NVIDIA - TensorRT

TensorRT実行プロバイダー

TensorRT実行プロバイダーにより、ONNX Runtimeは汎用GPU加速と比較して同じハードウェア上でより優れた推論パフォーマンスを提供します。

ONNX RuntimeのTensorRT実行プロバイダーは、NVIDIAのTensorRT深層学習推論エンジンを利用して、NVIDIA GPUファミリーでONNXモデルを加速します。MicrosoftとNVIDIAは、TensorRT実行プロバイダーをONNX Runtimeに統合するために密接に協力しました。

目次

インストール

Onnx RuntimeのGPU（CUDA/TensorRT）バージョンを選択してください：https://onnxruntime.ai/docs/install。Jetpack用のプリビルドパッケージとDockerイメージは[Jetson Zoo](https://elinux.org/Jetson_Zoo#ONNX_Runtime)で利用できます。

ソースからのビルド

ビルド手順を参照してください。

要件

注意：

バージョン1.19以降、ONNX Runtime GPUパッケージの配布時にCUDA 12がデフォルトバージョンになります。

ORT 1.22以降、CUDA 12 GPUパッケージのみがリリースされます。

ONNX Runtime	TensorRT	CUDA
main	10.9	12.0-12.8
1.22	10.9	12.0-12.8
1.21	10.8	12.0-12.8, 11.8
1.20	10.4	12.0-12.6, 11.8
1.19	10.2	12.0-12.6, 11.8
1.18	10.0	11.8, 12.0-12.6
1.17	8.6	11.8, 12.0-12.6
1.16	8.6	11.8
1.15	8.6	11.8
1.14	8.5	11.6
1.12-1.13	8.4	11.4
1.11	8.2	11.4
1.10	8.0	11.4
1.9	8.0	11.4
1.7-1.8	7.2	11.0.3
1.5-1.6	7.1	10.2
1.2-1.4	7.0	10.1
1.0-1.1	6.0	10.0

CUDA/cuDNNバージョンの詳細については、CUDA EP要件を参照してください。

使用方法

C/C++

Ort::Env env = Ort::Env{ORT_LOGGING_LEVEL_ERROR, "Default"};
Ort::SessionOptions sf;
int device_id = 0;
Ort::ThrowOnError(OrtSessionOptionsAppendExecutionProvider_Tensorrt(sf, device_id));
Ort::ThrowOnError(OrtSessionOptionsAppendExecutionProvider_CUDA(sf, device_id));
Ort::Session session(env, model_path, sf);

C APIの詳細はこちらにあります。

Python

TensorRT実行プロバイダーを使用するには、InferenceSessionをインスタンシエートする際にTensorRT実行プロバイダーを明示的に登録する必要があります。 TensorRTがサポートしないノードをCUDA実行プロバイダーに割り当てることを可能にするため、CUDAExecutionProviderも登録することをお勧めします。

import onnxruntime as ort
# TensorrtExecutionProviderがより高い優先度を持つように、providersを['TensorrtExecutionProvider', 'CUDAExecutionProvider']に設定します。
sess = ort.InferenceSession('model.onnx', providers=['TensorrtExecutionProvider', 'CUDAExecutionProvider'])

設定

TensorRT設定を構成する方法は2つあります。TensorRT実行プロバイダーセッションオプションまたは環境変数（非推奨）です。

以下は、TensorRT EPセッションオプションを設定するための例と異なるシナリオです：

Click below for Python API example:

import onnxruntime as ort

model_path = '<path to model>'

# 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
    }),
    ('CUDAExecutionProvider', {
        'device_id': 0,
        'arena_extend_strategy': 'kNextPowerOfTwo',
        'gpu_mem_limit': 2 * 1024 * 1024 * 1024,
        'cudnn_conv_algo_search': 'EXHAUSTIVE',
        'do_copy_in_default_stream': True,
    })
]

sess_opt = ort.SessionOptions()
sess = ort.InferenceSession(model_path, sess_options=sess_opt, providers=providers)

