Skip to content

Commit f1fb852

Browse files
dzierdzier
committed
Update README for 21.11 release
1 parent 3dfb68e commit f1fb852

File tree

1 file changed

+265
-2
lines changed

1 file changed

+265
-2
lines changed

‎README.md

+265-2
Original file line numberDiff line numberDiff line change
@@ -30,5 +30,268 @@
3030

3131
# Triton Inference Server
3232

33-
**NOTE: You are currently on the r21.11 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 2.16.0
42+
43+
* Added support for [LightGBM models with categorical
44+
features](https://github.com/triton-inference-server/fil_backend/tree/r21.11#categorical-feature-support)
45+
in FIL backend.
46+
47+
* Added [Jetson examples](docs/examples/jetson) in documentation.
48+
49+
* Completed proof of concept of [Inferentia
50+
support](https://github.com/triton-inference-server/python_backend/tree/r21.11/inferentia#readme).
51+
52+
* Added ARM Support for Model Analyzer.
53+
54+
## Features
55+
56+
* [Multiple deep-learning
57+
frameworks](https://github.com/triton-inference-server/backend). Triton
58+
can manage any number and mix of models (limited by system disk and
59+
memory resources). Triton supports TensorRT, TensorFlow GraphDef,
60+
TensorFlow SavedModel, ONNX, PyTorch TorchScript and OpenVINO model
61+
formats. Both TensorFlow 1.x and TensorFlow 2.x are
62+
supported. Triton also supports TensorFlow-TensorRT and
63+
ONNX-TensorRT integrated models.
64+
65+
* [Concurrent model
66+
execution](docs/architecture.md#concurrent-model-execution). Multiple
67+
models (or multiple instances of the same model) can run
68+
simultaneously on the same GPU or on multiple GPUs.
69+
70+
* [Dynamic batching](docs/architecture.md#models-and-schedulers). For
71+
models that support batching, Triton implements multiple scheduling
72+
and batching algorithms that combine individual inference requests
73+
together to improve inference throughput. These scheduling and
74+
batching decisions are transparent to the client requesting
75+
inference.
76+
77+
* [Extensible
78+
backends](https://github.com/triton-inference-server/backend). In
79+
addition to deep-learning frameworks, Triton provides a *backend
80+
API* that allows Triton to be extended with any model execution
81+
logic implemented in
82+
[Python](https://github.com/triton-inference-server/python_backend)
83+
or
84+
[C++](https://github.com/triton-inference-server/backend/blob/main/README.md#triton-backend-api),
85+
while still benefiting from the CPU and GPU support, concurrent
86+
execution, dynamic batching and other features provided by Triton.
87+
88+
* [Model pipelines](docs/architecture.md#ensemble-models). Triton
89+
*ensembles* represents a pipeline of one or more models and the
90+
connection of input and output tensors between those models. A
91+
single inference request to an ensemble will trigger the execution
92+
of the entire pipeline.
93+
94+
* [HTTP/REST and GRPC inference
95+
protocols](docs/inference_protocols.md) based on the community
96+
developed [KFServing
97+
protocol](https://github.com/kubeflow/kfserving/tree/master/docs/predict-api/v2).
98+
99+
* A [C API](docs/inference_protocols.md#c-api) allows Triton to be
100+
linked directly into your application for edge and other in-process
101+
use cases.
102+
103+
* [Metrics](docs/metrics.md) indicating GPU utilization, server
104+
throughput, and server latency. The metrics are provided in
105+
Prometheus data format.
106+
107+
## Documentation
108+
109+
[Triton Architecture](docs/architecture.md) gives a high-level
110+
overview of the structure and capabilities of the inference
111+
server. There is also an [FAQ](docs/faq.md). Additional documentation
112+
is divided into [*user*](#user-documentation) and
113+
[*developer*](#developer-documentation) sections. The *user*
114+
documentation describes how to use Triton as an inference solution,
115+
including information on how to configure Triton, how to organize and
116+
configure your models, how to use the C++ and Python clients, etc. The
117+
*developer* documentation describes how to build and test Triton and
118+
also how Triton can be extended with new functionality.
119+
120+
The Triton [Release
121+
Notes](https://docs.nvidia.com/deeplearning/triton-inference-server/release-notes/index.html)
122+
and [Support
123+
Matrix](https://docs.nvidia.com/deeplearning/dgx/support-matrix/index.html)
124+
indicate the required versions of the NVIDIA Driver and CUDA, and also
125+
describe supported GPUs.
126+
127+
### User Documentation
128+
129+
* [QuickStart](docs/quickstart.md)
