models with tensor parallelism sizes greater than 1 and the default "distributed_executor_backend" setting when using explicit model control mode. In attempt to load a vllm model (tp > 1) in explicit mode, users could potentially see failure at initialize step: could not acquire lock for <_io.BufferedWriter name='<stdout>'> at interpreter shutdown, possibly due to daemon threads. For the default model control mode, after server shutdown, vllm related sub-processes are not killed. Related vllm issue: vllm-project/vllm#6766 . Please specify "distributed_executor_backend":"ray" in the model.json when deploying vllm models with tensor parallelism > 1.

• When loading models with file override, multiple model configuration files are not supported. Users must provide the model configuration by setting parameter "config" : "<JSON>" instead of custom configuration file in the following format: "file:configs/<model-config-name>.pbtxt" : "<base64-encoded-file-content>".

• 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.

• 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. NVIDIA recommends 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 pytorch/pytorch#66930 for more information.

• Triton cannot retrieve GPU metrics with MIG-enabled GPU devices.

• 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.

• Python backend support for Windows is limited and does not currently support the following features:

  • GPU tensors
  • CPU and GPU-related metrics
  • Custom execution environments
  • The model load/unload APIs

Ubuntu 24.04 builds of the client libraries and examples are included in this release in the attached v2.60.0_ubuntu2404.clients.tar.gz file. The SDK is also available for as an Ubuntu 24.04 based NGC Container. The SDK container includes the client libraries and examples, Performance Analyzer and Model Analyzer. Some components are also available in the tritonclient pip package. See Getting the Client Libraries for more information on each of these options.

[!NOTE]
There is no Windows release for 25.08, the latest release is 25.01.

A release of Triton for IGX is provided in the attached tar file: tritonserver2.60.0-igpu.tar.

• This release supports CUDA 12.9, TensorRT 10.11.0.33, Onnx Runtime 1.23.0+1d1712fdaf, PyTorch 2.8.0a+34c6371d24, Python 3.12 and as well as ensembles.
• ONNX Runtime backend does not support the OpenVINO and TensorRT execution providers. The CUDA execution provider is in Beta.
• System shared memory is supported on Jetson. CUDA shared memory is not supported.
• GPU metrics, GCS storage, S3 storage and Azure storage are not supported.

The tar file contains the Triton server executable and shared libraries and also the C++ and Python client libraries and examples. For more information on how to install and use Triton on JetPack refer to jetson.md.

The wheel for the Python client library is present in the tar file and can be installed by running the following command:

python3 -m pip install --upgrade clients/python/tritonclient-2.59.0-py3-none-manylinux2014_aarch64.whl[all]

A release of Triton for AGX Systems is provided in the attached tar file: tritonserver2.60.0-agx.tar.

[!NOTE]
There is no Jetson AGX release for 25.08, requires DCGM version 4 to be installed in order to use GPU metrics.
Please use following command to install DCGM 4:

curl -o /tmp/cuda-keyring.deb \
         https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2404/x86_64/cuda-keyring_1.1-1_all.deb \
     && apt install /tmp/cuda-keyring.deb \
     && rm /tmp/cuda-keyring.deb \
     && apt update \
     && apt install --yes --no-install-recommends \
                  datacenter-gpu-manager-4-core=1:4.4.0-1

• This release supports CUDA 13.0, TensorRT 10.13.2.6, Onnx Runtime 1.23.0+1d1712fdaf, PyTorch [2.8.0a0+34c6371](https://docs.n...

• Fixed vulnerabilities in the Triton Inference Server.

• There was no python wheels packages released as part of 25.07 release

• TensorRT calibration cache may require size adjustment in some cases, which was observed for the IGX platform.

• The core Python binding may incur an additional D2H and H2D copy if the backend and frontend both specify device memory to be used for response tensors.

• A segmentation fault related to DCGM and NSCQ may be encountered during server shutdown on NVSwitch systems. A possible workaround for this issue is to disable the collection of GPU metrics tritonserver --allow-gpu-metrics false ...

• vLLM backend currently does not take advantage of the vLLM v0.6 performance improvement when metrics are enabled.

• When using TensorRT models, if auto-complete configuration is disabled and is_non_linear_format_io:true for reformat-free tensors is not provided in the model configuration, the model may not load successfully.

