Skip to content

Commit dd5d189

Browse files
dzierdzier
committed
Update README for 21.12 release
1 parent 5505030 commit dd5d189

File tree

1 file changed

+267
-2
lines changed

1 file changed

+267
-2
lines changed

‎README.md

+267-2
Original file line numberDiff line numberDiff line change
@@ -30,5 +30,270 @@
3030

3131
# Triton Inference Server
3232

33-
**NOTE: You are currently on the r21.12 branch which tracks stabilization
34-
towards the next release. This branch is not usable during stabilization.**
33+
Triton Inference Server provides a cloud and edge inferencing solution
34+
optimized for both CPUs and GPUs. Triton supports an HTTP/REST and
35+
GRPC protocol that allows remote clients to request inferencing for
36+
any model being managed by the server. For edge deployments, Triton is
37+
available as a shared library with a C API that allows the full
38+
functionality of Triton to be included directly in an
39+
application.
40+
41+
## What's New in 2.17.0
42+
43+
* Improved [Inferentia support](https://github.com/triton-inference-server/python_backend/tree/r21.12/inferentia#readme)
44+
to use Neuron Runtime 2.x and multiple instances.
45+
46+
* Models from MLflow can now be deployed to Triton with the [MLflow
47+
plugin](deploy/mlflow-triton-plugin).
48+
49+
* The preview release of TorchTRT models is now supported. PyTorch models
50+
optimized using TensorRT can now be loaded into Triton in the same way as
51+
other PyTorch models.
52+
53+
* At the end of each Model Analyzer phase, an example command line will be
54+
printed to run the next phase.
55+
56+
## Features
57+
58+
* [Multiple deep-learning
59+
frameworks](https://github.com/triton-inference-server/backend). Triton
60+
can manage any number and mix of models (limited by system disk and
61+
memory resources). Triton supports TensorRT, TensorFlow GraphDef,
62+
TensorFlow SavedModel, ONNX, PyTorch TorchScript and OpenVINO model
63+
formats. Both TensorFlow 1.x and TensorFlow 2.x are
64+
supported. Triton also supports TensorFlow-TensorRT and
65+
ONNX-TensorRT integrated models.
66+
67+
* [Concurrent model
68+
execution](docs/architecture.md#concurrent-model-execution). Multiple
69+
models (or multiple instances of the same model) can run
70+
simultaneously on the same GPU or on multiple GPUs.
71+
72+
* [Dynamic batching](docs/architecture.md#models-and-schedulers). For
73+
models that support batching, Triton implements multiple scheduling
74+
and batching algorithms that combine individual inference requests
75+
together to improve inference throughput. These scheduling and
76+
batching decisions are transparent to the client requesting
77+
inference.
78+
79+
* [Extensible
80+
backends](https://github.com/triton-inference-server/backend). In
81+
addition to deep-learning frameworks, Triton provides a *backend
82+
API* that allows Triton to be extended with any model execution
83+
logic implemented in
84+
[Python](https://github.com/triton-inference-server/python_backend)
85+
or
86+
[C++](https://github.com/triton-inference-server/backend/blob/main/README.md#triton-backend-api),
87+
while still benefiting from the CPU and GPU support, concurrent
88+
execution, dynamic batching and other features provided by Triton.
89+
90+
* [Model pipelines](docs/architecture.md#ensemble-models). Triton
91+
*ensembles* represents a pipeline of one or more models and the
92+
connection of input and output tensors between those models. A
93+
single inference request to an ensemble will trigger the execution
94+
of the entire pipeline.
95+
96+
* [HTTP/REST and GRPC inference
97+
protocols](docs/inference_protocols.md) based on the community
98+
developed [KFServing
99+
protocol](https://github.com/kubeflow/kfserving/tree/master/docs/predict-api/v2).
100+
101+
* A [C API](docs/inference_protocols.md#c-api) allows Triton to be
102+
linked directly into your application for edge and other in-process
103+
use cases.
104+
105+
* [Metrics](docs/metrics.md) indicating GPU utilization, server
106+
throughput, and server latency. The metrics are provided in
107+
Prometheus data format.
108+
109+
## Documentation
110+
111+
[Triton Architecture](docs/architecture.md) gives a high-level
112+
overview of the structure and capabilities of the inference
113+
server. There is also an [FAQ](docs/faq.md). Additional documentation
114+
is divided into [*user*](#user-documentation) and
115+
[*developer*](#developer-documentation) sections. The *user*
116+
documentation describes how to use Triton as an inference solution,
117+
including information on how to configure Triton, how to organize and
118+
configure your models, how to use the C++ and Python clients, etc. The
119+
*developer* documentation describes how to build and test Triton and
120+
