saymrwulf/onnxruntime
1
0
Fork 0
mirror of https://github.com/saymrwulf/onnxruntime.git synced 2026-06-03 23:49:44 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions
ONNX Runtime: cross-platform, high performance ML inferencing and training accelerator
10038 commits 2856 branches 57 tags 1.3 GiB
C++ 87.3%
Python 3.1%
C 2.4%
PureBasic 2%
C# 1%
Other 3.7%
Find a file
pengwa 43a5147e01
Memory optimization refactor and refinement (#17481) 
### Memory optimization refactor and refinement

Currently memory optimizer runs graph transformations and print
recompute opportunities in INFO level, while ORT backend has many many
INFO level logs making users hard to find those information. So we are
looking for a Python binding API to retrieve the memory optimization
opportunities instead of depending on the MemoryOptimizer's default
logging.
Then we can print ORTModule feature statistics using this information. 
Also, with such an API, we can create an ORT session created, where
allocation plan is done, the analysis will consider buffer reuse as
well. This can void giving some recomputation subgraphs that are reusing
other subgraphs' output buffers.

Check
https://github.com/microsoft/onnxruntime/blob/pengwa/add_devinfo_level/docs/Memory_Optimizer.md
for the new flow using `MemoryOptimizer`.

This pull requests made following refactoring:
1. Print the log in ORTModule Python script, along with ORTModule
feature enabling stats. This is implemented by exposing an API
`get_serialized_ortmodule_memory_stat` to retrieve the memory
optimization opportunities.
2. We are analyzing memory optimization opportunities considering ORT
memory planning. This is done by firstly creating the execution graph
without enabling MemoryOptimizer, then we call
`execution_agent.get_serialized_ortmodule_memory_stat` which internally
will consider the session memory allocation planner when analyzing
memory optimization opportunity. As a direct result, the memory
optimization opportunities can show those stashed activations that are
reusing other buffers.
3. Move recompute analysis logic from memory_optimizer.h/cc to
recompute_analysis.h/cc.
4. Abstract optimization strategies for their own implementation. This
will make introducing new strategies (for example compression and
decompression ) easier.

New logging matrix (INFO Level), in WARNING level, the details will NOT
show.
```
2023-09-13 13:25:09,249 orttraining.rank-0 [WARNING] -
***** ONNX Runtime Training (ORTModule) is accelerating your model *****

ORTModule is enabled with following features ON/OFF for [training] mode:

  ATen Executor         :   ON    :   Dispatch ATen operators to ORT's ATen executor
  Cast Propagation      :   ON    :   Level 1 enabled
  Custom Function       :   ON    :   Support custom torch.autograd.Function export and execution
  Memory Optimizer      :   ON    :   RecomputeConfig: Reshape+Where+BiasSoftmax+:1:-1,Cast+:1:-1, ProbeLevel: 1, available configs:
                                      Config                                                      Freq    Saving(B)       Saving Symbolic(Bytes)
   - Plan 1             :   ON    :   Reshape+Where+BiasSoftmax+:1:-1                             5       671,088,640     640.0*inputs_input_ids_dim0*inputs_input_ids_dim1**2
   - Plan 2             :   ON    :   Cast+:1:-1                                                  6       402,587,648     inputs_input_ids_dim0*inputs_input_ids_dim1*(384.0*inputs_input_ids_dim1 - 64.0)
   - Plan 3             :   OFF   :   Reshape+Where+:1:-1                                         1       134,217,728     128.0*inputs_input_ids_dim0*inputs_input_ids_dim1**2
   - Plan 4             :   OFF   :   BiasSoftmax+:1:-1                                           1       134,086,656     128.0*inputs_input_ids_dim0*inputs_input_ids_dim1*(inputs_input_ids_dim1 - 1)
   - Plan 5             :   OFF   :   BiasGelu+:1:-1                                              6       125,808,640     inputs_input_ids_dim0*(122880.0*inputs_input_ids_dim1 - 20480.0)
   - Plan 6             :   OFF   :   FusedMatMul+:1:-1                                           6       125,808,640     inputs_input_ids_dim0*(122880.0*inputs_input_ids_dim1 - 20480.0)
   - Plan 7             :   OFF   :   FusedMatMul+Add+FusedMatMul+Add+Add+Add+:1:-1               5       26,214,400      25600.0*inputs_input_ids_dim0*inputs_input_ids_dim1
   - Plan 8             :   OFF   :   Add+:1:-1                                                   1       5,237,760       5120.0*inputs_input_ids_dim0*(inputs_input_ids_dim1 - 1)
   - Plan 9             :   OFF   :   Reshape+Unsqueeze+Unsqueeze+Cast+Sub+Mul+Cast+:1:-1         1       4,096           4.0*inputs_input_ids_dim0*inputs_input_ids_dim1
   - Plan 10            :   OFF   :   Cast+:2:-1                                                  1       2,048           2.0*inputs_input_ids_dim0*inputs_input_ids_dim1
  Compute Optimizer     :   ON    :   Enable/Disable with env ORTMODULE_ENABLE_COMPUTE_OPTIMIZER=1/0
   - FLOPReduction      :   ON    :   Reduce FLOPs by upstreaming shrinking-sized ops
  Auto Fallback         :   ON    :   Fallback to PyTorch when encountering unsupported ops
  TritonOp Enabled      :   OFF   :   ORT will switch to Triton for executing some ops to further accelerate training.
  ZeRO Stage3 Support   :   OFF   :   Enable/Disable with env ORTMODULE_ENABLE_ZERO_STAGE3=1/0

Total ORT initialization overhead is 10.73s where export takes 8.39s.
Other overhead details:  graph builder init takes 0.06s, runtime detection takes 0.01s, graph building takes 0.31s, session creation takes 1.96s

Versions: ONNX Runtime - 1.16.0+cu118, ONNX - 1.11.0

Note 1: use comma to enable multiple plans at the same time.
  export ORTMODULE_MEMORY_OPT_CONFIG=<plan1 config>,<plan2 config>,...
Note 2: saving is calculated based on the 1st batch symbolic dim values:
  inputs_input_ids_dim0=1,
  inputs_input_ids_dim1=1024,
  inputs_attention_mask_dim0=1,
  inputs_attention_mask_dim1=1024,
  inputs_labels_dim0=1,
  inputs_labels_dim1=1024,

************************************************************************
```