• When using Python models in decoupled mode, users need to ensure that the ResponseSender goes out of scope or is properly cleaned up before unloading the model to guarantee that the unloading process executes correctly.

• Restart support was temporarily removed for Python models.

• Triton Inference Server with vLLM backend currently does not support running vLLM models with tensor parallelism sizes greater than 1 and the default "distributed_executor_backend" setting when using explicit model control mode. In attempt to load a vllm model (tp > 1) in explicit mode, users could potentially see failure at initialize step: could not acquire lock for <_io.BufferedWriter name='<stdout>'> at interpreter shutdown, possibly due to daemon threads. For the default model control mode, after server shutdown, vllm related sub-processes are not killed. Related vllm issue: vllm-project/vllm#6766 . Please specify "distributed_executor_backend":"ray" in the model.json when deploying vllm models with tensor parallelism > 1.

• When loading models with file override, multiple model configuration files are not supported. Users must provide the model configuration by setting parameter "config" : "<JSON>" instead of custom configuration file in the following format: "file:configs/<model-config-name>.pbtxt" : "<base64-encoded-file-content>".

• 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.

• 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. NVIDIA recommends 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 pytorch/pytorch#66930 for more information.

• Triton cannot retrieve GPU metrics with MIG-enabled GPU devices.

• 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.

• Python backend support for Windows is limited and does not currently support the following features:

  • GPU tensors
  • CPU and GPU-related metrics
  • Custom execution environments
  • The model load/unload APIs

Ubuntu 24.04 builds of the client libraries and examples are included in this release in the attached v2.59.1_ubuntu2404.clients.tar.gz file. The SDK is also available for as an Ubuntu 24.04 based NGC Container. The SDK container includes the client libraries and examples, Performance Analyzer and Model Analyzer. Some components are also available in the tritonclient pip package. See Getting the Client Libraries for more information on each of these options.

[!NOTE]
There is no Windows release for 25.07, the latest release is 25.01.

A release of Triton for IGX is provided in the attached tar file: tritonserver2.59.1-igpu.tar.

• This release supports TensorRT 10.11.0.33, Onnx Runtime 1.22.0, PyTorch 2.8.0a0+5228986c39.nv25.6, Python 3.12 and as well as ensembles.
• ONNX Runtime backend does not support the OpenVINO and TensorRT execution providers. The CUDA execution provider is in Beta.
• System shared memory is supported on Jetson. CUDA shared memory is not supported.
• GPU metrics, GCS storage, S3 storage and Azure storage are not supported.

The tar file contains the Triton server executable and shared libraries and also the C++ and Python client libraries and examples. For more information on how to install and use Triton on JetPack refer to jetson.md.

The wheel for the Python client library is present in the tar file and can be installed by running the following command:

python3 -m pip install --upgrade clients/python/tritonclient-2.59.0-py3-none-manylinux2014_aarch64.whl[all]

The Triton TensorRT-LLM container is built from the 25.04 image nvcr.io/nvidia/tritonserver:25.04-py3-min. Please refer to the support matrix and compatibility.md for all dependency versions related to 25.04. However, the packages listed below have different versions than those specified in the support matrix.

• Improved ensemble model performance in scenarios that allow out-of-order responses by increasing maximum throughput and reducing latency.

• TensorRT calibration cache may require size adjustment in some cases, which was observed for the IGX platform.

• The core Python binding may incur an additional D2H and H2D copy if the backend and frontend both specify device memory to be used for response tensors.

• A segmentation fault related to DCGM and NSCQ may be encountered during server shutdown on NVSwitch systems. A possible workaround for this issue is to disable the collection of GPU metrics tritonserver --allow-gpu-metrics false ...

• vLLM backend currently does not take advantage of the vLLM v0.6 performance improvement when metrics are enabled.

• When using TensorRT models, if auto-complete configuration is disabled and is_non_linear_format_io:true for reformat-free tensors is not provided in the model configuration, the model may not load successfully.

• When using Python models in decoupled mode, users need to ensure that the ResponseSender goes out of scope or is properly cleaned up before unloading the model to guarantee that the unloading process executes correctly.

• Restart support was temporarily removed for Python models.

