Skip to content

Commit 5e5db26

Browse files
dzierdzier
committed
Update README for 21.10 release
1 parent 84ef605 commit 5e5db26

File tree

1 file changed

+274
-2
lines changed

1 file changed

+274
-2
lines changed

‎README.md

+274-2
Original file line numberDiff line numberDiff line change
@@ -30,5 +30,277 @@
3030

3131
# Triton Inference Server
3232

33-
**NOTE: You are currently on the r21.10 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+
The current release of the Triton Inference Server is 2.14.0 and
42+
corresponds to the 21.09 release of the tritonserver container on
43+
[NVIDIA GPU Cloud (NGC)](https://ngc.nvidia.com). The branch for this
44+
release is
45+
[r21.09](https://github.com/triton-inference-server/server/tree/r21.09).
46+
47+
## What's New in 2.15.0
48+
49+
* [Rate limiter](docs/rate_limiter.md) is now available and manages the rate at
50+
which requests are scheduled on model instances by Triton.
51+
52+
* A beta version of Triton is available for ARM SBSA.
53+
54+
* Windows Triton build now supports HTTP protocol.
55+
56+
* Triton added support for caching responses to inference requests.
57+
58+
* [Sequence IDs](docs/protocol/extension_sequence.md) can now accept strings.
59+
60+
* Container composer tool can generate [CPU-only Triton
61+
containers](docs/compose.md#cpu-only-container-composition).
62+
63+
## Features
64+
65+
* [Multiple deep-learning
66+
frameworks](https://github.com/triton-inference-server/backend). Triton
67+
can manage any number and mix of models (limited by system disk and
68+
memory resources). Triton supports TensorRT, TensorFlow GraphDef,
69+
TensorFlow SavedModel, ONNX, PyTorch TorchScript and OpenVINO model
70+
formats. Both TensorFlow 1.x and TensorFlow 2.x are
71+
supported. Triton also supports TensorFlow-TensorRT and
72+
ONNX-TensorRT integrated models.
73+
74+
* [Concurrent model
75+
execution](docs/architecture.md#concurrent-model-execution). Multiple
76+
models (or multiple instances of the same model) can run
77+
simultaneously on the same GPU or on multiple GPUs.
78+
79+
* [Dynamic batching](docs/architecture.md#models-and-schedulers). For
80+
models that support batching, Triton implements multiple scheduling
81+
and batching algorithms that combine individual inference requests
82+
together to improve inference throughput. These scheduling and
83+
batching decisions are transparent to the client requesting
84+
inference.
85+
86+
* [Extensible
87+
backends](https://github.com/triton-inference-server/backend). In
88+
addition to deep-learning frameworks, Triton provides a *backend
89+
API* that allows Triton to be extended with any model execution
90+
logic implemented in
91+
[Python](https://github.com/triton-inference-server/python_backend)
92+
or
93+
[C++](https://github.com/triton-inference-server/backend/blob/main/README.md#triton-backend-api),
94+
while still benefiting from the CPU and GPU support, concurrent
95+
execution, dynamic batching and other features provided by Triton.
96+
97+
* [Model pipelines](docs/architecture.md#ensemble-models). Triton
98+
*ensembles* represents a pipeline of one or more models and the
99+
connection of input and output tensors between those models. A
100+
single inference request to an ensemble will trigger the execution
101+
of the entire pipeline.
102+
103+
* [HTTP/REST and GRPC inference
104+
protocols](docs/inference_protocols.md) based on the community
105+
developed [KFServing
106+
protocol](https://github.com/kubeflow/kfserving/tree/master/docs/predict-api/v2).
107+
108+
* A [C API](docs/inference_protocols.md#c-api) allows Triton to be
109+
linked directly into your application for edge and other in-process
110+
use cases.
111+
112+
* [Metrics](docs/metrics.md) indicating GPU utilization, server
113+
throughput, and server latency. The metrics are provided in
114+
Prometheus data format.
115+
116+
## Documentation
117+
118+
[Triton Architecture](docs/architecture.md) gives a high-level
119+
overview of the structure and capabilities of the inference
120+
server. There is also an [FAQ](docs/faq.md). Additional documentation
121+
is divided into [*user*](#user-documentation) and
122+
[*developer*](#developer-documentation) sections. The *user*
123+
documentation describes how to use Triton as an inference solution,
124+
including information on how to configure Triton, how to organize and
125+
configure your models, how to use the C++ and Python clients, etc. The
126+
*developer* documentation describes how to build and test Triton and
127+
also how Triton can be extended with new functionality.
128+
129+
The Triton [Release
130+
Notes](https://docs.nvidia.com/deeplearning/triton-inference-server/release-notes/index.html)