also how Triton can be extended with new functionality.
121+
122+
The Triton [Release
123+
Notes](https://docs.nvidia.com/deeplearning/triton-inference-server/release-notes/index.html)
124+
and [Support
125+
Matrix](https://docs.nvidia.com/deeplearning/dgx/support-matrix/index.html)
126+
indicate the required versions of the NVIDIA Driver and CUDA, and also
127+
describe supported GPUs.
128+
129+
### User Documentation
130+
131+
* [QuickStart](docs/quickstart.md)
132+
* [Install](docs/quickstart.md#install-triton-docker-image)
133+
* [Run](docs/quickstart.md#run-triton)
134+
* [Model Repository](docs/model_repository.md)
135+
* [Model Configuration](docs/model_configuration.md)
136+
* [Model Management](docs/model_management.md)
137+
* [Custom Operations](docs/custom_operations.md)
138+
* [Client Libraries and Examples](https://github.com/triton-inference-server/client)
139+
* [Optimization](docs/optimization.md)
140+
* [Model Analyzer](docs/model_analyzer.md)
141+
* [Performance Analyzer](docs/perf_analyzer.md)
142+
* [Metrics](docs/metrics.md)
143+
* [Rate Limiter](docs/rate_limiter.md)
144+
* [Jetson and JetPack](docs/jetson.md)
145+
146+
The [quickstart](docs/quickstart.md) walks you through all the steps
147+
required to install and run Triton with an example image
148+
classification model and then use an example client application to
149+
perform inferencing using that model. The quickstart also demonstrates
150+
how [Triton supports both GPU systems and CPU-only
151+
systems](docs/quickstart.md#run-triton).
152+
153+
The first step in using Triton to serve your models is to place one or
154+
more models into a [model
155+
repository](docs/model_repository.md). Optionally, depending on the type
156+
of the model and on what Triton capabilities you want to enable for
157+
the model, you may need to create a [model
158+
configuration](docs/model_configuration.md) for the model. If your
159+
model has [custom operations](docs/custom_operations.md) you will need
160+
to make sure they are loaded correctly by Triton.
161+
162+
After you have your model(s) available in Triton, you will want to
163+
send inference and other requests to Triton from your *client*
164+
application. The [Python and C++ client
165+
libraries](https://github.com/triton-inference-server/client) provide
166+
APIs to simplify this communication. There are also a large number of
167+
[client examples](https://github.com/triton-inference-server/client)
168+
that demonstrate how to use the libraries. You can also send
169+
HTTP/REST requests directly to Triton using the [HTTP/REST JSON-based
170+
protocol](docs/inference_protocols.md#httprest-and-grpc-protocols) or
171+
[generate a GRPC client for many other
172+
languages](https://github.com/triton-inference-server/client).
173+
174+
Understanding and [optimizing performance](docs/optimization.md) is an
175+
important part of deploying your models. The Triton project provides
176+
the [Performance Analyzer](docs/perf_analyzer.md) and the [Model
177+
Analyzer](docs/model_analyzer.md) to help your optimization
178+
efforts. Specifically, you will want to optimize [scheduling and
179+
batching](docs/architecture.md#models-and-schedulers) and [model
180+
instances](docs/model_configuration.md#instance-groups) appropriately
181+
for each model. You can also enable cross-model prioritization using
182+
[rate limiter](docs/rate_limiter.md) which manages the rate at which
183+
requests are scheduled on model instances. You may also want to
184+
consider [ensembling multiple models and pre/post-processing](docs/architecture.md#ensemble-models)
185+
into a pipeline. In some cases you may find [individual inference
186+
request trace data](docs/trace.md) useful when optimizing. A
187+
[Prometheus metrics endpoint](docs/metrics.md) allows you to visualize
188+
and monitor aggregate inference metrics.
189+
190+
NVIDIA publishes a number of [deep learning
191+
examples](https://github.com/NVIDIA/DeepLearningExamples) that use
192+
Triton.
193+
194+
As part of your deployment strategy you may want to [explicitly manage
195+
what models are available by loading and unloading
196+
models](docs/model_management.md) from a running Triton server. If you
197+
are using Kubernetes for deployment there are simple examples of how
198+
to deploy Triton using Kubernetes and Helm, one for
199+
[GCP](deploy/gcp/README.md) and one for [AWS](deploy/aws/README.md).
200+
201+
The [version 1 to version 2 migration
202+
information](docs/v1_to_v2.md) is helpful if you are moving to
203+
version 2 of Triton from previously using version 1.
204+
205+
### Developer Documentation
206+
207+
* [Build](docs/build.md)
208+
* [Protocols and APIs](docs/inference_protocols.md).
209+
* [Backends](https://github.com/triton-inference-server/backend)
210+
