Welcome to Triton Model Navigator, an inference toolkit designed for optimizing and deploying Deep Learning models with a focus on NVIDIA GPUs. The Triton Model Navigator streamlines the process of moving models and pipelines implemented in PyTorch, TensorFlow, and/or ONNX to TensorRT.
The Triton Model Navigator automates several critical steps, including model export, conversion, correctness testing, and profiling. By providing a single entry point for various supported frameworks, users can efficiently search for the best deployment option using the per-framework optimize function. The resulting optimized models are ready for deployment on either PyTriton or Triton Inference Server.
The distinct capabilities of Triton Model Navigator are summarized in the feature matrix:
| Feature | Description |
|---|---|
| Ease-of-use | Single line of code to run all possible optimization paths directly from your source code |
| Wide Framework Support | Compatible with various machine learning frameworks including PyTorch, TensorFlow, and ONNX |
| Models Optimization | Enhance the performance of models such as ResNET and BERT for efficient inference deployment |
| Pipelines Optimization | Streamline Python code pipelines for models such as Stable Diffusion and Whisper using Inplace Optimization, exclusive to PyTorch |
| Model Export and Conversion | Automate the process of exporting and converting models between various formats with focus on TensorRT and Torch-TensorRT |
| Correctness Testing | Ensures the converted model produce correct outputs validating against the original model |
| Performance Profiling | Profiles models to select the optimal format based on performance metrics such as latency and throughput to optimize target hardware utilization |
| Models Deployment | Automates models and pipelines deployment on PyTriton and Triton Inference Server through dedicated API |
Learn more about Triton Model Navigator features in documentation.
Before proceeding with the installation of Triton Model Navigator, ensure your system meets the following criteria:
- Operating System: Linux (Ubuntu 20.04+ recommended)
- Python: Version
3.8or newer - NVIDIA GPU
You can use NGC Containers for PyTorch and TensorFlow which contain all necessary dependencies:
The Triton Model Navigator can be installed from pypi.org.
For installing with PyTorch dependencies, use:
pip install -U --extra-index-url https://pypi.ngc.nvidia.com triton-model-navigator[torch]For installing with TensorFlow dependencies, use:
pip install -U --extra-index-url https://pypi.ngc.nvidia.com triton-model-navigator[tensorflow]The default CUDA version for ONNXRuntime since 1.19.0 is CUDA 12. To install with CUDA 11 support use following extra index url:
.. --extra-index-url https://aiinfra.pkgs.visualstudio.com/PublicPackages/_packaging/onnxruntime-cuda-11/pypi/simple/ ..The quick start section provides examples of possible optimization and deployment paths provided in Triton Model Navigator.
The Inplace Optimize allows seamless optimization of models for deployment, such as converting them to TensorRT, without requiring any changes to the original Python pipelines.
The below code presents Stable Diffusion pipeline optimization. But first, before you run the example install the required packages:
pip install transformers diffusers torchThen, initialize the pipeline and wrap the model components with nav.Module::
import model_navigator as nav
from transformers.modeling_outputs import BaseModelOutputWithPooling
from diffusers import DPMSolverMultistepScheduler, StableDiffusionPipeline
def get_pipeline():
# Initialize Stable Diffusion pipeline and wrap modules for optimization
pipe = StableDiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-2-1")
pipe.scheduler = DPMSolverMultistepScheduler.from_config(pipe.scheduler.config)
pipe = pipe.to("cuda")
pipe.text_encoder = nav.Module(
pipe.text_encoder,
name="clip",
output_mapping=lambda output: BaseModelOutputWithPooling(**output),
)
pipe.unet = nav.Module(
pipe.unet,
name="unet",
)
pipe.vae.decoder = nav.Module(
pipe.vae.decoder,
name="vae",
)
return pipePrepare a simple dataloader:
# Please mind, the first element in tuple need to be a batch size
def get_dataloader():
return [(1, "a photo of an astronaut riding a horse on mars")]Execute model optimization:
pipe = get_pipeline()
dataloader = get_dataloader()
nav.optimize(pipe, dataloader)Once the pipeline has been optimized, you can load explicit the most performant version of the modules executing:
nav.load_optimized()At this point, you can simply use the original pipeline to generate prediction with optimized models directly in Python:
pipe.to("cuda")
images = pipe(["a photo of an astronaut riding a horse on mars"])
image = images[0][0]
image.save("an_astronaut_riding_a_horse.png")An example of how to serve a Stable Diffusion pipeline through PyTriton can be found here.
Please read Error isolation when running Python script when you plan to place code in Python script.
Triton Model Navigator support also optimization path for deployment on Triton. This path is supported for nn.Module, keras.Model or ONNX files which inputs are tensors.
To optimize ResNet50 model from TorchHub run the following code:
import torch
import model_navigator as nav
# Initialize the model
resnet50 = torch.hub.load('NVIDIA/DeepLearningExamples:torchhub', 'nvidia_resnet50', pretrained=True).eval()
# Wrap model in nav.Module
resnet50 = nav.Module(resnet50, name="resnet50")
# Optimize Torch model loaded from TorchHub
nav.optimize(resnet50, dataloader=[(1, [torch.randn(1, 3, 256, 256)])])Once optimization is done, creating a model store for deployment on Triton is simple as following code:
import pathlib
# Generate the model store from optimized model
resnet50.triton_model_store(
model_repository_path=pathlib.Path("model_repository"),
)Please read Error isolation when running Python script when you plan to place code in Python script.
Triton Model Navigator enhances models and pipelines and provides a uniform method for profiling any Python function, callable, or model. At present, our support is limited strictly to static batch profiling scenarios.
As an example, we will use a simple function that simply sleeps for 50ms:
import time
def custom_fn(input_):
# wait 50ms
time.sleep(0.05)
return input_Let's provide a dataloader we will use for profiling:
# Tuple of batch size and data sample
dataloader = [(1, ["This is example input"])]Finally, run the profiling of the function with prepared dataloader:
nav.profile(custom_fn, dataloader)Important: Please review below section to prevent unexpected issues when running optimize.
For better error isolation, some conversions and exports are run in separate child processes using multiprocessing in
the spawn mode. This means that everything in a global scope will be run in a child process. You can encounter
unexpected issue when the optimization code is place in Python script and executed as:
python optimize.pyTo prevent nested optimization, you have to either put the optimize code in:
if __name__ == "__main__":
# optimization goes hereor
import multiprocessing as mp
if mp.current_process().name == "MainProcess":
# optimization goes hereIf none of the above works for you, you can run all optimization in a single process at the cost of error isolation by setting the following environment variable:
NAVIGATOR_USE_MULTIPROCESSING=FalseWe offer comprehensive, step-by-step guides that showcase the utilization of the Triton Model Navigator’s diverse features. These guides are designed to elucidate the processes of optimization, profiling, testing, and deployment of models using PyTriton and Triton Inference Server.