NNFusion v0.4 Release Candidate

Fix variable name typo (#334)

Co-authored-by: Wenxiang Hu <[email protected]>

NNFusion v0.3 Release

Major Feature

• Support end-to-end BERT model training (in ONNX format) on real dataset
• Add new operator fusion passes for transformer-based model optimization
• Provide C++ and JSON interfaces for extending custom operators
• Support a new HLSL code generator

Others

• Update related documentations
• Fix bugs

中文版本说明快捷通道-->#105 (comment)

NNFusion v0.2 Release

Major Features

• Support the use of Python interface to accelerate the training and inference of PyTorch model
• Support low-precision and mixed-precision model compilation, e.g., fp16
• Provide auto kernel tuner integration:
  • Add Antares IR for 60+ ops
  • Support auto tuning via Antares tuning service
• Support parallel training via SuperScaler
• Enable local kernel cache through kernel database

Others

• Update related documentations
• Some enhancements on user experiences and bug fix

中文版本说明快捷通道-->#105 (comment)

NNFusion v0.1 Release

• Build and Installation:

  • Support out-of-box installation with docker image
  • Support source code install on native system and docker
  • Support devices like CUDA GPUs, and ROCm GPUs.

• Models, Framework and Operators:

  • Support DNN model formats including TensorFlow and ONNX
  • Support commonly used models including AlexNet, VGG11, ResNet50, seq2seq, BERT, etc.
  • Support more than 100 commonly used operators.

• Model Compilation and Execution:

  • Provide a full-stack optimization mechanism, including data-flow graph optimizations, model-specific kernel selection, kernel co-scheduling, etc.
  • Provide ahead-of-time and source-to-source(model-to-code) compilation to reduce runtime overhead
  • Remove third-party library or framework dependencies

• Usability:

  • Provide command line tool nnfusion
  • Provide tools for users to freeze TensorFlow and PyTorch models
  • Provide flexible way to customize optimization through direct code modification on generated code

中文版本说明快捷通道--> #72 (comment)