* [Repository Agents](docs/repository_agents.md)
211+
* [Test](docs/test.md)
212+
213+
Triton can be [built using
214+
Docker](docs/build.md#building-triton-with-docker) or [built without
215+
Docker](docs/build.md#building-triton-without-docker). After building
216+
you should [test Triton](docs/test.md).
217+
218+
It is also possible to [create a Docker image containing a customized
219+
Triton](docs/compose.md) that contains only a subset of the backends.
220+
221+
The Triton project also provides [client libraries for Python and
222+
C++](https://github.com/triton-inference-server/client) that make it
223+
easy to communicate with the server. There are also a large number of
224+
[example clients](https://github.com/triton-inference-server/client)
225+
that demonstrate how to use the libraries. You can also develop your
226+
own clients that directly communicate with Triton using [HTTP/REST or
227+
GRPC protocols](docs/inference_protocols.md). There is also a [C
228+
API](docs/inference_protocols.md) that allows Triton to be linked
229+
directly into your application.
230+
231+
A [Triton backend](https://github.com/triton-inference-server/backend)
232+
is the implementation that executes a model. A backend can interface
233+
with a deep learning framework, like PyTorch, TensorFlow, TensorRT or
234+
ONNX Runtime; or it can interface with a data processing framework
235+
like [DALI](https://github.com/triton-inference-server/dali_backend);
236+
or you can extend Triton by [writing your own
237+
backend](https://github.com/triton-inference-server/backend) in either
238+
[C/C++](https://github.com/triton-inference-server/backend/blob/main/README.md#triton-backend-api)
239+
or
240+
[Python](https://github.com/triton-inference-server/python_backend).
241+
242+
A [Triton repository agent](docs/repository_agents.md) extends Triton
243+
with new functionality that operates when a model is loaded or
244+
unloaded. You can introduce your own code to perform authentication,
245+
decryption, conversion, or similar operations when a model is loaded.
246+
247+
## Papers and Presentation
248+
249+
* [Maximizing Deep Learning Inference Performance with NVIDIA Model
250+
Analyzer](https://developer.nvidia.com/blog/maximizing-deep-learning-inference-performance-with-nvidia-model-analyzer/).
251+
252+
* [High-Performance Inferencing at Scale Using the TensorRT Inference
253+
Server](https://developer.nvidia.com/gtc/2020/video/s22418).
254+
255+
* [Accelerate and Autoscale Deep Learning Inference on GPUs with
256+
KFServing](https://developer.nvidia.com/gtc/2020/video/s22459).
257+
258+
* [Deep into Triton Inference Server: BERT Practical Deployment on
259+
NVIDIA GPU](https://developer.nvidia.com/gtc/2020/video/s21736).
260+
261+
* [Maximizing Utilization for Data Center Inference with TensorRT
262+
Inference Server](https://on-demand-gtc.gputechconf.com/gtcnew/sessionview.php?sessionName=s9438-maximizing+utilization+for+data+center+inference+with+tensorrt+inference+server).
263+
264+
* [NVIDIA TensorRT Inference Server Boosts Deep Learning
265+
Inference](https://devblogs.nvidia.com/nvidia-serves-deep-learning-inference/).
266+
267+
* [GPU-Accelerated Inference for Kubernetes with the NVIDIA TensorRT
268+
Inference Server and
269+
Kubeflow](https://www.kubeflow.org/blog/nvidia_tensorrt/).
270+
271+
* [Deploying NVIDIA Triton at Scale with MIG and Kubernetes](https://developer.nvidia.com/blog/deploying-nvidia-triton-at-scale-with-mig-and-kubernetes/).
272+
273+
## Contributing
274+
275+
Contributions to Triton Inference Server are more than welcome. To
276+
contribute make a pull request and follow the guidelines outlined in
277+
[CONTRIBUTING.md](CONTRIBUTING.md). If you have a backend, client,
278+
example or similar contribution that is not modifying the core of
279+
Triton, then you should file a PR in the [contrib
280+
repo](https://github.com/triton-inference-server/contrib).
281+
282+
## Reporting problems, asking questions
283+
284+
We appreciate any feedback, questions or bug reporting regarding this
285+
project. When help with code is needed, follow the process outlined in
286+
the Stack Overflow (<https://stackoverflow.com/help/mcve>)
287+
document. Ensure posted examples are:
288+
289+
* minimal – use as little code as possible that still produces the
290+
same problem
291+
292+
* complete – provide all parts needed to reproduce the problem. Check
293+
if you can strip external dependency and still show the problem. The
294+
less time we spend on reproducing problems the more time we have to
295+
fix it
296+
297+
* verifiable – test the code you're about to provide to make sure it
298+
reproduces the problem. Remove all other problems that are not
299+
related to your request/question.

0 commit comments

Comments
 (0)