• Triton Inference Server with vLLM backend currently does not support running vLLM models with tensor parallelism sizes greater than 1 and the default "distributed_executor_backend" setting when using explicit model control mode. In attempt to load a vllm model (tp > 1) in explicit mode, users could potentially see failure at initialize step: could not acquire lock for <_io.BufferedWriter name='<stdout>'> at interpreter shutdown, possibly due to daemon threads. For the default model control mode, after server shutdown, vllm related sub-processes are not killed. Related vllm issue: vllm-project/vllm#6766 . Please specify "distributed_executor_backend":"ray" in the model.json when deploying vllm models with tensor parallelism > 1.

• When loading models with file override, multiple model configuration files are not supported. Users must provide the model configuration by setting parameter "config" : "<JSON>" instead of custom configuration file in the following format: "file:configs/<model-config-name>.pbtxt" : "<base64-encoded-file-content>".

• 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.

• 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. NVIDIA recommends 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 pytorch/pytorch#66930 for more information.

• Triton cannot retrieve GPU metrics with MIG-enabled GPU devices.

• 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.

• Python backend support for Windows is limited and does not currently support the following features:

  • GPU tensors
  • CPU and GPU-related metrics
  • Custom execution environments
  • The model load/unload APIs

Ubuntu 24.04 builds of the client libraries and examples are included in this release in the attached v2.59.0_ubuntu2404.clients.tar.gz file. The SDK is also available for as an Ubuntu 24.04 based NGC Container. The SDK container includes the client libraries and examples, Performance Analyzer and Model Analyzer. Some components are also available in the tritonclient pip package. See Getting the Client Libraries for more information on each of these options.

[!NOTE]
There is no Windows release for 25.06, the latest release is 25.01.

A release of Triton for IGX is provided in the attached tar file: tritonserver2.59.0-igpu.tar.

• This release supports TensorRT 10.11.0.33, Onnx Runtime 1.22.0, PyTorch 2.8.0a0+5228986c39.nv25.6, Python 3.12 and as well as ensembles.
• ONNX Runtime backend does not support the OpenVINO and TensorRT execution providers. The CUDA execution provider is in Beta.
• System shared memory is supported on Jetson. CUDA shared memory is not supported.
• GPU metrics, GCS storage, S3 storage and Azure storage are not supported.

The tar file contains the Triton server executable and shared libraries and also the C++ and Python client libraries and examples. For more information on how to install and use Triton on JetPack refer to jetson.md.

The wheel for the Python client library is present in the tar file and can be installed by running the following command:

python3 -m pip install --upgrade clients/python/tritonclient-2.59.0-py3-none-manylinux2014_aarch64.whl[all]

The Triton TensorRT-LLM container is built from the 25.04 image nvcr.io/nvidia/tritonserver:25.04-py3-min. Please refer to the support matrix and compatibility.md for all dependency versions related to 25.04. However, the packages listed below have different versions than those specified in the support matrix.

• Optional “execution_context_allocation_strategy” parameter in the TensorRT backend configuration allows selection of memory allocation behavior.
• Support Tool calling functionality with Llama 3 and Mistral models in OpenAI frontend.
• Improvements around memory allocation and various bug fixes.
• GenAI-Perf now offers a new configuration file alongside the command line.
• GenAI-Perf now collects GPU metrics from /metrics endpoint exposed by DCGM Exporter.
• GenAI-Perf supports new Power, Utilization, Ecc, Errors and PCie metrics.

• vLLM backend for 25.05 might be unstable with the vLLM V1 architecture. We recommend switching to V0 for this release, by setting VLLM_USE_V1 environment variable to 0. However, users should be aware that vLLM's V0 API is affected by vulnerabilities.

• vLLM containers include vllm version 0.8.4 which is affected by vulnerabilities.
  Workarounds:
  Prior to the fix, your options include:

  • Do not expose the vLLM host to a network where any untrusted connections may reach the host.
  • Ensure that only the other vLLM hosts are able to connect to the TCP port used for the XPUB socket. Note that port used is random.

• The core Python binding may incur an additional D2H and H2D copy if the backend and frontend both specify device memory to be used for response tensors.

• A segmentation fault related to DCGM and NSCQ may be encountered during server shutdown on NVSwitch systems. A possible workaround for this issue is to disable the collection of GPU metrics tritonserver --allow-gpu-metrics false ...