If DEVINFO level is enabled, then more details about the memory
optimizations are printed.
```

MemoryInsight Summary - User config: BiasGelu+:1:-1,Cast+:2:-1
==========================================================================================================================================
|Freq   | Memory Optimization Opportunities (Clustered by node-level activation patterns)                                                |
|_ _ _ _|_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ |
|3      |For each row options are mutually exclusive, only one of them can be enabled.                                                   |
|       |                                                                                                                                |
|       |>>Option 1     : Recompute subgraph FusedMatMul+Add+Reshape+                                                                    |
|       |  Status       : Disabled. Enable with export ORTMODULE_MEMORY_OPT_CONFIG=FusedMatMul+Add+Reshape+:1:-1                         |
|       |  Stashed Activations:                                                                                                          |
|       |   - ReuseFreq :  Output 0(3),                                                                                                  |
|       |   - Output 0  : [inputs_input_ids_dim0 x inputs_input_ids_dim1 x 32 x 240 x ], byte/elem: 2, 100% saved                        |
|_ _ _ _|_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ |
|2      |For each row options are mutually exclusive, only one of them can be enabled.                                                   |
|       |                                                                                                                                |
|       |>>Option 1     : Recompute subgraph Reshape+                                                                                    |
|       |  Status       : Disabled. Enable with export ORTMODULE_MEMORY_OPT_CONFIG=Reshape+:1:-1                                         |
|       |  Stashed Activations:                                                                                                          |
|       |   - ReuseFreq :  Output 0(2),                                                                                                  |
|       |   - Output 0  : [ x 2560 x ], byte/elem: 2, 100% saved                                                                         |
|_ _ _ _|_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ |
|2      |For each row options are mutually exclusive, only one of them can be enabled.                                                   |
|       |                                                                                                                                |
|       |>>Option 1     : Recompute subgraph FusedMatMul+                                                                                |
|       |  Status       : Disabled. Enable with export ORTMODULE_MEMORY_OPT_CONFIG=FusedMatMul+:1:-1                                     |
|       |  Stashed Activations:                                                                                                          |
|       |   - Output 0  : [inputs_input_ids_dim0 x inputs_input_ids_dim1 x 10240 x ], byte/elem: 2, 100% saved                           |
|_ _ _ _|_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ |
|2      |For each row options are mutually exclusive, only one of them can be enabled.                                                   |
|       |                                                                                                                                |
|       |>>Option 1     : Recompute subgraph Cast+                                                                                       |
|       |  Status       : Disabled. Enable with export ORTMODULE_MEMORY_OPT_CONFIG=Cast+:1:-1                                            |
|       |  Stashed Activations:                                                                                                          |
|       |   - Output 0  : [inputs_input_ids_dim0 x 32 x inputs_input_ids_dim1 x inputs_input_ids_dim1 x ], byte/elem: 2, 100% saved      |
|_ _ _ _|_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ |
|2      |For each row options are mutually exclusive, only one of them can be enabled.                                                   |
|       |                                                                                                                                |
|       |>>Option 1     : Recompute subgraph Reshape+Where+BiasSoftmax+                                                                  |
|       |  Status       : Disabled. Enable with export ORTMODULE_MEMORY_OPT_CONFIG=Reshape+Where+BiasSoftmax+:1:-1                       |
|       |  Stashed Activations:                                                                                                          |
|       |   - Output 0  : [inputs_input_ids_dim0 x 32 x inputs_input_ids_dim1 x inputs_input_ids_dim1 x ], byte/elem: 4, 100% saved      |
|_ _ _ _|_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ |
|2      |For each row options are mutually exclusive, only one of them can be enabled.                                                   |
|       |                                                                                                                                |
|       |>>Option 1     : Recompute subgraph BiasGelu+                                                                                   |
|       |  Status       : Enabled, requested count=-1, actual applied count=2                                                            |
|       |  Stashed Activations:                                                                                                          |
|       |   - Output 0  : [inputs_input_ids_dim0 x inputs_input_ids_dim1 x 10240 x ], byte/elem: 2, 100% saved                           |
|_ _ _ _|_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ |
|2      |For each row options are mutually exclusive, only one of them can be enabled.                                                   |
|       |                                                                                                                                |
|       |>>Option 1     : Recompute subgraph FusedMatMul+Add+FusedMatMul+Add+Add+Add+                                                    |
|       |  Status       : Disabled. Enable with export ORTMODULE_MEMORY_OPT_CONFIG=FusedMatMul+Add+FusedMatMul+Add+Add+Add+:1:-1         |
|       |  Stashed Activations:                                                                                                          |
|       |   - Output 0  : [inputs_input_ids_dim0 x inputs_input_ids_dim1 x 2560 x ], byte/elem: 2, 100% saved                            |
|_ _ _ _|_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ |
|1      |For each row options are mutually exclusive, only one of them can be enabled.                                                   |
