Skip to content

Commit f6ad9ca

Browse files
dzierdzier
committed
Update README for 22.02 release
1 parent 3a18515 commit f6ad9ca

File tree

1 file changed

+332
-2
lines changed

1 file changed

+332
-2
lines changed

‎README.md

+332-2
Original file line numberDiff line numberDiff line change
@@ -30,5 +30,335 @@
3030

3131
# Triton Inference Server
3232

33-
**NOTE: You are currently on the r22.02 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 2.19.0
42+
43+
* Enabled [full-fledged SSL support in HTTP C++ client
44+
library](https://github.com/triton-inference-server/client/tree/r22.02#ssltls).
45+
46+
* New [cache metrics](docs/metrics.md) added to Prometheus metrics.
47+
48+
* PyTorch Backend now supports [passing inputs in the form of a dictionary of
49+
tensors](docs/model_configuration.md#special-conventions-for-pytorch-backend).
50+
51+
## Features
52+
53+
* [Deep learning
54+
frameworks](https://github.com/triton-inference-server/backend).
55+
Triton supports TensorRT, TensorFlow GraphDef, TensorFlow
56+
SavedModel, ONNX, PyTorch TorchScript and OpenVINO model
57+
formats. Both TensorFlow 1.x and TensorFlow 2.x are
58+
supported. Triton also supports TensorFlow-TensorRT, ONNX-TensorRT
59+
and PyTorch-TensorRT integrated models.
60+
61+
* [Machine learning
62+
frameworks](https://github.com/triton-inference-server/fil_backend).
63+
Triton supports popular machine learning frameworks such as XGBoost,
64+
LightGBM, Scikit-Learn and cuML using the [RAPIDS Forest Inference
65+
Library](https://medium.com/rapids-ai/rapids-forest-inference-library-prediction-at-100-million-rows-per-second-19558890bc35).
66+
67+
* [Concurrent model
68+
execution](docs/architecture.md#concurrent-model-execution). Triton
69+
can simultaneously run multiple models (or multiple instances of the
70+
same model) using the same or different deep-learning and
71+
machine-learning frameworks.
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 full CPU and GPU support, concurrent
89+
execution, dynamic batching and other features provided by Triton.
90+
91+
* Model pipelines using
92+
[Ensembling](docs/architecture.md#ensemble-models) or [Business
93+
Logic Scripting
94+
(BLS)](https://github.com/triton-inference-server/python_backend#business-logic-scripting).
95+
A Triton *ensemble* represents a pipeline of one or more models and
96+
the connection of input and output tensors between those
97+
models. *BLS* allows a pipeline along with extra business logic to
98+
be represented in Python. In both cases a single inference request
99+
will trigger the execution of the entire pipeline.
100+
101+
* [HTTP/REST and GRPC inference
102+
protocols](docs/inference_protocols.md) based on the community
103+
developed [KServe
104+
protocol](https://github.com/kserve/kserve/tree/master/docs/predict-api/v2).
105+
106+
* A [C API](docs/inference_protocols.md#c-api) allows Triton to be
107+
linked directly into your application for edge and other in-process
108+
use cases.
109+
110+
* [Metrics](docs/metrics.md) indicating GPU utilization, server
111+
throughput, and server latency. The metrics are provided in
112+
Prometheus data format.
113+
114+
## Documentation
115+
116+
[Triton Architecture](docs/architecture.md) gives a high-level
117+
overview of the structure and capabilities of the inference
118+
server. There is also an [FAQ](docs/faq.md). Additional documentation
119+
is divided into [*user*](#user-documentation) and
120+
[*developer*](#developer-documentation) sections. The *user*
121+
documentation describes how to use Triton as an inference solution,
122+
including information on how to configure Triton, how to organize and
123+
configure your models, how to use the C++ and Python clients, etc. The
124+
*developer* documentation describes how to build and test Triton and
125+
also how Triton can be extended with new functionality.
126+
127+
The Triton [Release
128+
Notes](https://docs.nvidia.com/deeplearning/triton-inference-server/release-notes/index.html)
129+
and [Support
130+
Matrix](https://docs.nvidia.com/deeplearning/dgx/support-matrix/index.html)
131+
indicate the required versions of the NVIDIA Driver and CUDA, and also
132+
describe supported GPUs.
133+
134+
### User Documentation
135+
136+
* [QuickStart](docs/quickstart.md)
137+
* [Install Triton](docs/quickstart.md#install-triton-docker-image)
138+
* [Create Model Repository](docs/quickstart.md#create-a-model-repository)
139+
* [Run Triton](docs/quickstart.md#run-triton)
140+
