With the TensorRT execution provider, the ONNX Runtime delivers better inferencing performance on the same hardware compared to generic GPU acceleration.
The TensorRT execution provider in the ONNX Runtime makes use of NVIDIA's [TensorRT](https://developer.nvidia.com/tensorrt) Deep Learning inferencing engine to accelerate ONNX model in their family of GPUs. Microsoft and NVIDIA worked closely to integrate the TensorRT execution provider with ONNX Runtime.
Please select the GPU (CUDA/TensorRT) version of OnnxRuntime: https://onnxruntime.ai/docs/install. Pre-built packages and Docker images are available for Jetpack in the [Jetson Zoo](https://elinux.org/Jetson_Zoo#ONNX_Runtime).
To use TensorRT execution provider, you must explicitly register TensorRT execution provider when instantiating the `InferenceSession`.
Note that it is recommended you also register `CUDAExecutionProvider` to allow Onnx Runtime to assign nodes to CUDA execution provider that TensorRT does not support.
There are two ways to configure TensorRT settings, either by [TensorRT Execution Provider Session Option](./TensorRT-ExecutionProvider.md#execution-provider-options) or [Environment Variables(deprecated)](./TensorRT-ExecutionProvider.md#environment-variablesdeprecated).
Here are examples and different [scenarios](./TensorRT-ExecutionProvider.md#scenario) to set TensorRT EP session options:
#### Click below for Python API example:
<details>
```python
import onnxruntime as ort
model_path = '<pathtomodel>'
# note: for bool type options in python API, set them as False/True
providers = [
('TensorrtExecutionProvider', {
'device_id': 0, # Select GPU to execute
'trt_max_workspace_size': 2147483648, # Set GPU memory usage limit
'trt_fp16_enable': True, # Enable FP16 precision for faster inference
| Use heuristics to speed up engine builds | [trt_build_heuristics_enable](./TensorRT-ExecutionProvider.md#trt_build_heuristics_enable) | bool |
| Enable sparsity to leverage zero values | [trt_sparsity_enable](./TensorRT-ExecutionProvider.md#trt_sparsity_enable) | bool |
| Enable Deep Learning Accelerator (DLA) on edge SoC | [trt_dla_enable](./TensorRT-ExecutionProvider.md#trt_dla_enable) | bool |
| Specify which DLA core to use | [trt_dla_core](./TensorRT-ExecutionProvider.md#trt_dla_core) | int |
| | | |
| **Subgraph and Graph Optimization** | | |
| Limit partitioning iterations for model conversion | [trt_max_partition_iterations](./TensorRT-ExecutionProvider.md#trt_max_partition_iterations) | int |
| Set minimum size for subgraphs in partitioning | [trt_min_subgraph_size](./TensorRT-ExecutionProvider.md#trt_min_subgraph_size) | int |
| Force sequential engine builds under multi-GPU | [trt_force_sequential_engine_build](./TensorRT-ExecutionProvider.md#trt_force_sequential_engine_build) | bool |
| | | |
| **Advanced Configuration and Profiling** | | |
| Enable sharing of context memory between subgraphs | [trt_context_memory_sharing_enable](./TensorRT-ExecutionProvider.md#trt_context_memory_sharing_enable) | bool |
| Force layer norm calculations to FP32 | [trt_layer_norm_fp32_fallback](./TensorRT-ExecutionProvider.md#trt_layer_norm_fp32_fallback) | bool |
| Capture CUDA graph for reduced launch overhead | [trt_cuda_graph_enable](./TensorRT-ExecutionProvider.md#trt_cuda_graph_enable) | bool |
| Set optimization level for TensorRT builder | [trt_builder_optimization_level](./TensorRT-ExecutionProvider.md#trt_builder_optimization_level) | int |
| Set number of auxiliary streams for computation | [trt_auxiliary_streams](./TensorRT-ExecutionProvider.md#trt_auxiliary_streams) | int |
TensorRT configurations can be set by execution provider options. It's useful when each model and inference session have their own configurations. In this case, execution provider option settings will override any environment variable settings. All configurations should be set explicitly, otherwise default value will be taken.
* Description: define the compute stream for the inference to run on. It implicitly sets the `has_user_compute_stream` option. It cannot be set through `UpdateTensorRTProviderOptions`, but rather `UpdateTensorRTProviderOptionsWithValue`.
* This cannot be used in combination with an external allocator.
* This can also be set using the python API.