Click below for C++ API example:

Ort::SessionOptions session_options;

const auto& api = Ort::GetApi();
OrtTensorRTProviderOptionsV2* tensorrt_options;
Ort::ThrowOnError(api.CreateTensorRTProviderOptions(&tensorrt_options));

std::vector<const char*> option_keys = {
    "device_id",
    "trt_max_workspace_size",
    "trt_max_partition_iterations",
    "trt_min_subgraph_size",
    "trt_fp16_enable",
    "trt_int8_enable",
    "trt_int8_use_native_calibration_table",
    "trt_dump_subgraphs",
    // below options are strongly recommended !
    "trt_engine_cache_enable",
    "trt_engine_cache_path",
    "trt_timing_cache_enable",
    "trt_timing_cache_path",
};
std::vector<const char*> option_values = {
    "1",
    "2147483648",
    "10",
    "5",
    "1",
    "1",
    "1",
    "1",
    "1",
    "1",
    "/path/to/cache",
    "1",
    "/path/to/cache", // can be same as the engine cache folder
};

Ort::ThrowOnError(api.UpdateTensorRTProviderOptions(tensorrt_options,
                                                    option_keys.data(), option_values.data(), option_keys.size()));


cudaStream_t cuda_stream;
cudaStreamCreate(&cuda_stream);
// this implicitly sets "has_user_compute_stream"
Ort::ThrowOnError(api.UpdateTensorRTProviderOptionsWithValue(cuda_options, "user_compute_stream", cuda_stream))

session_options.AppendExecutionProvider_TensorRT_V2(*tensorrt_options);
/// below code can be used to print all options
OrtAllocator* allocator;
char* options;
Ort::ThrowOnError(api.GetAllocatorWithDefaultOptions(&allocator));
Ort::ThrowOnError(api.GetTensorRTProviderOptionsAsString(tensorrt_options,          allocator, &options));

シナリオ

シナリオ	TensorRT EPセッションオプション	タイプ
デバイスとコンピュート設定
実行用のGPU IDを指定	device_id	int
GPU操作用のカスタムコンピュートストリームを設定	user_compute_stream	string
Engine Caching and Compatibility
Enable caching of TensorRT engines	trt_engine_cache_enable	bool
Set path to store cached TensorRT engines	trt_engine_cache_path	string
Set prefix for cached engine files	trt_engine_cache_prefix	string
Maximize engine compatibility across Ampere+ GPUs	trt_engine_hw_compatible	bool
Precision and Performance
Set TensorRT EP GPU memory usage limit	trt_max_workspace_size	int
Enable FP16 precision for faster performance	trt_fp16_enable	bool
Enable INT8 precision for quantized inference	trt_int8_enable	bool
Name INT8 calibration table for non-QDQ models	trt_int8_calibration_table_name	string
Use native TensorRT calibration tables	trt_int8_use_native_calibration_table	bool
Use heuristics to speed up engine builds	trt_build_heuristics_enable	bool
Enable sparsity to leverage zero values	trt_sparsity_enable	bool
Enable Deep Learning Accelerator (DLA) on edge SoC	trt_dla_enable	bool
Specify which DLA core to use	trt_dla_core	int
Subgraph and Graph Optimization
Limit partitioning iterations for model conversion	trt_max_partition_iterations	int
Set minimum size for subgraphs in partitioning	trt_min_subgraph_size	int
Dump optimized subgraphs for debugging	trt_dump_subgraphs	bool
Force sequential engine builds under multi-GPU	trt_force_sequential_engine_build	bool
Exclude specific op types from running on TRT	trt_op_types_to_exclude	string
Advanced Configuration and Profiling
Enable sharing of context memory between subgraphs	trt_context_memory_sharing_enable	bool
Force layer norm calculations to FP32	trt_layer_norm_fp32_fallback	bool
Capture CUDA graph for reduced launch overhead	trt_cuda_graph_enable	bool
Set optimization level for TensorRT builder	trt_builder_optimization_level	int
Set number of auxiliary streams for computation	trt_auxiliary_streams	int
Specify tactics sources for TensorRT	trt_tactic_sources	string
Add additional plugin library paths for TensorRT	trt_extra_plugin_lib_paths	string
Enable detailed logging of build steps	trt_detailed_build_log	bool
Timing cache
Enable use of timing cache to speed up builds	trt_timing_cache_enable	bool
Set path for storing timing cache	trt_timing_cache_path	string
Force use of timing cache regardless of GPU match	trt_force_timing_cache	bool
Dynamic Shape Profiling
Define min shapes	trt_profile_min_shapes	string
Define max shapes	trt_profile_max_shapes	string
Define optimal shapes	trt_profile_opt_shapes	string