131+
and [Support
132+
Matrix](https://docs.nvidia.com/deeplearning/dgx/support-matrix/index.html)
133+
indicate the required versions of the NVIDIA Driver and CUDA, and also
134+
describe supported GPUs.
135+
136+
### User Documentation
137+
138+
* [QuickStart](docs/quickstart.md)
139+
* [Install](docs/quickstart.md#install-triton-docker-image)
140+
* [Run](docs/quickstart.md#run-triton)
141+
* [Model Repository](docs/model_repository.md)
142+
* [Model Configuration](docs/model_configuration.md)
143+
* [Model Management](docs/model_management.md)
144+
* [Custom Operations](docs/custom_operations.md)
145+
* [Client Libraries and Examples](https://github.com/triton-inference-server/client)
146+
* [Optimization](docs/optimization.md)
147+
* [Model Analyzer](docs/model_analyzer.md)
148+
* [Performance Analyzer](docs/perf_analyzer.md)
149+
* [Metrics](docs/metrics.md)
150+
* [Rate Limiter](docs/rate_limiter.md)
151+
* [Jetson and JetPack](docs/jetson.md)
152+
153+
The [quickstart](docs/quickstart.md) walks you through all the steps
154+
required to install and run Triton with an example image
155+
classification model and then use an example client application to
156+
perform inferencing using that model. The quickstart also demonstrates
157+
how [Triton supports both GPU systems and CPU-only
158+
systems](docs/quickstart.md#run-triton).
159+
160+
The first step in using Triton to serve your models is to place one or
161+
more models into a [model
162+
repository](docs/model_repository.md). Optionally, depending on the type
163+
of the model and on what Triton capabilities you want to enable for
164+
the model, you may need to create a [model
165+
configuration](docs/model_configuration.md) for the model. If your
166+
model has [custom operations](docs/custom_operations.md) you will need
167+
to make sure they are loaded correctly by Triton.
168+
169+
After you have your model(s) available in Triton, you will want to
170+
send inference and other requests to Triton from your *client*
171+
application. The [Python and C++ client
172+
libraries](https://github.com/triton-inference-server/client) provide
173+
APIs to simplify this communication. There are also a large number of
174+
[client examples](https://github.com/triton-inference-server/client)
175+
that demonstrate how to use the libraries. You can also send
176+
HTTP/REST requests directly to Triton using the [HTTP/REST JSON-based
177+
protocol](docs/inference_protocols.md#httprest-and-grpc-protocols) or
178+
[generate a GRPC client for many other
179+
languages](https://github.com/triton-inference-server/client).
180+
181+
Understanding and [optimizing performance](docs/optimization.md) is an
182+
important part of deploying your models. The Triton project provides
183+
the [Performance Analyzer](docs/perf_analyzer.md) and the [Model
184+
Analyzer](docs/model_analyzer.md) to help your optimization
185+
efforts. Specifically, you will want to optimize [scheduling and
186+
batching](docs/architecture.md#models-and-schedulers) and [model
187+
instances](docs/model_configuration.md#instance-groups) appropriately
188+
for each model. You can also enable cross-model prioritization using
189+
[rate limiter](docs/rate_limiter.md) which manages the rate at which
190+
requests are scheduled on model instances. You may also want to
191+
consider [ensembling multiple models and pre/post-processing](docs/architecture.md#ensemble-models)
192+
into a pipeline. In some cases you may find [individual inference
193+
request trace data](docs/trace.md) useful when optimizing. A
194+
[Prometheus metrics endpoint](docs/metrics.md) allows you to visualize
195+
and monitor aggregate inference metrics.
196+
197+
NVIDIA publishes a number of [deep learning
198+
examples](https://github.com/NVIDIA/DeepLearningExamples) that use
199+
Triton.
200+
201+
As part of your deployment strategy you may want to [explicitly manage
202+
what models are available by loading and unloading
203+
models](docs/model_management.md) from a running Triton server. If you
204+
are using Kubernetes for deployment there are simple examples of how
205+
to deploy Triton using Kubernetes and Helm, one for
206+
[GCP](deploy/gcp/README.md) and one for [AWS](deploy/aws/README.md).
207+
208+
The [version 1 to version 2 migration
209+
information](docs/v1_to_v2.md) is helpful if you are moving to
210+
version 2 of Triton from previously using version 1.
211+
212+
### Developer Documentation
213+
214+
* [Build](docs/build.md)
215+
* [Protocols and APIs](docs/inference_protocols.md).
216+
* [Backends](https://github.com/triton-inference-server/backend)
217+
* [Repository Agents](docs/repository_agents.md)
218+
* [Test](docs/test.md)
219+