* i.e The cuda stream captured from pytorch can be passed into ORT-TRT. Click below to check sample code:
* To take advantage of user compute stream, it is recommended to use [I/O Binding](https://onnxruntime.ai/docs/api/python/api_summary.html#data-on-device) to bind inputs and outputs to tensors in device.
* Description: maximum number of iterations allowed in model partitioning for TensorRT.
* If target model can't be successfully partitioned when the maximum number of iterations is reached, the whole model will fall back to other execution providers such as CUDA or CPU.
> Note: not all Nvidia GPUs support FP16 precision.
##### trt_int8_enable
* Description: enable INT8 mode in TensorRT.
> Note: not all Nvidia GPUs support INT8 precision.
##### trt_int8_calibration_table_name
* Description: specify INT8 calibration table file for non-QDQ models in INT8 mode.
> Note: calibration table should not be provided for QDQ model because TensorRT doesn't allow calibration table to be loded if there is any Q/DQ node in the model. By default the name is empty.
##### trt_int8_use_native_calibration_table
* Description: select what calibration table is used for non-QDQ models in INT8 mode.
> Note: Please copy up-to-date calibration table file to `trt_engine_cache_path` before inference. Calibration table is specific to models and calibration data sets. Whenever new calibration table is generated, old file in the path should be cleaned up or be replaced.
* The purpose of using engine caching is to save engine build time in the case that TensorRT may take long time to optimize and build engine.
* Engine will be cached when it's built for the first time so next time when new inference session is created the engine can be loaded directly from cache. In order to validate that the loaded engine is usable for current inference, engine profile is also cached and loaded along with engine. If current input shapes are in the range of the engine profile, the loaded engine can be safely used. Otherwise if input shapes are out of range, profile cache will be updated to cover the new shape and engine will be recreated based on the new profile (and also refreshed in the engine cache).
* Note each engine is created for specific settings such as model path/name, precision (FP32/FP16/INT8 etc), workspace, profiles etc, and specific GPUs and it's not portable, so it's essential to make sure those settings are not changing, otherwise the engine needs to be rebuilt and cached again.
* Description: specify path for TensorRT engine and profile files if `trt_engine_cache_enable` is `True`, or path for INT8 calibration table file if `trt_int8_enable` is `True`.
* Description: customize engine cache prefix when `trt_engine_cache_enable` is `True`.
* ORT-TRT will only reuse existing engine cache with customized prefix if the same prefix is assigned in `trt_engine_cache_prefix`. If this option is empty, new engine cache with default prefix will be generated.
##### trt_dump_subgraphs
* Description: dumps the subgraphs that are transformed into TRT engines in onnx format to the filesystem.
* Description: this will capture a [CUDA graph](https://developer.nvidia.com/blog/cuda-graphs/) which can drastically help for a network with many small layers as it reduces launch overhead on the CPU.
* Description: control if sparsity can be used by TRT.
* Check `--sparsity` in `trtexec` command-line flags for [details](https://docs.nvidia.com/deeplearning/tensorrt/developer-guide/index.html#trtexec-flags).
> WARNING: levels below 3 do not guarantee good engine performance, but greatly improve build time. Default 3, valid range [0-5]. Check `--builderOptimizationLevel` in `trtexec` command-line flags for [details](https://docs.nvidia.com/deeplearning/tensorrt/developer-guide/index.html#trtexec-flags).
* Description: set maximum number of auxiliary streams per inference stream.
* Setting this value to 0 will lead to optimal memory usage.
* Default -1 = heuristics.
* Check `--maxAuxStreams` in `trtexec` command-line flags for [details](https://docs.nvidia.com/deeplearning/tensorrt/developer-guide/index.html#trtexec-flags).
* Description: specify extra TensorRT plugin library paths.
* ORT TRT by default supports any TRT plugins registered in TRT registry in TRT plugin library (i.e., `libnvinfer_plugin.so`).
* Moreover, if users want to use other TRT plugins that are not in TRT plugin library,
* for example, FasterTransformer has many TRT plugin implementations for different models, user can specify like this `ORT_TENSORRT_EXTRA_PLUGIN_LIB_PATHS=libvit_plugin.so;libvit_int8_plugin.so`.
* Description: build with dynamic shapes using a profile with the min/max/opt shapes provided.
* The format of the profile shapes is `input_tensor_1:dim_1xdim_2x...,input_tensor_2:dim_3xdim_4x...,...`
* These three flags should all be provided in order to enable explicit profile shapes feature.