130+
* [Install](docs/quickstart.md#install-triton-docker-image)
131+
* [Run](docs/quickstart.md#run-triton)
132+
* [Model Repository](docs/model_repository.md)
133+
* [Model Configuration](docs/model_configuration.md)
134+
* [Model Management](docs/model_management.md)
135+
* [Custom Operations](docs/custom_operations.md)
136+
* [Client Libraries and Examples](https://github.com/triton-inference-server/client)
137+
* [Optimization](docs/optimization.md)
138+
* [Model Analyzer](docs/model_analyzer.md)
139+
* [Performance Analyzer](docs/perf_analyzer.md)
140+
* [Metrics](docs/metrics.md)
141+
* [Rate Limiter](docs/rate_limiter.md)
142+
* [Jetson and JetPack](docs/jetson.md)
143+
144+
The [quickstart](docs/quickstart.md) walks you through all the steps
145+
required to install and run Triton with an example image
146+
classification model and then use an example client application to
147+
perform inferencing using that model. The quickstart also demonstrates
148+
how [Triton supports both GPU systems and CPU-only
149+
systems](docs/quickstart.md#run-triton).
150+
151+
The first step in using Triton to serve your models is to place one or
152+
more models into a [model
153+
repository](docs/model_repository.md). Optionally, depending on the type
154+
of the model and on what Triton capabilities you want to enable for
155+
the model, you may need to create a [model
156+
configuration](docs/model_configuration.md) for the model. If your
157+
model has [custom operations](docs/custom_operations.md) you will need
158+
to make sure they are loaded correctly by Triton.
159+
160+
After you have your model(s) available in Triton, you will want to
161+
send inference and other requests to Triton from your *client*
162+
application. The [Python and C++ client
163+
libraries](https://github.com/triton-inference-server/client) provide
164+
APIs to simplify this communication. There are also a large number of
165+
[client examples](https://github.com/triton-inference-server/client)
166+
that demonstrate how to use the libraries. You can also send
167+
HTTP/REST requests directly to Triton using the [HTTP/REST JSON-based
168+
protocol](docs/inference_protocols.md#httprest-and-grpc-protocols) or
169+
[generate a GRPC client for many other
170+
languages](https://github.com/triton-inference-server/client).
171+
172+
Understanding and [optimizing performance](docs/optimization.md) is an
173+
important part of deploying your models. The Triton project provides
174+
the [Performance Analyzer](docs/perf_analyzer.md) and the [Model
175+
Analyzer](docs/model_analyzer.md) to help your optimization
176+
efforts. Specifically, you will want to optimize [scheduling and
177+
batching](docs/architecture.md#models-and-schedulers) and [model
178+
instances](docs/model_configuration.md#instance-groups) appropriately
179+
for each model. You can also enable cross-model prioritization using
180+
[rate limiter](docs/rate_limiter.md) which manages the rate at which
181+
requests are scheduled on model instances. You may also want to
182+
consider [ensembling multiple models and pre/post-processing](docs/architecture.md#ensemble-models)
183+
into a pipeline. In some cases you may find [individual inference
184+
request trace data](docs/trace.md) useful when optimizing. A
185+
[Prometheus metrics endpoint](docs/metrics.md) allows you to visualize
186+
and monitor aggregate inference metrics.
187+
188+
NVIDIA publishes a number of [deep learning
189+
examples](https://github.com/NVIDIA/DeepLearningExamples) that use
190+
Triton.
191+
192+
As part of your deployment strategy you may want to [explicitly manage
193+
what models are available by loading and unloading
194+
models](docs/model_management.md) from a running Triton server. If you
195+
are using Kubernetes for deployment there are simple examples of how
196+
to deploy Triton using Kubernetes and Helm, one for
197+
[GCP](deploy/gcp/README.md) and one for [AWS](deploy/aws/README.md).
198+
199+
The [version 1 to version 2 migration
200+
information](docs/v1_to_v2.md) is helpful if you are moving to
201+
version 2 of Triton from previously using version 1.
202+
203+
### Developer Documentation
204+
205+
* [Build](docs/build.md)
206+
* [Protocols and APIs](docs/inference_protocols.md).
207+
* [Backends](https://github.com/triton-inference-server/backend)
208+
* [Repository Agents](docs/repository_agents.md)
209+
* [Test](docs/test.md)
210+
211+
Triton can be [built using
212+
Docker](docs/build.md#building-triton-with-docker) or [built without
213+