|       |                                                                                                                                |
|       |>>Option 1     : Recompute subgraph Reshape+Where+                                                                              |
|       |  Status       : Disabled. Enable with export ORTMODULE_MEMORY_OPT_CONFIG=Reshape+Where+:1:-1                                   |
|       |  Stashed Activations:                                                                                                          |
|       |   - Output 0  : [inputs_input_ids_dim0 x 32 x inputs_input_ids_dim1 x inputs_input_ids_dim1 x ], byte/elem: 4, 100% saved      |
|_ _ _ _|_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ |
|1      |For each row options are mutually exclusive, only one of them can be enabled.                                                   |
|       |                                                                                                                                |
|       |>>Option 1     : Recompute subgraph FusedMatMul+                                                                                |
|       |  Status       : Disabled. Enable with export ORTMODULE_MEMORY_OPT_CONFIG=FusedMatMul+:1:-1                                     |
|       |  Stashed Activations:                                                                                                          |
|       |   - Output 0  : [inputs_input_ids_dim0*(inputs_input_ids_dim1 - 1) x 10240 x ], byte/elem: 2, 100% saved                       |
|_ _ _ _|_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ |
|1      |For each row options are mutually exclusive, only one of them can be enabled.                                                   |
|       |                                                                                                                                |
|       |>>Option 1     : Recompute subgraph Cast+                                                                                       |
|       |  Status       : Disabled. Enable with export ORTMODULE_MEMORY_OPT_CONFIG=Cast+:1:-1                                            |
|       |  Stashed Activations:                                                                                                          |
|       |   - Output 0  : [inputs_input_ids_dim0 x 32 x inputs_input_ids_dim1 - 1 x inputs_input_ids_dim1 x ], byte/elem: 2, 100% saved  |
|_ _ _ _|_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ |
|1      |For each row options are mutually exclusive, only one of them can be enabled.                                                   |
|       |                                                                                                                                |
|       |>>Option 1     : Recompute subgraph Reshape+Unsqueeze+Unsqueeze+Cast+Sub+Mul+Cast+                                              |
|       |  Status       : Disabled. Enable with export ORTMODULE_MEMORY_OPT_CONFIG=Reshape+Unsqueeze+Unsqueeze+Cast+Sub+Mul+Cast+:1:-1   |
|       |  Stashed Activations:                                                                                                          |
|       |   - Output 0  : [inputs_input_ids_dim0 x 1 x 1 x inputs_input_ids_dim1 x ], byte/elem: 4, 100% saved                           |
|       |                                                                                                                                |
|       |>>Option 2     : RecomputeWithCompromise subgraph Cast+                                                                         |
|       |  Status       : Enabled, requested count=-1, actual applied count=1                                                            |
|       |  Stashed Activations:                                                                                                          |
|       |   - Output 0  : [inputs_input_ids_dim0 x 1 x 1 x inputs_input_ids_dim1 x ], byte/elem: 4, 50% saved                            |
|_ _ _ _|_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ |
|1      |For each row options are mutually exclusive, only one of them can be enabled.                                                   |
|       |                                                                                                                                |
|       |>>Option 1     : Recompute subgraph BiasSoftmax+                                                                                |
|       |  Status       : Disabled. Enable with export ORTMODULE_MEMORY_OPT_CONFIG=BiasSoftmax+:1:-1                                     |
|       |  Stashed Activations:                                                                                                          |
|       |   - Output 0  : [inputs_input_ids_dim0 x 32 x inputs_input_ids_dim1 - 1 x inputs_input_ids_dim1 x ], byte/elem: 4, 100% saved  |
|_ _ _ _|_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ |
|1      |For each row options are mutually exclusive, only one of them can be enabled.                                                   |
|       |                                                                                                                                |
|       |>>Option 1     : Recompute subgraph BiasGelu+                                                                                   |
|       |  Status       : Enabled, requested count=-1, actual applied count=1                                                            |
|       |  Stashed Activations:                                                                                                          |
|       |   - Output 0  : [inputs_input_ids_dim0*(inputs_input_ids_dim1 - 1) x 10240 x ], byte/elem: 2, 100% saved                       |
|_ _ _ _|_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ |
|1      |For each row options are mutually exclusive, only one of them can be enabled.                                                   |
|       |                                                                                                                                |
|       |>>Option 1     : Recompute subgraph Add+                                                                                        |
|       |  Status       : Disabled. Enable with export ORTMODULE_MEMORY_OPT_CONFIG=Add+:1:-1                                             |
|       |  Stashed Activations:                                                                                                          |
|       |   - Output 0  : [inputs_input_ids_dim0*(inputs_input_ids_dim1 - 1) x 2560 x ], byte/elem: 2, 100% saved                        |
|_ _ _ _|_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ |
==========================================================================================================================================
Note: use comma as a separator for enabling more than one subgraphs.

************************************************************************