* Check [Explicit shape range for dynamic shape input](#explicit-shape-range-for-dynamic-shape-input) and TRT doc [optimization profiles](https://docs.nvidia.com/deeplearning/tensorrt/developer-guide/index.html#opt_profiles) for more details.
##### trt_engine_hw_compatible
* Description: enable Ampere+ hardware compatibility if `trt_engine_cache_enable` is enabled
* Hardware-compatible engines can be reused across all Ampere+ GPU environments (may have lower throughput and/or higher latency).
* Engines will be generated and loaded with `sm80+` name suffix, instead of actual compute capacity.
* Turing and former Nvidia GPU architecture and Nvidia Jetson Orin platform are not eligble to this option.
*`ORT_TENSORRT_MAX_PARTITION_ITERATIONS`: maximum number of iterations allowed in model partitioning for TensorRT. If target model can't be successfully partitioned when the maximum number of iterations is reached, the whole model will fall back to other execution providers such as CUDA or CPU. Default value: 1000.
*`ORT_TENSORRT_MIN_SUBGRAPH_SIZE`: minimum node size in a subgraph after partitioning. Subgraphs with smaller size will fall back to other execution providers. Default value: 1.
*`ORT_TENSORRT_INT8_CALIBRATION_TABLE_NAME`: Specify INT8 calibration table file for non-QDQ models in INT8 mode. Note calibration table should not be provided for QDQ model because TensorRT doesn't allow calibration table to be loded if there is any Q/DQ node in the model. By default the name is empty.
*`ORT_TENSORRT_INT8_USE_NATIVE_CALIBRATION_TABLE`: Select what calibration table is used for non-QDQ models in INT8 mode. If 1, native TensorRT generated calibration table is used; if 0, ONNXRUNTIME tool generated calibration table is used. Default value: 0.
* **Note: Please copy up-to-date calibration table file to `ORT_TENSORRT_CACHE_PATH` before inference. Calibration table is specific to models and calibration data sets. Whenever new calibration table is generated, old file in the path should be cleaned up or be replaced.**
*`ORT_TENSORRT_ENGINE_CACHE_ENABLE`: Enable TensorRT engine caching. The purpose of using engine caching is to save engine build time in the case that TensorRT may take long time to optimize and build engine. Engine will be cached when it's built for the first time so next time when new inference session is created the engine can be loaded directly from cache. In order to validate that the loaded engine is usable for current inference, engine profile is also cached and loaded along with engine. If current input shapes are in the range of the engine profile, the loaded engine can be safely used. Otherwise if input shapes are out of range, profile cache will be updated to cover the new shape and engine will be recreated based on the new profile (and also refreshed in the engine cache). Note each engine is created for specific settings such as model path/name, precision (FP32/FP16/INT8 etc), workspace, profiles etc, and specific GPUs and it's not portable, so it's essential to make sure those settings are not changing, otherwise the engine needs to be rebuilt and cached again. 1: enabled, 0: disabled. Default value: 0.
* **Warning: Please clean up any old engine and profile cache files (.engine and .profile) if any of the following changes:**
* Model changes (if there are any changes to the model topology, opset version, operators etc.)
* ORT version changes (i.e. moving from ORT version 1.8 to 1.9)
* TensorRT version changes (i.e. moving from TensorRT 7.0 to 8.0)
* Hardware changes. (Engine and profile files are not portable and optimized for specific Nvidia hardware)
*`ORT_TENSORRT_CACHE_PATH`: Specify path for TensorRT engine and profile files if `ORT_TENSORRT_ENGINE_CACHE_ENABLE` is 1, or path for INT8 calibration table file if ORT_TENSORRT_INT8_ENABLE is 1.
*`ORT_TENSORRT_DUMP_SUBGRAPHS`: Dumps the subgraphs that are transformed into TRT engines in onnx format to the filesystem. This can help debugging subgraphs, e.g. by using `trtexec --onnx my_model.onnx` and check the outputs of the parser. 1: enabled, 0: disabled. Default value: 0.
*`ORT_TENSORRT_FORCE_TIMING_CACHE_ENABLE`: Force the TensorRT timing cache to be used even if device profile does not match. Default 0 = false, nonzero = true.
*`ORT_TENSORRT_DETAILED_BUILD_LOG_ENABLE`: Enable detailed build step logging on TensorRT EP with timing for each engine build. Default 0 = false, nonzero = true.
*`ORT_TENSORRT_BUILD_HEURISTICS_ENABLE`: Build engine using heuristics to reduce build time. Default 0 = false, nonzero = true.
