Skip to content

Commit 6733c3f

Browse files
committed
Update README for 21.07 release
1 parent 731ea66 commit 6733c3f

File tree

1 file changed

+262
-2
lines changed

1 file changed

+262
-2
lines changed

README.md

+262-2
Original file line numberDiff line numberDiff line change
@@ -30,5 +30,265 @@
3030

3131
# Triton Inference Server
3232

33-
**NOTE: You are currently on the r21.07 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 2.12.0
42+
43+
* Added support for CPU in RAPIDS FIL Backend.
44+
45+
* Inference requests using the C API are now allowed to provide multiple copies
46+
of an input tensor in different memories. Triton will choose the most
47+
performant copy to use depending on where the inference request is executed.
48+
49+
* For ONNX models using TensorRT acceleration, the tensorrt_accelerator option
50+
in the model configuration can now specify precision and workspace size.
51+
https://github.com/triton-inference-server/server/blob/main/docs/optimization.md#onnx-with-tensorrt-optimization
52+
53+
* Model Analyzer added an offline mode, which prioritizes throughput over
54+
latency for offline inferencing scenarios. A new set of reports and graphs are
55+
created to better analyze the offline use case.
56+
57+
## Features
58+
59+
* [Multiple deep-learning
60+
frameworks](https://github.com/triton-inference-server/backend). Triton
61+
can manage any number and mix of models (limited by system disk and
62+
memory resources). Triton supports TensorRT, TensorFlow GraphDef,
63+
TensorFlow SavedModel, ONNX, PyTorch TorchScript and OpenVINO model
64+
formats. Both TensorFlow 1.x and TensorFlow 2.x are
65+
supported. Triton also supports TensorFlow-TensorRT and
66+
ONNX-TensorRT integrated models.
67+
68+
* [Concurrent model
69+
execution](docs/architecture.md#concurrent-model-execution). Multiple
70+
models (or multiple instances of the same model) can run
71+
simultaneously on the same GPU or on multiple GPUs.
72+
73+
* [Dynamic batching](docs/architecture.md#models-and-schedulers). For
74+
models that support batching, Triton implements multiple scheduling
75+
and batching algorithms that combine individual inference requests
76+
together to improve inference throughput. These scheduling and
77+
batching decisions are transparent to the client requesting
78+
inference.
79+
80+
* [Extensible
81+
backends](https://github.com/triton-inference-server/backend). In
82+
addition to deep-learning frameworks, Triton provides a *backend
83+
API* that allows Triton to be extended with any model execution
84+
logic implemented in
85+
[Python](https://github.com/triton-inference-server/python_backend)
86+
or
87+
[C++](https://github.com/triton-inference-server/backend/blob/main/README.md#triton-backend-api),
88+
while still benefiting from the CPU and GPU support, concurrent
89+
execution, dynamic batching and other features provided by Triton.
90+
91+
* [Model pipelines](docs/architecture.md#ensemble-models). Triton
92+
*ensembles* represents a pipeline of one or more models and the
93+
connection of input and output tensors between those models. A
94+
single inference request to an ensemble will trigger the execution
95+
of the entire pipeline.
96+
97+
* [HTTP/REST and GRPC inference
98+
protocols](docs/inference_protocols.md) based on the community
99+
developed [KFServing
100+
protocol](https://github.com/kubeflow/kfserving/tree/master/docs/predict-api/v2).
101+
102+
* A [C API](docs/inference_protocols.md#c-api) allows Triton to be
103+
linked directly into your application for edge and other in-process
104+
use cases.
105+
106+
* [Metrics](docs/metrics.md) indicating GPU utilization, server
107+
throughput, and server latency. The metrics are provided in
108+
Prometheus data format.
109+
110+
## Documentation
111+
112+
[Triton Architecture](docs/architecture.md) gives a high-level
113+
overview of the structure and capabilities of the inference
114+
server. There is also an [FAQ](docs/faq.md). Additional documentation
115+
is divided into [*user*](#user-documentation) and
116+
[*developer*](#developer-documentation) sections. The *user*
117+
documentation describes how to use Triton as an inference solution,
118+
including information on how to configure Triton, how to organize and
119+
configure your models, how to use the C++ and Python clients, etc. The
120+
*developer* documentation describes how to build and test Triton and
121+
also how Triton can be extended with new functionality.
122+
123+
The Triton [Release
124+