```


### Motivation and Context
<!-- - Why is this change required? What problem does it solve?
- If it fixes an open issue, please link to the issue here. -->
2023-11-23 11:39:00 +08:00
.config Update tsaoptions.json: update the email alias (#13448) 2022-10-26 15:56:16 -07:00
.devcontainer Remove two lines in the Dockerfile for Github Codespace (#12278) 2022-07-21 20:52:17 -07:00
.gdn Update win-ci-pipeline.yml: enable xnnpack tests (#16244) 2023-06-14 19:12:42 -07:00
.github Update stale.yml to fix start-date bug (#18376) 2023-11-09 16:04:31 -08:00
.pipelines Bump DirectML version from 1.12.0 to 1.12.1 (#17225) 2023-08-20 09:55:38 -07:00
.vscode Remove deprecated vscode settings (#18349) 2023-11-10 18:00:35 -08:00
cgmanifests onboard MoE (#18279) 2023-11-14 16:48:51 -08:00
cmake Memory optimization refactor and refinement (#17481) 2023-11-23 11:39:00 +08:00
csharp Fix 4 more bad delegates missing the attribute that cause iOS AOT errors at runtime (#18390) 2023-11-14 14:00:21 +10:00
dockerfiles Update dockerfiles/Dockerfile.source to avoid installing onnx (#17975) 2023-10-20 09:24:21 -07:00
docs Memory optimization refactor and refinement (#17481) 2023-11-23 11:39:00 +08:00
include/onnxruntime/core Memory optimization refactor and refinement (#17481) 2023-11-23 11:39:00 +08:00
java [java] Make the backing byte buffer in an OrtValue accessible (#16578) 2023-10-17 10:03:49 -07:00
js [js/webgpu] Add BatchNormalization Op (#18468) 2023-11-22 15:58:06 -08:00
objectivec Objective-C Add Support to Create and Query String ORTValues (#16764) 2023-07-20 17:39:29 -07:00
onnxruntime Memory optimization refactor and refinement (#17481) 2023-11-23 11:39:00 +08:00
orttraining Memory optimization refactor and refinement (#17481) 2023-11-23 11:39:00 +08:00
rust Fix rust compile issues and add GH action to run build validations and tests (#18346) 2023-11-09 04:26:02 -08:00
samples Removed all the deprecated python training code and related tests and utils (#18333) 2023-11-17 18:19:21 -08:00
tools [js/web] use Chrome in CI for npm tests (#18522) 2023-11-21 18:03:57 -08:00
winml Bump linter versions (#18341) 2023-11-08 13:04:40 -08:00
.clang-format Prevent GSL_SUPPRESS arguments from being modified by clang-format (#17242) 2023-08-22 18:26:53 -07:00
.clang-tidy Create clang-tidy CI (#12653) 2022-09-30 08:05:38 -07:00
.dockerignore
.gitattributes
.gitignore remove 'lib/' from .gitignore (#15613) 2023-04-24 18:43:32 -07:00
.gitmodules Remove onnxruntime extensions from list of gitmodules (#17615) 2023-09-19 17:12:14 -07:00
.lintrunner.toml FP16 optimizer automatically detect DeepSpeed compatibility (#18084) 2023-10-25 15:11:02 +08:00
build.bat try to find patch.exe in git default installation folder (#17106) 2023-08-10 21:48:13 -07:00
build.sh Upgrade old Python version in packaging pipeline (#16667) 2023-07-17 08:24:47 -07:00
CITATION.cff Fix CITATION.cff and add automatic validation of your citation metadata (#10478) 2022-04-13 10:03:52 -07:00
CODEOWNERS Add owners for public facing API files (#15288) 2023-03-30 17:16:15 -07:00