• vLLM backend currently does not take advantage of the vLLM v0.6 performance improvement when metrics are enabled.

• When using TensorRT models, if auto-complete configuration is disabled and is_non_linear_format_io:true for reformat-free tensors is not provided in the model configuration, the model may not load successfully.

• When using Python models in decoupled mode, users need to ensure that the ResponseSender goes out of scope or is properly cleaned up before unloading the model to guarantee that the unloading process executes correctly.

• Restart support was temporarily removed for Python models.

• Triton Inference Server with vLLM backend currently does not support running vLLM models with tensor parallelism sizes greater than 1 and the default "distributed_executor_backend" setting when using explicit model control mode. In attempt to load a vllm model (tp > 1) in explicit mode, users could potentially see failure at initialize step: could not acquire lock for <_io.BufferedWriter name='<stdout>'> at interpreter shutdown, possibly due to daemon threads. For the default model control mode, after server shutdown, vllm related sub-processes are not killed. Related vllm issue: vllm-project/vllm#6766 . Please specify "distributed_executor_backend":"ray" in the model.json when deploying vllm models with tensor parallelism > 1.

• When loading models with file override, multiple model configuration files are not supported. Users must provide the model configuration by setting parameter "config" : "<JSON>" instead of custom configuration file in the following format: "file:configs/<model-config-name>.pbtxt" : "<base64-encoded-file-content>".

• 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.

• 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. NVIDIA recommends 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 pytorch/pytorch#66930 for more information.

• Triton cannot retrieve GPU metrics with MIG-enabled GPU devices.

• 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.

• Python backend support for Windows is limited and does not currently support the following features:

  • GPU tensors
  • CPU and GPU-related metrics
  • Custom execution environments
  • The model load/unload APIs

Ubuntu 24.04 builds of the client libraries and examples are included in this release in the attached v2.58.0_ubuntu2404.clients.tar.gz file. The SDK is also available for as an Ubuntu 24.04 based NGC Container. The SDK container includes the client libraries and examples, Performance Analyzer and Model Analyzer. Some components are also available in the tritonclient pip package. See Getting the Client Libraries for more information on each of these options.

[!NOTE]
There is no Windows release for 25.05, the latest release is 25.01.

A release of Triton for IGX is provided in the attached tar file: tritonserver2.58.0-igpu.tar.

• This release supports TensorRT 10.10.0.31, Onnx Runtime 1.22.0, PyTorch 2.8.0a0+5228986c39.nv25.5, Python 3.12 and as well as ensembles.
• ONNX Runtime backend does not support the OpenVINO and TensorRT execution providers. The CUDA execution provider is in Beta.
• System shared memory is supported on Jetson. CUDA shared memory is not supported.
• GPU metrics, GCS storage, S3 storage and Azure storage are not supported.

The tar file contains the Triton server executable and shared libraries and also the C++ and Python client libraries and examples. For more information on how to install and use Triton on JetPack refer to jetson.md.

The wheel for the Python client library is present in the tar file and can be installed by running the following command:

python3 -m pip install --upgrade clients/python/tritonclient-2.58.0-py3-none-manylinux2014_aarch64.whl[all]

The Triton TensorRT-LLM container is built from the 25.03 image nvcr.io/nvidia/tritonserver:25.03-py3-min. Please refer to the support matrix and compatibility.md for all dependency versions related to 25.03. However, the packa...

• Exposed gRPC infer thread count as a server option.
• Improved server stability during the gRPC client cancellation.
• Improved server stability in tracing mode.
• Added BLS decoupled request cancellation in the Python Backend
• GenAI-Perf now offers a new configuration file alongside the command line.
• GenAI-Perf now supports the Huggingface TGI generated endpoint.
• GenAI-Perf added a Token per second per user (TPS/user) metric.
• GenAI-Perf metric parsing speed was increased by 60%.

• vLLM backend for 25.04 might be unstable with the vLLM V1 architecture. We recommend switching to V0 for this release, by setting VLLM_USE_V1 environment variable to 0. However, users should be aware that vLLM's V0 API is affected by vulnerabilities.

• vLLM containers include vllm version 0.8.1 which is affected by new vulnerabilities.
  Workarounds:
  Prior to the fix, your options include:

  • Do not expose the vLLM host to a network where any untrusted connections may reach the host.
  • Ensure that only the other vLLM hosts are able to connect to the TCP port used for the XPUB socket. Note that port used is random.