* [Model Repository](docs/model_repository.md)
141+
* [Cloud Storage](docs/model_repository.md#model-repository-locations)
142+
* [File Organization](docs/model_repository.md#model-files)
143+
* [Model Versioning](docs/model_repository.md#model-versions)
144+
* [Model Configuration](docs/model_configuration.md)
145+
* [Required Model Configuration](docs/model_configuration.md#minimal-model-configuration)
146+
* [Maximum Batch Size - Batching and Non-Batching Models](docs/model_configuration.md#maximum-batch-size)
147+
* [Input and Output Tensors](docs/model_configuration.md#inputs-and-outputs)
148+
* [Tensor Datatypes](docs/model_configuration.md#datatypes)
149+
* [Tensor Reshape](docs/model_configuration.md#reshape)
150+
* [Shape Tensor](docs/model_configuration.md#shape-tensors)
151+
* [Auto-Generate Required Model Configuration](docs/model_configuration.md#auto-generated-model-configuration)
152+
* [Version Policy](docs/model_configuration.md#version-policy)
153+
* [Instance Groups](docs/model_configuration.md#instance-groups)
154+
* [Specifying Multiple Model Instances](docs/model_configuration.md#multiple-model-instances)
155+
* [CPU and GPU Instances](docs/model_configuration.md#cpu-model-instance)
156+
* [Configuring Rate Limiter](docs/model_configuration.md#rate-limiter-configuration)
157+
* [Optimization Settings](docs/model_configuration.md#optimization_policy)
158+
* [Framework-Specific Optimization](docs/optimization.md#framework-specific-optimization)
159+
* [ONNX-TensorRT](docs/optimization.md#onnx-with-tensorrt-optimization)
160+
* [ONNX-OpenVINO](docs/optimization.md#onnx-with-openvino-optimization)
161+
* [TensorFlow-TensorRT](docs/optimization.md#tensorflow-with-tensorrt-optimization)
162+
* [TensorFlow-Mixed-Precision](docs/optimization.md#tensorflow-automatic-fp16-optimization)
163+
* [NUMA Optimization](docs/optimization.md#numa-optimization)
164+
* [Scheduling and Batching](docs/model_configuration.md#scheduling-and-batching)
165+
* [Default Scheduler - Non-Batching](docs/model_configuration.md#default-scheduler)
166+
* [Dynamic Batcher](docs/model_configuration.md#dynamic-batcher)
167+
* [How to Configure Dynamic Batcher](docs/model_configuration.md#recommended-configuration-process)
168+
* [Delayed Batching](docs/model_configuration.md#delayed-batching)
169+
* [Preferred Batch Size](docs/model_configuration.md#preferred-batch-sizes)
170+
* [Preserving Request Ordering](docs/model_configuration.md#preserve-ordering)
171+
* [Priority Levels](docs/model_configuration.md#priority-levels)
172+
* [Queuing Policies](docs/model_configuration.md#queue-policy)
173+
* [Ragged Batching](docs/ragged_batching.md)
174+
* [Sequence Batcher](docs/model_configuration.md#sequence-batcher)
175+
* [Stateful Models](docs/architecture.md#stateful-models)
176+
* [Control Inputs](docs/architecture.md#control-inputs)
177+
* [Implicit State - Stateful Inference Using a Stateless Model](docs/architecture.md#implicit-state-management)
178+
* [Sequence Scheduling Strategies](docs/architecture.md#scheduling-strateties)
179+
* [Direct](docs/architecture.md#direct)
180+
* [Oldest](docs/architecture.md#oldest)
181+
* [Rate Limiter](docs/rate_limiter.md)
182+
* [Model Warmup](docs/model_configuration.md#model-warmup)
183+
* [Inference Request/Response Cache](docs/model_configuration.md#response-cache)
184+
* Model Pipeline
185+
* [Model Ensemble](docs/architecture.md#ensemble-models)
186+
* [Business Logic Scripting (BLS)](https://github.com/triton-inference-server/python_backend#business-logic-scripting)
187+
* [Model Management](docs/model_management.md)
188+
* [Explicit Model Loading and Unloading](docs/model_management.md#model-control-mode-explicit)
189+
* [Modifying the Model Repository](docs/model_management.md#modifying-the-model-repository)
190+
* [Metrics](docs/metrics.md)
191+
* [Framework Custom Operations](docs/custom_operations.md)
192+
* [TensorRT](docs/custom_operations.md#tensorrt)
193+
* [TensorFlow](docs/custom_operations.md#tensorflow)
194+
* [PyTorch](docs/custom_operations.md#pytorch)
195+
* [ONNX](docs/custom_operations.md#onnx)
196+
* [Client Libraries and Examples](https://github.com/triton-inference-server/client)
197+
* [C++ HTTP/GRPC Libraries](https://github.com/triton-inference-server/client#client-library-apis)
198+
* [Python HTTP/GRPC Libraries](https://github.com/triton-inference-server/client#client-library-apis)
199+
* [Java HTTP Library](https://github.com/triton-inference-server/client/src/java)
200+