CONTRIBUTING.md Fix link to High Level Design (#11786) 2023-02-28 11:05:54 -08:00
lgtm.yml Fix lgtm C++ error (#13613) 2022-11-10 10:06:22 -08:00
LICENSE
NuGet.config
ort.wprp
ORT_icon_for_light_bg.png
packages.config Bump DirectML version from 1.12.0 to 1.12.1 (#17225) 2023-08-20 09:55:38 -07:00
pyproject.toml [ORTModule] ATen Efficient Attention and Triton Flash Attention (#17959) 2023-10-27 10:29:27 +08:00
README.md add third-party pipeline status to README.md (#16155) 2023-05-31 22:14:39 -07:00
requirements-dev.txt ONNX 1.15 integration (#17125) 2023-09-26 14:44:48 -07:00
requirements-doc.txt
requirements-lintrunner.txt Bump linter versions (#18341) 2023-11-08 13:04:40 -08:00
requirements-training.txt ONNX 1.15 integration (#17125) 2023-09-26 14:44:48 -07:00
requirements.txt.in Add additional python requirements (#11522) 2022-05-20 16:16:18 -07:00
SECURITY.md Microsoft mandatory file (#11619) 2022-05-25 13:56:10 -07:00
setup.py Update setup.py: replace libcudart.so.12.0 with libcudart.so.12 (#18501) 2023-11-19 22:06:32 -08:00
ThirdPartyNotices.txt Flash Attention v2 MHA (#17227) 2023-08-31 13:52:21 -07:00
VERSION_NUMBER Bump Up Version to 1.17.0 (#17587) 2023-09-20 11:02:58 +08:00

README.md

ONNX Runtime is a cross-platform inference and training machine-learning accelerator.

ONNX Runtime inference can enable faster customer experiences and lower costs, supporting models from deep learning frameworks such as PyTorch and TensorFlow/Keras as well as classical machine learning libraries such as scikit-learn, LightGBM, XGBoost, etc. ONNX Runtime is compatible with different hardware, drivers, and operating systems, and provides optimal performance by leveraging hardware accelerators where applicable alongside graph optimizations and transforms. Learn more →

ONNX Runtime training can accelerate the model training time on multi-node NVIDIA GPUs for transformer models with a one-line addition for existing PyTorch training scripts. Learn more →

Get Started & Resources

Builtin Pipeline Status

System Inference Training
Windows Build Status
Build Status
Build Status
Linux Build Status
Build Status
Build Status
Build Status
Build Status		 Build Status
Build Status
Build Status
Mac Build Status
Android Build Status
iOS Build Status
Web Build Status
Other Build Status
Build Status

Third-party Pipeline Status

System Inference Training
Linux Build Status

Data/Telemetry

Windows distributions of this project may collect usage data and send it to Microsoft to help improve our products and services. See the privacy statement for more details.

Contributions and Feedback

We welcome contributions! Please see the contribution guidelines.

For feature requests or bug reports, please file a GitHub Issue.

For general discussion or questions, please use GitHub Discussions.

Code of Conduct

This project has adopted the Microsoft Open Source Code of Conduct. For more information see the Code of Conduct FAQ or contact opencode@microsoft.com with any additional questions or comments.

License

This project is licensed under the MIT License.