220+
Triton can be [built using
221+
Docker](docs/build.md#building-triton-with-docker) or [built without
222+
Docker](docs/build.md#building-triton-without-docker). After building
223+
you should [test Triton](docs/test.md).
224+
225+
It is also possible to [create a Docker image containing a customized
226+
Triton](docs/compose.md) that contains only a subset of the backends.
227+
228+
The Triton project also provides [client libraries for Python and
229+
C++](https://github.com/triton-inference-server/client) that make it
230+
easy to communicate with the server. There are also a large number of
231+
[example clients](https://github.com/triton-inference-server/client)
232+
that demonstrate how to use the libraries. You can also develop your
233+
own clients that directly communicate with Triton using [HTTP/REST or
234+
GRPC protocols](docs/inference_protocols.md). There is also a [C
235+
API](docs/inference_protocols.md) that allows Triton to be linked
236+
directly into your application.
237+
238+
A [Triton backend](https://github.com/triton-inference-server/backend)
239+
is the implementation that executes a model. A backend can interface
240+
with a deep learning framework, like PyTorch, TensorFlow, TensorRT or
241+
ONNX Runtime; or it can interface with a data processing framework
242+
like [DALI](https://github.com/triton-inference-server/dali_backend);
243+
or you can extend Triton by [writing your own
244+
backend](https://github.com/triton-inference-server/backend) in either
245+
[C/C++](https://github.com/triton-inference-server/backend/blob/main/README.md#triton-backend-api)
246+
or
247+
[Python](https://github.com/triton-inference-server/python_backend).
248+
249+
A [Triton repository agent](docs/repository_agents.md) extends Triton
250+
with new functionality that operates when a model is loaded or
251+
unloaded. You can introduce your own code to perform authentication,
252+
decryption, conversion, or similar operations when a model is loaded.
253+
254+
## Papers and Presentation
255+
256+
* [Maximizing Deep Learning Inference Performance with NVIDIA Model
257+
Analyzer](https://developer.nvidia.com/blog/maximizing-deep-learning-inference-performance-with-nvidia-model-analyzer/).
258+
259+
* [High-Performance Inferencing at Scale Using the TensorRT Inference
260+
Server](https://developer.nvidia.com/gtc/2020/video/s22418).
261+
262+
* [Accelerate and Autoscale Deep Learning Inference on GPUs with
263+
KFServing](https://developer.nvidia.com/gtc/2020/video/s22459).
264+
265+
* [Deep into Triton Inference Server: BERT Practical Deployment on
266+
NVIDIA GPU](https://developer.nvidia.com/gtc/2020/video/s21736).
267+
268+
* [Maximizing Utilization for Data Center Inference with TensorRT
269+
Inference Server](https://on-demand-gtc.gputechconf.com/gtcnew/sessionview.php?sessionName=s9438-maximizing+utilization+for+data+center+inference+with+tensorrt+inference+server).
270+
271+
* [NVIDIA TensorRT Inference Server Boosts Deep Learning
272+
Inference](https://devblogs.nvidia.com/nvidia-serves-deep-learning-inference/).
273+
274+
* [GPU-Accelerated Inference for Kubernetes with the NVIDIA TensorRT
275+
Inference Server and
276+
Kubeflow](https://www.kubeflow.org/blog/nvidia_tensorrt/).
277+
278+
* [Deploying NVIDIA Triton at Scale with MIG and Kubernetes](https://developer.nvidia.com/blog/deploying-nvidia-triton-at-scale-with-mig-and-kubernetes/).
279+
280+
## Contributing
281+
282+
Contributions to Triton Inference Server are more than welcome. To
283+
contribute make a pull request and follow the guidelines outlined in
284+
[CONTRIBUTING.md](CONTRIBUTING.md). If you have a backend, client,
285+
example or similar contribution that is not modifying the core of
286+
Triton, then you should file a PR in the [contrib
287+
repo](https://github.com/triton-inference-server/contrib).
288+
289+
## Reporting problems, asking questions
290+
291+
We appreciate any feedback, questions or bug reporting regarding this
292+
project. When help with code is needed, follow the process outlined in
293+
the Stack Overflow (<https://stackoverflow.com/help/mcve>)
294+
document. Ensure posted examples are:
295+
296+
* minimal – use as little code as possible that still produces the
297+
same problem
298+
299+
* complete – provide all parts needed to reproduce the problem. Check
300+
if you can strip external dependency and still show the problem. The
301+
less time we spend on reproducing problems the more time we have to
302+
fix it
303+
304+
* verifiable – test the code you're about to provide to make sure it
305+
reproduces the problem. Remove all other problems that are not
306+
related to your request/question.

0 commit comments

Comments
 (0)