Note: for bool type options, assign them with True/False in python, or 1/0 in C++.

Execution Provider Options

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.

device_id

• Description: GPU device ID.
• Default value: 0

user_compute_stream

• 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:

    import onnxruntime as ort
    import torch
    ...
    sess = ort.InferenceSession('model.onnx')
    if torch.cuda.is_available():
        s = torch.cuda.Stream()
        option = {"user_compute_stream": str(s.cuda_stream)}
        sess.set_providers(["TensorrtExecutionProvider"], [option])
        options = sess.get_provider_options()
    
        assert "TensorrtExecutionProvider" in options
        assert options["TensorrtExecutionProvider"].get("user_compute_stream", "") == str(s.cuda_stream)
        assert options["TensorrtExecutionProvider"].get("has_user_compute_stream", "") == "1"
    ...

• To take advantage of user compute stream, it is recommended to use I/O Binding to bind inputs and outputs to tensors in device.

trt_max_workspace_size

• Description: maximum workspace size for TensorRT engine.

• Default value: 1073741824 (1GB).

trt_max_partition_iterations

• 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.
• Default value: 1000.

trt_min_subgraph_size

• Description: minimum node size in a subgraph after partitioning.

• Subgraphs with smaller size will fall back to other execution providers.

• Default value: 1.

trt_fp16_enable

• Description: enable FP16 mode in TensorRT.

  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.

  • If True, native TensorRT generated calibration table is used;
  • If False, ONNXRUNTIME tool generated calibration table is used.

  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.

trt_dla_enable

• Description: enable DLA (Deep Learning Accelerator).

  Note: Not all Nvidia GPUs support DLA.

trt_dla_core

• Description: specify DLA core to execute on. Default value: 0.

trt_engine_cache_enable

• Description: 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.

  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)

trt_engine_cache_path

• 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.

trt_engine_cache_prefix

• 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.
  • This can help debugging subgraphs, e.g. by using trtexec --onnx my_model.onnx and check the outputs of the parser.

trt_force_sequential_engine_build

• Description: sequentially build TensorRT engines across provider instances in multi-GPU environment.

trt_context_memory_sharing_enable

• Description: share execution context memory between TensorRT subgraphs.

trt_layer_norm_fp32_fallback

• Description: force Pow + Reduce ops in layer norm to FP32.

trt_timing_cache_enable

• Description: enable TensorRT timing cache.
  • Check Timing cache for details.

trt_timing_cache_path

• Description: specify path for TensorRT timing cache if trt_timing_cache_enable is True.
  • Not specifying a trt_timing_cache_path will result in using the working directory

trt_force_timing_cache

• Description: force the TensorRT timing cache to be used even if device profile does not match.
  • A perfect match is only the exact same GPU model as the on that produced the timing cache.

trt_detailed_build_log

• Description: enable detailed build step logging on TensorRT EP with timing for each engine build.

trt_build_heuristics_enable

• Description: build engine using heuristics to reduce build time.

trt_cuda_graph_enable

• Description: this will capture a CUDA graph which can drastically help for a network with many small layers as it reduces launch overhead on the CPU.

trt_sparsity_enable

• Description: control if sparsity can be used by TRT.
  • Check --sparsity in trtexec command-line flags for details.

trt_builder_optimization_level

• Description: 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.