*`ORT_TENSORRT_SPARSITY_ENABLE`: Control if sparsity can be used by TRT. Default 0 = false, 1 = true. Check `--sparsity` in `trtexec` command-line flags for [details](https://docs.nvidia.com/deeplearning/tensorrt/developer-guide/index.html#trtexec-flags).
*`ORT_TENSORRT_BUILDER_OPTIMIZATION_LEVEL`: Set the builder optimization level. WARNING: levels below 3 do not guarantee good engine performance, but greatly improve build time. Default 3, valid range [0-5]. Check `--builderOptimizationLevel` in `trtexec` command-line flags for [details](https://docs.nvidia.com/deeplearning/tensorrt/developer-guide/index.html#trtexec-flags).
*`ORT_TENSORRT_AUXILIARY_STREAMS`: Set maximum number of auxiliary streams per inference stream. Setting this value to 0 will lead to optimal memory usage. Default -1 = heuristics. Check `--maxAuxStreams` in `trtexec` command-line flags for [details](https://docs.nvidia.com/deeplearning/tensorrt/developer-guide/index.html#trtexec-flags).
*`ORT_TENSORRT_TACTIC_SOURCES`: Specify the tactics to be used by adding (+) or removing (-) tactics from the default tactic sources (default = all available tactics) e.g. "-CUDNN,+CUBLAS" available keys: "CUBLAS", "CUBLAS_LT", "CUDNN" or "EDGE_MASK_CONVOLUTIONS".
*`ORT_TENSORRT_EXTRA_PLUGIN_LIB_PATHS`: Specify extra TensorRT plugin library paths. ORT TRT by default supports any TRT plugins registered in TRT registry in TRT plugin library (i.e., `libnvinfer_plugin.so`). Moreover, if users want to use other TRT plugins that are not in TRT plugin library, for example, FasterTransformer has many TRT plugin implementations for different models, user can specify like this `ORT_TENSORRT_EXTRA_PLUGIN_LIB_PATHS=libvit_plugin.so;libvit_int8_plugin.so`.
*`ORT_TENSORRT_PROFILE_MIN_SHAPES`, `ORT_TENSORRT_PROFILE_MAX_SHAPES` and `ORT_TENSORRT_PROFILE_OPT_SHAPES` : Build with dynamic shapes using a profile with the min/max/opt shapes provided. The format of the profile shapes is "input_tensor_1:dim_1xdim_2x...,input_tensor_2:dim_3xdim_4x...,..." and these three flags should all be provided in order to enable explicit profile shapes feature. Check [Explicit shape range for dynamic shape input](#explicit-shape-range-for-dynamic-shape-input) and TRT doc [optimization profiles](https://docs.nvidia.com/deeplearning/tensorrt/developer-guide/index.html#opt_profiles) for more details.
# Engine cache files must be invalidated if there are any changes to the model, ORT version, TensorRT version or if the underlying hardware changes. Engine files are not portable across devices.
With the following command, the embedded engine model (`model_ctx.onnx`) will be generated along with the engine cache in the same directory.
Note: The example does not specify `trt_engine_cache_path` because `onnxruntime_perf_test` requires a specific folder structure to run the inference. However, we still recommend specifying `trt_engine_cache_path` to better organize the caches.
Once the inference is complete, the embedded engine model is saved to disk. User can then run this model just like the original one, but with a significantly quicker session creation time.
### More about Embedded engine model / EPContext model
* One constraint is that the entire model needs to be TRT eligible
* When running the embedded engine model, the default setting is `trt_ep_context_embed_mode=0`, where the engine cache path is embedded and TRT EP will look for the engine cache on the disk. Alternatively, users can set `trt_ep_context_embed_mode=1`, embedding the entire engine binary data as a string in the model. However, this mode increases initialization time due to ORT graph optimization hashing the long string. Therefore, we recommend using `trt_ep_context_embed_mode=0`.
* The default name of an embedded engine model will have `_ctx.onnx` appended to the end. Users can specify `trt_ep_context_file_path=my_ep_context_model.onnx` to overwrite this default name.
* If an embedded engine is used the library **`nvinfer_builder_resource` of TensorRT is not required**, which is by far the largest library. This enables the case of shipping a minimal set of libraries in the case that a fixed set of models is used which are packaged as precompield engine.
* Besides everything that embedded engines enable to accelerate the load time, they also **enable packaging an externally compiled engine** using e.g. `trtexec`. A [python script](https://github.com/microsoft/onnxruntime/blob/main/onnxruntime/python/tools/tensorrt/gen_trt_engine_wrapper_onnx_model.py) that is capable of packaging such a precompiled engine into an ONNX file is included in the python tools.