Docker](docs/build.md#building-triton-without-docker). After building
214+
you should [test Triton](docs/test.md).
215+
216+
It is also possible to [create a Docker image containing a customized
217+
Triton](docs/compose.md) that contains only a subset of the backends.
218+
219+
The Triton project also provides [client libraries for Python and
220+
C++](https://github.com/triton-inference-server/client) that make it
221+
easy to communicate with the server. There are also a large number of
222+
[example clients](https://github.com/triton-inference-server/client)
223+
that demonstrate how to use the libraries. You can also develop your
224+
own clients that directly communicate with Triton using [HTTP/REST or
225+
GRPC protocols](docs/inference_protocols.md). There is also a [C
226+
API](docs/inference_protocols.md) that allows Triton to be linked
227+
directly into your application.
228+
229+
A [Triton backend](https://github.com/triton-inference-server/backend)
230+
is the implementation that executes a model. A backend can interface
231+
with a deep learning framework, like PyTorch, TensorFlow, TensorRT or
232+
ONNX Runtime; or it can interface with a data processing framework
233+
like [DALI](https://github.com/triton-inference-server/dali_backend);
234+
or you can extend Triton by [writing your own
235+
backend](https://github.com/triton-inference-server/backend) in either
236+
[C/C++](https://github.com/triton-inference-server/backend/blob/main/README.md#triton-backend-api)
237+
or
238+
[Python](https://github.com/triton-inference-server/python_backend).
239+
240+
A [Triton repository agent](docs/repository_agents.md) extends Triton
241+
with new functionality that operates when a model is loaded or
242+
unloaded. You can introduce your own code to perform authentication,
243+
decryption, conversion, or similar operations when a model is loaded.
244+
245+
## Papers and Presentation
246+
247+
* [Maximizing Deep Learning Inference Performance with NVIDIA Model
248+
Analyzer](https://developer.nvidia.com/blog/maximizing-deep-learning-inference-performance-with-nvidia-model-analyzer/).
249+
250+
* [High-Performance Inferencing at Scale Using the TensorRT Inference
251+
Server](https://developer.nvidia.com/gtc/2020/video/s22418).
252+
253+
* [Accelerate and Autoscale Deep Learning Inference on GPUs with
254+
KFServing](https://developer.nvidia.com/gtc/2020/video/s22459).
255+
256+
* [Deep into Triton Inference Server: BERT Practical Deployment on
257+
NVIDIA GPU](https://developer.nvidia.com/gtc/2020/video/s21736).
258+
259+
* [Maximizing Utilization for Data Center Inference with TensorRT
260+
Inference Server](https://on-demand-gtc.gputechconf.com/gtcnew/sessionview.php?sessionName=s9438-maximizing+utilization+for+data+center+inference+with+tensorrt+inference+server).
261+
262+
* [NVIDIA TensorRT Inference Server Boosts Deep Learning
263+
Inference](https://devblogs.nvidia.com/nvidia-serves-deep-learning-inference/).
264+
265+
* [GPU-Accelerated Inference for Kubernetes with the NVIDIA TensorRT
266+
Inference Server and
267+
Kubeflow](https://www.kubeflow.org/blog/nvidia_tensorrt/).
268+
269+
* [Deploying NVIDIA Triton at Scale with MIG and Kubernetes](https://developer.nvidia.com/blog/deploying-nvidia-triton-at-scale-with-mig-and-kubernetes/).
270+
271+
## Contributing
272+
273+
Contributions to Triton Inference Server are more than welcome. To
274+
contribute make a pull request and follow the guidelines outlined in
275+
[CONTRIBUTING.md](CONTRIBUTING.md). If you have a backend, client,
276+
example or similar contribution that is not modifying the core of
277+
Triton, then you should file a PR in the [contrib
278+
repo](https://github.com/triton-inference-server/contrib).
279+
280+
## Reporting problems, asking questions
281+
282+
We appreciate any feedback, questions or bug reporting regarding this
283+
project. When help with code is needed, follow the process outlined in
284+
the Stack Overflow (<https://stackoverflow.com/help/mcve>)
285+
document. Ensure posted examples are:
286+
287+
* minimal – use as little code as possible that still produces the
288+
same problem
289+
290+
* complete – provide all parts needed to reproduce the problem. Check
291+
if you can strip external dependency and still show the problem. The
292+
less time we spend on reproducing problems the more time we have to
293+
fix it
294+
295+
* verifiable – test the code you're about to provide to make sure it
296+
reproduces the problem. Remove all other problems that are not
297+
related to your request/question.

0 commit comments

Comments
 (0)