trt_auxiliary_streams

• 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.

trt_tactic_sources

• Description: 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”.

trt_extra_plugin_lib_paths

• 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.

trt_profile_min_shapes

trt_profile_max_shapes

trt_profile_opt_shapes

• 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 and TRT doc optimization 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.

trt_op_types_to_exclude

• Description: exclude specific op types from running on TRT. (Available in ORT 1.21.0)
  • The format is op_type_1,op_type_2,op_type_3...
  • One use case is to mitigate the performance issue mentioned below, it allows users to prevent DDS ops from running on TensorRT, ensuring they are executed by CUDA EP or CPU EP instead:

    ./onnxruntime_perf_test -r 1 -e tensorrt -i "trt_op_types_to_exclude|NonMaxSuppression,NonZero,RoiAlign" /path/to/onnx/your_model.onnx

  • Another use case is experimenting assigning ops to CUDA EP vs TRT EP

Environment Variables(deprecated)

Following environment variables can be set for TensorRT execution provider. Click below for more details.

• ORT_TENSORRT_MAX_WORKSPACE_SIZE: maximum workspace size for TensorRT engine. Default value: 1073741824 (1GB).

• 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_FP16_ENABLE: Enable FP16 mode in TensorRT. 1: enabled, 0: disabled. Default value: 0. Note not all Nvidia GPUs support FP16 precision.

• ORT_TENSORRT_INT8_ENABLE: Enable INT8 mode in TensorRT. 1: enabled, 0: disabled. Default value: 0. Note not all Nvidia GPUs support INT8 precision.

• 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_DLA_ENABLE: Enable DLA (Deep Learning Accelerator). 1: enabled, 0: disabled. Default value: 0. Note not all Nvidia GPUs support DLA.

• ORT_TENSORRT_DLA_CORE: Specify DLA core to execute on. Default value: 0.

• 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_SEQUENTIAL_ENGINE_BUILD: Sequentially build TensorRT engines across provider instances in multi-GPU environment. 1: enabled, 0: disabled. Default value: 0.

• ORT_TENSORRT_CONTEXT_MEMORY_SHARING_ENABLE: Share execution context memory between TensorRT subgraphs. Default 0 = false, nonzero = true.

• ORT_TENSORRT_LAYER_NORM_FP32_FALLBACK: Force Pow + Reduce ops in layer norm to FP32. Default 0 = false, nonzero = true.

• ORT_TENSORRT_TIMING_CACHE_ENABLE: Enable TensorRT timing cache. Default 0 = false, nonzero = true. Check Timing cache for details.

• 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.

• 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.

• 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.

• 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 and TRT doc optimization profiles for more details.

One can override default values by setting environment variables. e.g. on Linux:

# Override default max workspace size to 2GB
export ORT_TENSORRT_MAX_WORKSPACE_SIZE=2147483648

# Override default maximum number of iterations to 10
export ORT_TENSORRT_MAX_PARTITION_ITERATIONS=10

# Override default minimum subgraph node size to 5
export ORT_TENSORRT_MIN_SUBGRAPH_SIZE=5

# Enable FP16 mode in TensorRT
export ORT_TENSORRT_FP16_ENABLE=1

# Enable INT8 mode in TensorRT
export ORT_TENSORRT_INT8_ENABLE=1

# Use native TensorRT calibration table
export ORT_TENSORRT_INT8_USE_NATIVE_CALIBRATION_TABLE=1

# Enable TensorRT engine caching
export ORT_TENSORRT_ENGINE_CACHE_ENABLE=1
# Please Note warning above. This feature is experimental.
# 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.

# Specify TensorRT cache path
export ORT_TENSORRT_CACHE_PATH="/path/to/cache"

# Dump out subgraphs to run on TensorRT
export ORT_TENSORRT_DUMP_SUBGRAPHS=1

# Enable context memory sharing between TensorRT subgraphs. Default 0 = false, nonzero = true
export ORT_TENSORRT_CONTEXT_MEMORY_SHARING_ENABLE=1

TensorRT EP Caches

There are three major TRT EP caches:

• TRT timing cache
• TRT engine cache
• Embedded engine model / EPContext model

Caches can help reduce session creation time from minutes to seconds

Following numbers are measured from initializing session with TRT EP for SD UNet model.