When/if using [onnxruntime_perf_test](https://github.com/microsoft/onnxruntime/tree/main/onnxruntime/test/perftest#onnxruntime-performance-test), use the flag `-e tensorrt`. Check below for sample.
If some operators in the model are not supported by TensorRT, ONNX Runtime will partition the graph and only send supported subgraphs to TensorRT execution provider. Because TensorRT requires that all inputs of the subgraphs have shape specified, ONNX Runtime will throw error if there is no input shape info. In this case please run shape inference for the entire model first by running script [here](https://github.com/microsoft/onnxruntime/blob/main/onnxruntime/python/tools/symbolic_shape_infer.py) (Check below for sample).
ORT TRT can leverage the TRT plugins which come with TRT plugin library in official release. To use TRT plugins, firstly users need to create the custom node (a one-to-one mapping to TRT plugin) with a registered plugin name and `trt.plugins` domain in the ONNX model. So, ORT TRT can recognize this custom node and pass the node together with the subgraph to TRT. Please see following python example to create a new custom node in the ONNX model:
"DisentangledAttention_TRT", # The registered name is from https://github.com/NVIDIA/TensorRT/blob/main/plugin/disentangledAttentionPlugin/disentangledAttentionPlugin.cpp#L36
["input1", "input2", "input3"],
["output"],
"DisentangledAttention_TRT",
domain="trt.plugins", # The domain has to be "trt.plugins"
Note: If users want to use TRT plugins that are not in the TRT plugin library in official release, please see the ORT TRT provider option `trt_extra_plugin_lib_paths` for more details.
Enabling `trt_timing_cache_enable` will enable ORT TRT to use TensorRT timing cache to accelerate engine build time on a device with the same compute capability. This will work across models as it simply stores kernel latencies for specific configurations and cubins (TRT 9.0+). Those files are usually very small (only a few KB or MB) which makes them very easy to ship with an application to accelerate the build time on the user end.
_Note:_ A timing cache can be used across one [GPU compute capability](https://developer.nvidia.com/cuda-gpus) similar to an engine. Nonetheless the preferred way is to use one per GPU model, but practice shows that sharing across one compute capability works well in most cases.
ORT TRT lets you explicitly specify min/max/opt shapes for each dynamic shape input through three provider options, `trt_profile_min_shapes`, `trt_profile_max_shapes` and `trt_profile_opt_shapes`. If these three provider options are not specified
and model has dynamic shape input, ORT TRT will determine the min/max/opt shapes for the dynamic shape input based on incoming input tensor. The min/max/opt shapes are required for TRT optimization profile (An optimization profile describes a range of dimensions for each TRT network input and the dimensions that the auto-tuner will use for optimization. When using runtime dimensions, you must create at least one optimization profile at build time.)
To use the engine cache built with optimization profiles specified by explicit shape ranges, user still needs to provide those three provider options as well as engine cache enable flag.
ORT TRT will firstly compare the shape ranges of those three provider options with the shape ranges saved in the .profile file, and then rebuild the engine if the shape ranges don't match.
# you can find engine cache and profile cache are saved in the 'trt_engine_cache_path' directory, e.g.
# TensorrtExecutionProvider_TRTKernel_graph_torch_jit_1843998305741310361_0_0_fp16.engine and TensorrtExecutionProvider_TRTKernel_graph_torch_jit_1843998305741310361_0_0_fp16.profile.
Please note that there is a constraint of using this explicit shape range feature, i.e., all the dynamic shape inputs should be provided with corresponding min/max/opt shapes.
1. Download the Faster R-CNN onnx model from the ONNX model zoo [here](https://github.com/onnx/models/tree/main/validated/vision/object_detection_segmentation/faster-rcnn).
2. Infer shapes in the model by running the [shape inference script](https://github.com/microsoft/onnxruntime/blob/main/onnxruntime/python/tools/symbolic_shape_infer.py)
> Download sample test data with model from [model zoo](https://github.com/onnx/models/tree/main/validated/vision/object_detection_segmentation/faster-rcnn), and put test_data_set folder next to your inferred model
Please see [this Notebook](https://github.com/microsoft/onnxruntime/blob/main/docs/python/notebooks/onnx-inference-byoc-gpu-cpu-aks.ipynb) for an example of running a model on GPU using ONNX Runtime through Azure Machine Learning Services.