• The core Python binding may incur an additional D2H and H2D copy if the backend and frontend both specify device memory to be used for response tensors.

• A segmentation fault related to DCGM and NSCQ may be encountered during server shutdown on NVSwitch systems. A possible workaround for this issue is to disable the collection of GPU metrics tritonserver --allow-gpu-metrics false ...

• vLLM backend currently does not take advantage of the vLLM v0.6 performance improvement when metrics are enabled.

• When using TensorRT models, if auto-complete configuration is disabled and is_non_linear_format_io:true for reformat-free tensors is not provided in the model configuration, the model may not load successfully.

• When using Python models in decoupled mode, users need to ensure that the ResponseSender goes out of scope or is properly cleaned up before unloading the model to guarantee that the unloading process executes correctly.

• Restart support was temporarily removed for Python models.

• Triton Inference Server with vLLM backend currently does not support running vLLM models with tensor parallelism sizes greater than 1 and the default "distributed_executor_backend" setting when using explicit model control mode. In attempt to load a vllm model (tp > 1) in explicit mode, users could potentially see failure at initialize step: could not acquire lock for <_io.BufferedWriter name='<stdout>'> at interpreter shutdown, possibly due to daemon threads. For the default model control mode, after server shutdown, vllm related sub-processes are not killed. Related vllm issue: vllm-project/vllm#6766 . Please specify "distributed_executor_backend":"ray" in the model.json when deploying vllm models with tensor parallelism > 1.

• When loading models with file override, multiple model configuration files are not supported. Users must provide the model configuration by setting parameter "config" : "<JSON>" instead of custom configuration file in the following format: "file:configs/<model-config-name>.pbtxt" : "<base64-encoded-file-content>".

• 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.

• 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. NVIDIA recommends 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 pytorch/pytorch#66930 for more information.

• Triton cannot retrieve GPU metrics with MIG-enabled GPU devices.

• 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.

• Python backend support for Windows is limited and does not currently support the following features:

  • GPU tensors
  • CPU and GPU-related metrics
  • Custom execution environments
  • The model load/unload APIs

Ubuntu 24.04 builds of the client libraries and examples are included in this release in the attached v2.57.0_ubuntu2404.clients.tar.gz file. The SDK is also available for as an Ubuntu 24.04 based NGC Container. The SDK container includes the client libraries and examples, Performance Analyzer and Model Analyzer. Some components are also available in the tritonclient pip package. See Getting the Client Libraries for more information on each of these options.

[!NOTE]
There is no Windows release for 25.04, the latest release is 25.01.

A release of Triton for IGX is provided in the attached tar file: tritonserver2.57.0-igpu.tar.

• This release supports TensorRT 10.9.0.34, Onnx Runtime 1.21.0, PyTorch 2.7.0a0+79aa17489c.nv25.4, Python 3.12 and as well as ensembles.
• ONNX Runtime backend does not support the OpenVINO and TensorRT execution providers. The CUDA execution provider is in Beta.
• System shared memory is supported on Jetson. CUDA shared memory is not supported.
• GPU metrics, GCS storage, S3 storage and Azure storage are not supported.

The tar file contains the Triton server executable and shared libraries and also the C++ and Python client libraries and examples. For more information on how to install and use Triton on JetPack refer to jetson.md.

The wheel for the Python client library is present in the tar file and can be installed by running the following command:

python3 -m pip install --upgrade clients/python/tritonclient-2.57.0-py3-none-manylinux2014_aarch64.whl[all]

The Triton TensorRT-LLM container is built from the 25.03 image nvcr.io/nvidia/tritonserver:25.03-py3-min. Please refer to the support matrix and compatibility.md for all dependency versions related to 25.03. However, the packages listed below have different versions...

• The Tensorflow Backend has been deprecated starting in 25.03. The last release of Triton Inference Server with the Tensorflow Backend is 25.02. Users wishing to continue using the Tensorflow Backend in 25.03 and later can build the Tensorflow Backend from source and install the result into the /opt/tritonserver/backends/ directory.
• The “XX.YY-tf2-python-py3” container will no longer be available starting in 25.03. See the Tensorflow Backend deprecation.
• Added generate and generate_stream inference types to SageMaker server. Customers can choose which inference types - infer (default), generate or generate_stream using SAGEMAKER_TRITON_INFERENCE_TYPE environment variable during server launch.
• In an effort to allow quick, on-demand metric retrieval for external load balancers such as the Kubernetes Inference Gateway API, Triton when used with TRT-LLM can include live KV-cache utilization and capacity metrics in the HTTP response header when processing inference requests.