• No cache (default)  – 384 seconds
  • The first run (warmup) can be very long because building engine involves exhaustive profiling for every kernels to select the optimal one.
• Timing cache used – 42 seconds
  • Keep layer-profiling information and reuse them to expedite build time
  • Timing cache can be shared across multiple models if layers are the same
• Engine cache used – 9 seconds
  • Serialize engine from memory to disk for later use
  • Skip entire engine build and deserialize engine cache to memory
• Embedded engine used (no builder instantiation) - 1.9 seconds
  • The serialized engine cache is wrapped inside an ONNX model
  • No builder will be instantiated, nor engine will be built
  • Quickly load engine with less processes needed

How to set caches

• Use Timing cache (.timing):
  • trt_timing_cache_enable = true
  • trt_timing_cache_path = .\
  • trt_force_timing_cache = true (accept slight GPU mismatch within CC)
• Use Engine Cache (.engine):
  • trt_engine_cache_enable = true
  • trt_engine_cache_path = .\trt_engines
• Use Embed Engine (_ctx.onnx):
  • Get the embed engine model via warmup run with the original model
  • trt_engine_cache_enable = true
  • trt_dump_ep_context_model = true
  • trt_ep_context_file_path = .\
  • Will be generated with inputs/outputs identical to original model
  • Run the embed engine model as the original model !

The folder structure of the caches:

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.

$./onnxruntime_perf_test -e tensorrt -r 1 -i "trt_engine_cache_enable|true trt_dump_ep_context_model|true" /model_database/transformer_model/model.onnx

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.

$./onnxruntime_perf_test -e tensorrt -r 1 /model_database/transformer_model/model_ctx.onnx

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 that is capable of packaging such a precompiled engine into an ONNX file is included in the python tools.

Performance Tuning

For performance tuning, please see guidance on this page: ONNX Runtime Perf Tuning

When/if using onnxruntime_perf_test, use the flag -e tensorrt. Check below for sample.

Shape Inference for TensorRT Subgraphs

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 (Check below for sample).

TensorRT Plugins Support

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:

Click below for Python API example:

from onnx import TensorProto, helper

def generate_model(model_name):
    nodes = [
        helper.make_node(
            "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"
            factor=0.123,
            span=128,
        ),
    ]

    graph = helper.make_graph(
        nodes,
        "trt_plugin_custom_op",
        [  # input
            helper.make_tensor_value_info("input1", TensorProto.FLOAT, [12, 256, 256]),
            helper.make_tensor_value_info("input2", TensorProto.FLOAT, [12, 256, 256]),
            helper.make_tensor_value_info("input3", TensorProto.FLOAT, [12, 256, 256]),
        ],
        [  # output
            helper.make_tensor_value_info("output", TensorProto.FLOAT, [12, 256, 256]),
        ],
    )

    model = helper.make_model(graph)
    onnx.save(model, model_name)

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.

Timing cache

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 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.

The following examples shows build time reduction with timing cache:

Model	no Cache	with Cache
efficientnet-lite4-11	34.6 s	7.7 s
yolov4	108.62 s	9.4 s

Click below for Python example:

import onnxruntime as ort

ort.set_default_logger_severity(0) # Turn on verbose mode for ORT TRT
sess_options = ort.SessionOptions()

trt_ep_options = {
    "trt_timing_cache_enable": True,
}

sess = ort.InferenceSession(
    "my_model.onnx",
    providers=[
        ("TensorrtExecutionProvider", trt_ep_options),
        "CUDAExecutionProvider",
    ],
)

# Once inference session initialization is done (assume no dynamic shape input, otherwise you must wait until inference run is done)
# you can find timing cache is saved in the 'trt_engine_cache_path' directory, e.g., TensorrtExecutionProvider_cache_cc75.timing, please note
# that the name contains information of compute capability.

sess.run(
    None,
    {"input_ids": np.zeros((1, 77), dtype=np.int32)}
)

Explicit shape range for dynamic shape input

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.