Notes](https://docs.nvidia.com/deeplearning/triton-inference-server/release-notes/index.html)
125+
and [Support
126+
Matrix](https://docs.nvidia.com/deeplearning/dgx/support-matrix/index.html)
127+
indicate the required versions of the NVIDIA Driver and CUDA, and also
128+
describe supported GPUs.
129+
130+
### User Documentation
131+
132+
- [QuickStart](docs/quickstart.md)
133+
- [Install](docs/quickstart.md#install-triton-docker-image)
134+
- [Run](docs/quickstart.md#run-triton)
135+
- [Model Repository](docs/model_repository.md)
136+
- [Model Configuration](docs/model_configuration.md)
137+
- [Model Management](docs/model_management.md)
138+
- [Custom Operations](docs/custom_operations.md)
139+
- [Client Libraries and Examples](https://github.com/triton-inference-server/client)
140+
- [Optimization](docs/optimization.md)
141+
- [Model Analyzer](docs/model_analyzer.md)
142+
- [Performance Analyzer](docs/perf_analyzer.md)
143+
- [Metrics](docs/metrics.md)
144+
145+
The [quickstart](docs/quickstart.md) walks you through all the steps
146+
required to install and run Triton with an example image
147+
classification model and then use an example client application to
148+
perform inferencing using that model. The quickstart also demonstrates
149+
how [Triton supports both GPU systems and CPU-only
150+
systems](docs/quickstart.md#run-triton).
151+
152+
The first step in using Triton to serve your models is to place one or
153+
more models into a [model
154+
repository](docs/model_repository.md). Optionally, depending on the type
155+
of the model and on what Triton capabilities you want to enable for
156+
the model, you may need to create a [model
157+
configuration](docs/model_configuration.md) for the model. If your
158+
model has [custom operations](docs/custom_operations.md) you will need
159+
to make sure they are loaded correctly by Triton.
160+
161+
After you have your model(s) available in Triton, you will want to
162+
send inference and other requests to Triton from your *client*
163+
application. The [Python and C++ client
164+
libraries](https://github.com/triton-inference-server/client) provide
165+
APIs to simplify this communication. There are also a large number of
166+
[client examples](https://github.com/triton-inference-server/client)
167+
that demonstrate how to use the libraries. You can also send
168+
HTTP/REST requests directly to Triton using the [HTTP/REST JSON-based
169+
protocol](docs/inference_protocols.md#httprest-and-grpc-protocols) or
170+
[generate a GRPC client for many other
171+
languages](https://github.com/triton-inference-server/client).
172+
173+
Understanding and [optimizing performance](docs/optimization.md) is an
174+
important part of deploying your models. The Triton project provides
175+
the [Performance Analyzer](docs/perf_analyzer.md) and the [Model
176+
Analyzer](docs/model_analyzer.md) to help your optimization
177+
efforts. Specifically, you will want to optimize [scheduling and
178+
batching](docs/architecture.md#models-and-schedulers) and [model
179+
instances](docs/model_configuration.md#instance-groups) appropriately
180+
for each model. You may also want to consider [ensembling multiple
181+
models and pre/post-processing](docs/architecture.md#ensemble-models)
182+
into a pipeline. In some cases you may find [individual inference
183+
request trace data](docs/trace.md) useful when optimizing. A
184+
[Prometheus metrics endpoint](docs/metrics.md) allows you to visualize
185+
and monitor aggregate inference metrics.
186+
187+
NVIDIA publishes a number of [deep learning
188+
examples](https://github.com/NVIDIA/DeepLearningExamples) that use
189+
Triton.
190+
191+
As part of your deployment strategy you may want to [explicitly manage
192+
what models are available by loading and unloading
193+
models](docs/model_management.md) from a running Triton server. If you
194+
are using Kubernetes for deployment there are simple examples of how
195+
to deploy Triton using Kubernetes and Helm, one for
196+
[GCP](deploy/gcp/README.md) and one for [AWS](deploy/aws/README.md).
197+
198+
The [version 1 to version 2 migration
199+
information](docs/v1_to_v2.md) is helpful if you are moving to
200+
version 2 of Triton from previously using version 1.
201+
202+
### Developer Documentation
203+
204+
- [Build](docs/build.md)
205+
- [Protocols and APIs](docs/inference_protocols.md).
206+
- [Backends](https://github.com/triton-inference-server/backend)
207+
- [Repository Agents](docs/repository_agents.md)
208+
- [Test](docs/test.md)
209+
210+
