58af36b49a
### Fuse ScaledSum and its backward BatchScale For deberta models, there is a pattern a / scalar_0 + b / scalar_1 + c / scalar_2 We can fuse this into ScaledSum operator, taking 2(or 3) inputs, and 2(or 3) attributes scalar, generating one output. For the backward, the gradient of a, b and c will be computed with BatchScale. ### Benchmark on 8x32GV100 ```bash torchrun --nproc_per_node=8 examples/onnxruntime/training/language-modeling/run_mlm.py --model_name_or_path microsoft/deberta-v3-large --dataset_name wikitext --dataset_config_name wikitext-2-raw-v1 --num_train_epochs 10 --do_train --overwrite_output_dir --output_dir ./outputs/ --seed 1137 --fp16 --report_to none --optim adamw_ort_fused --max_steps 400 --logging_steps 1 --use_module_with_loss --deepspeed aml_ds_config_zero_1.json --per_device_train_batch_size 10 ``` #### Main Branch ``` Total overhead: 127954ms where export takes 116489ms. epoch = 14.29 train_loss = 4.9803 train_runtime = 0:10:27.29 train_samples = 2223 train_samples_per_second = 51.013 train_steps_per_second = 0.638 throughput per GPU = 14.29* 2223/ (627.29 - 127.954) / 8 (gpu) = 7.952 samples/second ``` #### This PR ``` Total overhead: 128761ms where export takes 118510ms. ***** train metrics ***** epoch = 14.29 train_loss = 4.6144 train_runtime = 0:10:04.31 train_samples = 2223 train_samples_per_second = 52.953 train_steps_per_second = 0.662 throughput per GPU = 14.29*2223 / (604.31 - 128.761) / 8 = 8.350 samples/second ``` 5.x% performance gains. |
||
|---|---|---|
| .config | ||
| .devcontainer | ||
| .gdn | ||
| .github | ||
| .pipelines | ||
| .vscode | ||
| cgmanifests | ||
| cmake | ||
| csharp | ||
| dockerfiles | ||
| docs | ||
| include/onnxruntime/core | ||
| java | ||
| js | ||
| objectivec | ||
| onnxruntime | ||
| orttraining | ||
| rust | ||
| samples | ||
| swift/OnnxRuntimeBindingsTests | ||
| tools | ||
| winml | ||
| .clang-format | ||
| .clang-tidy | ||
| .dockerignore | ||
| .gitattributes | ||
| .gitignore | ||
| .gitmodules | ||
| .lintrunner.toml | ||
| build.bat | ||
| build.sh | ||
| CITATION.cff | ||
| CODEOWNERS | ||
| CONTRIBUTING.md | ||
| lgtm.yml | ||
| LICENSE | ||
| NuGet.config | ||
| ort.wprp | ||
| ORT_icon_for_light_bg.png | ||
| Package.swift | ||
| packages.config | ||
| pyproject.toml | ||
| README.md | ||
| requirements-dev.txt | ||
| requirements-doc.txt | ||
| requirements-lintrunner.txt | ||
| requirements-training.txt | ||
| requirements.txt.in | ||
| SECURITY.md | ||
| setup.py | ||
| ThirdPartyNotices.txt | ||
| VERSION_NUMBER | ||
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
-
General Information: onnxruntime.ai
-
Usage documention and tutorials: onnxruntime.ai/docs
-
YouTube video tutorials: youtube.com/@ONNXRuntime
-
Companion sample repositories:
- ONNX Runtime Inferencing: microsoft/onnxruntime-inference-examples
- ONNX Runtime Training: microsoft/onnxruntime-training-examples
Builtin Pipeline Status
| System | Inference | Training |
|---|---|---|
| Windows | ||
| Linux | ||
| Mac | ||
| Android | ||
| iOS | ||
| Web | ||
| Other |
Third-party Pipeline Status
| System | Inference | Training |
|---|---|---|
| Linux |  |
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.