Click below for Python example:

import onnxruntime as ort

ort.set_default_logger_severity(0) # Turn on verbose mode for ORT TRT
sess_options = ort.SessionOptions()

trt_ep_options = {
    "trt_fp16_enable": True,
    "trt_engine_cache_enable": True,
    "trt_profile_min_shapes": "sample:2x4x64x64,encoder_hidden_states:2x77x768",
    "trt_profile_max_shapes": "sample:32x4x64x64,encoder_hidden_states:32x77x768",
    "trt_profile_opt_shapes": "sample:2x4x64x64,encoder_hidden_states:2x77x768",
}

sess = ort.InferenceSession(
    "my_model.onnx",
    providers=[
        ("TensorrtExecutionProvider", trt_ep_options),
        "CUDAExecutionProvider",
    ],
)

batch_size = 1
unet_dim = 4
max_text_len = 77
embed_dim = 768
latent_height = 64
latent_width = 64

args = {
    "sample": np.zeros(
        (2 * batch_size, unet_dim, latent_height, latent_width), dtype=np.float32
    ),
    "timestep": np.ones((1,), dtype=np.float32),
    "encoder_hidden_states": np.zeros(
        (2 * batch_size, max_text_len, embed_dim),
        dtype=np.float32,
    ),
}
sess.run(None, args)
# 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.

Data-dependant shape (DDS) ops

The DDS operations — NonMaxSuppression, NonZero, and RoiAlign — have output shapes that are only determined at runtime.

To ensure DDS ops are executed by TRT-EP/TRT instead of CUDA EP or CPU EP, please check the following:

• For TensorRT < 10.7: Build ORT with onnx-tensorrt OSS parser and use 10.X-GA-ORT-DDS branch.
• For TensorRT >= 10.7: By default, DDS ops will be executed by TRT.
• For ORT: By default, ORT relies on the TRT parser to decide if DDS ops run with TRT. However, note that ORT 1.20.1 and 1.20.2 will not run DDS ops with TRT due to a known performance issue.

Samples

This example shows how to run the Faster R-CNN model on TensorRT execution provider.

1. Download the Faster R-CNN onnx model from the ONNX model zoo here.

2. Infer shapes in the model by running the shape inference script

   python symbolic_shape_infer.py --input /path/to/onnx/model/model.onnx --output /path/to/onnx/model/new_model.onnx --auto_merge

3. To test model with sample input and verify the output, run onnx_test_runner under ONNX Runtime build directory.

   Models and test_data_set_ folder need to be stored under the same path. onnx_test_runner will test all models under this path.

   ./onnx_test_runner -e tensorrt /path/to/onnx/model/

4. To test on model performance, run onnxruntime_perf_test on your shape-inferred Faster-RCNN model

   Download sample test data with model from model zoo, and put test_data_set folder next to your inferred model

   # e.g.
   # -r: set up test repeat time
   # -e: set up execution provider
   # -i: set up params for execution provider options
   ./onnxruntime_perf_test -r 1 -e tensorrt -i "trt_fp16_enable|true" /path/to/onnx/your_inferred_model.onnx

Please see this Notebook for an example of running a model on GPU using ONNX Runtime through Azure Machine Learning Services.

Known Issues

• TensorRT 8.6 built-in parser and TensorRT oss parser behaves differently. Namely built-in parser cannot recognize some custom plugin ops while OSS parser can. See EfficientNMS_TRT missing attribute class_agnostic w/ TensorRT 8.6 .
• There is a performance issue for TensorRT versions 10.0 to 10.5 when running models, such as Faster-RCNN, that:
  • contain data-dependent shape (DDS) operations, like NonMaxSuppression, NonZero, and RoiAlign, and
  • DDS ops are executed with TRT