* GRPC Generated Libraries
201+
* [go](https://github.com/triton-inference-server/client/tree/main/src/grpc_generated/go)
202+
* [Java/Scala](https://github.com/triton-inference-server/client/tree/main/src/grpc_generated/java)
203+
* [Performance Analysis](docs/optimization.md)
204+
* [Model Analyzer](docs/model_analyzer.md)
205+
* [Performance Analyzer](docs/perf_analyzer.md)
206+
* [Inference Request Tracing](docs/trace.md)
207+
* [Jetson and JetPack](docs/jetson.md)
208+
209+
The [quickstart](docs/quickstart.md) walks you through all the steps
210+
required to install and run Triton with an example image
211+
classification model and then use an example client application to
212+
perform inferencing using that model. The quickstart also demonstrates
213+
how [Triton supports both GPU systems and CPU-only
214+
systems](docs/quickstart.md#run-triton).
215+
216+
The first step in using Triton to serve your models is to place one or
217+
more models into a [model
218+
repository](docs/model_repository.md). Optionally, depending on the type
219+
of the model and on what Triton capabilities you want to enable for
220+
the model, you may need to create a [model
221+
configuration](docs/model_configuration.md) for the model. If your
222+
model has [custom operations](docs/custom_operations.md) you will need
223+
to make sure they are loaded correctly by Triton.
224+
225+
After you have your model(s) available in Triton, you will want to
226+
send inference and other requests to Triton from your *client*
227+
application. The [Python and C++ client
228+
libraries](https://github.com/triton-inference-server/client) provide
229+
APIs to simplify this communication. There are also a large number of
230+
[client examples](https://github.com/triton-inference-server/client)
231+
that demonstrate how to use the libraries. You can also send
232+
HTTP/REST requests directly to Triton using the [HTTP/REST JSON-based
233+
protocol](docs/inference_protocols.md#httprest-and-grpc-protocols) or
234+
[generate a GRPC client for many other
235+
languages](https://github.com/triton-inference-server/client).
236+
237+
Understanding and [optimizing performance](docs/optimization.md) is an
238+
important part of deploying your models. The Triton project provides
239+
the [Performance Analyzer](docs/perf_analyzer.md) and the [Model
240+
Analyzer](docs/model_analyzer.md) to help your optimization
241+
efforts. Specifically, you will want to optimize [scheduling and
242+
batching](docs/architecture.md#models-and-schedulers) and [model
243+
instances](docs/model_configuration.md#instance-groups) appropriately
244+
for each model. You can also enable cross-model prioritization using
245+
the [rate limiter](docs/rate_limiter.md) which manages the rate at
246+
which requests are scheduled on model instances. You may also want to
247+
consider combining multiple models and pre/post-processing into a
248+
pipeline using [ensembling](docs/architecture.md#ensemble-models) or
249+
[Business Logic Scripting
250+
(BLS)](https://github.com/triton-inference-server/python_backend#business-logic-scripting). A
251+
[Prometheus metrics endpoint](docs/metrics.md) allows you to visualize
252+
and monitor aggregate inference metrics.
253+
254+
NVIDIA publishes a number of [deep learning
255+
examples](https://github.com/NVIDIA/DeepLearningExamples) that use
256+
Triton.
257+
258+
As part of your deployment strategy you may want to [explicitly manage
259+
what models are available by loading and unloading
260+
models](docs/model_management.md) from a running Triton server. If you
261+
are using Kubernetes for deployment there are simple examples of how
262+
to deploy Triton using Kubernetes and Helm:
263+
[GCP](deploy/gcp/README.md), [AWS](deploy/aws/README.md), and [NVIDIA
264+
FleetCommand](deploy/fleetcommand/README.md)
265+
266+
The [version 1 to version 2 migration
267+
information](docs/v1_to_v2.md) is helpful if you are moving to
268+
version 2 of Triton from previously using version 1.
269+
270+
### Developer Documentation
271+
272+
* [Build](docs/build.md)
273+
* [Protocols and APIs](docs/inference_protocols.md).
274+
* [Backends](https://github.com/triton-inference-server/backend)
275+
* [Repository Agents](docs/repository_agents.md)
276+
* [Test](docs/test.md)
277+
278+
Triton can be [built using
279+
Docker](docs/build.md#building-triton-with-docker) or [built without
280+
Docker](docs/build.md#building-triton-without-docker). After building
281+
you should [test Triton](docs/test.md).