• The core Python binding may incur an additional D2H and H2D copy if the backend and frontend both specify device memory to be used for response tensors.

• A segmentation fault related to DCGM and NSCQ may be encountered during server shutdown on NVSwitch systems. A possible workaround for this issue is to disable the collection of GPU metrics tritonserver --allow-gpu-metrics false ...

• vLLM backend currently does not take advantage of the vLLM v0.6 performance improvement when metrics are enabled.

• Incorrect results are known to occur when using TensorRT (TRT) Backend for inference using int8 data type for I/O on the Blackwell GPU architecture.

• When using TensorRT models, if auto-complete configuration is disabled and is_non_linear_format_io:true for reformat-free tensors is not provided in the model configuration, the model may not load successfully.

• When using Python models in decoupled mode, users need to ensure that the ResponseSender goes out of scope or is properly cleaned up before unloading the model to guarantee that the unloading process executes correctly.

• Restart support was temporarily removed for Python models.

• Triton Inference Server with vLLM backend currently does not support running vLLM models with tensor parallelism sizes greater than 1 and the default "distributed_executor_backend" setting when using explicit model control mode. In attempt to load a vllm model (tp > 1) in explicit mode, users could potentially see failure at initialize step: could not acquire lock for <_io.BufferedWriter name='<stdout>'> at interpreter shutdown, possibly due to daemon threads. For the default model control mode, after server shutdown, vllm related sub-processes are not killed. Related vllm issue: vllm-project/vllm#6766 . Please specify "distributed_executor_backend":"ray" in the model.json when deploying vllm models with tensor parallelism > 1.

• When loading models with file override, multiple model configuration files are not supported. Users must provide the model configuration by setting parameter "config" : "<JSON>" instead of custom configuration file in the following format: "file:configs/<model-config-name>.pbtxt" : "<base64-encoded-file-content>".

• 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.

• 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. NVIDIA recommends 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 pytorch/pytorch#66930 for more information.

• Triton cannot retrieve GPU metrics with MIG-enabled GPU devices.

• 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.

• Python backend support for Windows is limited and does not currently support the following features:

  • GPU tensors
  • CPU and GPU-related metrics
  • Custom execution environments
  • The model load/unload APIs

Ubuntu 24.04 builds of the client libraries and examples are included in this release in the attached v2.56.0_ubuntu2404.clients.tar.gz file. The SDK is also available for as an Ubuntu 24.04 based NGC Container. The SDK container includes the client libraries and examples, Performance Analyzer and Model Analyzer. Some components are also available in the tritonclient pip package. See Getting the Client Libraries for more information on each of these options.

[!NOTE]
There is no Windows release for 25.03, the latest release is 25.01.

A release of Triton for IGX is provided in the attached tar file: tritonserver2.56.0-igpu.tgz.

• This release supports TensorRT 10.9.0.34, Onnx Runtime 1.21.0, PyTorch 2.7.0a0+7c8ec84dab.nv25.3, Python 3.12 and as well as ensembles.
• ONNX Runtime backend does not support the OpenVINO and TensorRT execution providers. The CUDA execution provider is in Beta.
• System shared memory is supported on Jetson. CUDA shared memory is not supported.
• GPU metrics, GCS storage, S3 storage and Azure storage are not supported.

The tar file contains the Triton server executable and shared libraries and also the C++ and Python client libraries and examples. For more information on how to install and use Triton on JetPack refer to jetson.md.

The wheel for the Python client library is present in the tar file and can be installed by running the following command:

python3 -m pip install --upgrade clients/python/tritonclient-2.56.0-py3-none-manylinux2014_aarch64.whl[all]

The Triton TensorRT-LLM container is built from the 25.03 image nvcr.io/nvidia/tritonserver:25.03-py3-min. Please refer to the support matrix for all dependency versions related to 25.03. However, the packages listed below have ...