Triton can be [built using
211+
Docker](docs/build.md#building-triton-with-docker) or [built without
212+
Docker](docs/build.md#building-triton-without-docker). After building
213+
you should [test Triton](docs/test.md).
214+
215+
It is also possible to [create a Docker image containing a customized
216+
Triton](docs/compose.md) that contains only a subset of the backends.
217+
218+
The Triton project also provides [client libraries for Python and
219+
C++](https://github.com/triton-inference-server/client) that make it
220+
easy to communicate with the server. There are also a large number of
221+
[example clients](https://github.com/triton-inference-server/client)
222+
that demonstrate how to use the libraries. You can also develop your
223+
own clients that directly communicate with Triton using [HTTP/REST or
224+
GRPC protocols](docs/inference_protocols.md). There is also a [C
225+
API](docs/inference_protocols.md) that allows Triton to be linked
226+
directly into your application.
227+
228+
A [Triton backend](https://github.com/triton-inference-server/backend)
229+
is the implementation that executes a model. A backend can interface
230+
with a deep learning framework, like PyTorch, TensorFlow, TensorRT or
231+
ONNX Runtime; or it can interface with a data processing framework
232+
like [DALI](https://github.com/triton-inference-server/dali_backend);
233+
or you can extend Triton by [writing your own
234+
backend](https://github.com/triton-inference-server/backend) in either
235+
[C/C++](https://github.com/triton-inference-server/backend/blob/main/README.md#triton-backend-api)
236+
or
237+
[Python](https://github.com/triton-inference-server/python_backend).
238+
239+
A [Triton repository agent](docs/repository_agents.md) extends Triton
240+
with new functionality that operates when a model is loaded or
241+
unloaded. You can introduce your own code to perform authentication,
242+
decryption, conversion, or similar operations when a model is loaded.
243+
244+
## Papers and Presentation
245+
246+
* [Maximizing Deep Learning Inference Performance with NVIDIA Model
247+
Analyzer](https://developer.nvidia.com/blog/maximizing-deep-learning-inference-performance-with-nvidia-model-analyzer/).
248+
249+
* [High-Performance Inferencing at Scale Using the TensorRT Inference
250+
Server](https://developer.nvidia.com/gtc/2020/video/s22418).
251+
252+
* [Accelerate and Autoscale Deep Learning Inference on GPUs with
253+
KFServing](https://developer.nvidia.com/gtc/2020/video/s22459).
254+
255+
* [Deep into Triton Inference Server: BERT Practical Deployment on
256+
NVIDIA GPU](https://developer.nvidia.com/gtc/2020/video/s21736).
257+
258+
* [Maximizing Utilization for Data Center Inference with TensorRT
259+
Inference Server](https://on-demand-gtc.gputechconf.com/gtcnew/sessionview.php?sessionName=s9438-maximizing+utilization+for+data+center+inference+with+tensorrt+inference+server).
260+
261+
* [NVIDIA TensorRT Inference Server Boosts Deep Learning
262+
Inference](https://devblogs.nvidia.com/nvidia-serves-deep-learning-inference/).
263+
264+
* [GPU-Accelerated Inference for Kubernetes with the NVIDIA TensorRT
265+
Inference Server and
266+
Kubeflow](https://www.kubeflow.org/blog/nvidia_tensorrt/).
267+
268+
## Contributing
269+
270+
Contributions to Triton Inference Server are more than welcome. To
271+
contribute make a pull request and follow the guidelines outlined in
272+
[CONTRIBUTING.md](CONTRIBUTING.md). If you have a backend, client,
273+
example or similar contribution that is not modifying the core of
274+
Triton, then you should file a PR in the [contrib
275+
repo](https://github.com/triton-inference-server/contrib).
276+
277+
## Reporting problems, asking questions
278+
279+
We appreciate any feedback, questions or bug reporting regarding this
280+
project. When help with code is needed, follow the process outlined in
281+
the Stack Overflow (https://stackoverflow.com/help/mcve)
282+
document. Ensure posted examples are:
283+
284+
* minimal – use as little code as possible that still produces the
285+
same problem
286+
287+
* complete – provide all parts needed to reproduce the problem. Check
288+
if you can strip external dependency and still show the problem. The
289+
less time we spend on reproducing problems the more time we have to
290+
fix it
291+
292+
* verifiable – test the code you're about to provide to make sure it
293+
reproduces the problem. Remove all other problems that are not
294+
related to your request/question.

0 commit comments

Comments
 (0)