282+
283+
It is also possible to [create a Docker image containing a customized
284+
Triton](docs/compose.md) that contains only a subset of the backends.
285+
286+
The Triton project also provides [client libraries for Python and
287+
C++](https://github.com/triton-inference-server/client) that make it
288+
easy to communicate with the server. There are also a large number of
289+
[example clients](https://github.com/triton-inference-server/client)
290+
that demonstrate how to use the libraries. You can also develop your
291+
own clients that directly communicate with Triton using [HTTP/REST or
292+
GRPC protocols](docs/inference_protocols.md). There is also a [C
293+
API](docs/inference_protocols.md) that allows Triton to be linked
294+
directly into your application.
295+
296+
A [Triton backend](https://github.com/triton-inference-server/backend)
297+
is the implementation that executes a model. A backend can interface
298+
with a deep learning framework, like PyTorch, TensorFlow, TensorRT or
299+
ONNX Runtime; or it can interface with a data processing framework
300+
like [DALI](https://github.com/triton-inference-server/dali_backend);
301+
or you can extend Triton by [writing your own
302+
backend](https://github.com/triton-inference-server/backend) in either
303+
[C/C++](https://github.com/triton-inference-server/backend/blob/main/README.md#triton-backend-api)
304+
or
305+
[Python](https://github.com/triton-inference-server/python_backend).
306+
307+
A [Triton repository agent](docs/repository_agents.md) extends Triton
308+
with new functionality that operates when a model is loaded or
309+
unloaded. You can introduce your own code to perform authentication,
310+
decryption, conversion, or similar operations when a model is loaded.
311+
312+
## Papers and Presentation
313+
314+
* [Maximizing Deep Learning Inference Performance with NVIDIA Model
315+
Analyzer](https://developer.nvidia.com/blog/maximizing-deep-learning-inference-performance-with-nvidia-model-analyzer/).
316+
317+
* [High-Performance Inferencing at Scale Using the TensorRT Inference
318+
Server](https://developer.nvidia.com/gtc/2020/video/s22418).
319+
320+
* [Accelerate and Autoscale Deep Learning Inference on GPUs with
321+
KFServing](https://developer.nvidia.com/gtc/2020/video/s22459).
322+
323+
* [Deep into Triton Inference Server: BERT Practical Deployment on
324+
NVIDIA GPU](https://developer.nvidia.com/gtc/2020/video/s21736).
325+
326+
* [Maximizing Utilization for Data Center Inference with TensorRT
327+
Inference Server](https://on-demand-gtc.gputechconf.com/gtcnew/sessionview.php?sessionName=s9438-maximizing+utilization+for+data+center+inference+with+tensorrt+inference+server).
328+
329+
* [NVIDIA TensorRT Inference Server Boosts Deep Learning
330+
Inference](https://devblogs.nvidia.com/nvidia-serves-deep-learning-inference/).
331+
332+
* [GPU-Accelerated Inference for Kubernetes with the NVIDIA TensorRT
333+
Inference Server and
334+
Kubeflow](https://www.kubeflow.org/blog/nvidia_tensorrt/).
335+
336+
* [Deploying NVIDIA Triton at Scale with MIG and Kubernetes](https://developer.nvidia.com/blog/deploying-nvidia-triton-at-scale-with-mig-and-kubernetes/).
337+
338+
## Contributing
339+
340+
Contributions to Triton Inference Server are more than welcome. To
341+
contribute make a pull request and follow the guidelines outlined in
342+
[CONTRIBUTING.md](CONTRIBUTING.md). If you have a backend, client,
343+
example or similar contribution that is not modifying the core of
344+
Triton, then you should file a PR in the [contrib
345+
repo](https://github.com/triton-inference-server/contrib).
346+
347+
## Reporting problems, asking questions
348+
349+
We appreciate any feedback, questions or bug reporting regarding this
350+
project. When help with code is needed, follow the process outlined in
351+
the Stack Overflow (<https://stackoverflow.com/help/mcve>)
352+
document. Ensure posted examples are:
353+
354+
* minimal – use as little code as possible that still produces the
355+
same problem
356+
357+
* complete – provide all parts needed to reproduce the problem. Check
358+
if you can strip external dependency and still show the problem. The
359+
less time we spend on reproducing problems the more time we have to
360+
fix it
361+
362+
* verifiable – test the code you're about to provide to make sure it
363+
reproduces the problem. Remove all other problems that are not
364+
related to your request/question.

0 commit comments

Comments
 (0)