AXera Pulsar2 Toolchain Instructions Manual

copyright

• 1. Copyright Notice

Pulsar2 Toolchain

• 1. Documentation modifications and release notes
  • 1.1. Modify history
• 2. Introduction to AXera NPU (Neutron)
  • 2.1. Hardware terminology
  • 2.2. EU types
  • 2.3. Offline compilation and runtime modes
  • 2.4. Neutron architectures
• 3. Pulsar2 Toolchain overview
  • 3.1. Introduction
  • 3.2. Guide to the content of subsequent chapters
• 4. Quick Start
  • 4.1. Development environment preparation
  • 4.2. Quick Start(AX650)
  • 4.3. Quick Start(AX620E)
  • 4.4. Quick Start(AX615)
  • 4.5. Quick Start(M57)
  • 4.6. Quick Start(AX637)
  • 4.7. Quick Start(AX8860)
• 5. Advanced Guide to Model Conversion
  • 5.1. Overview
  • 5.2. Detailed explanation of model compilation
  • 5.3. Detailed explanation of multi-core compilation
  • 5.4. Detailed explanation of multi-batch compilation
  • 5.5. Detailed explanation layer by layer
  • 5.6. Detailed explanation of loading custom data sets
  • 5.7. Cascaded model adaptation
  • 5.8. Multi-input model configuration quantitative data set
  • 5.9. Multi-input model configuration quantization data set (NumpyObject)
  • 5.10. Detailed explanation of mixed precision quantization
  • 5.11. Enter size modification
  • 5.12. Configure model additional input dimensions
  • 5.13. Operator attribute modification
  • 5.14. Constant data modification
  • 5.15. Set separate compilation options for subgraphs
  • 5.16. Transformer model configuration details
  • 5.17. Quantized ONNX model import
  • 5.18. Color space conversion configuration
  • 5.19. Detailed explanation of advanced quantitative strategy configuration
• 6. Advanced Guide to Simulation Running
  • 6.1. Overview
  • 6.2. Detailed explanation of simulation operation
  • 6.3. Simulation run example
• 7. Advanced Guide to Model Deployment
  • 7.1. Overview
  • 7.2. Other Examples
• 8. Configuration file details
  • 8.1. Profile overview
  • 8.2. Complete json configuration reference
  • 8.3. Quantitative parameter description
  • 8.4. Quantitative precision analysis parameter description
  • 8.5. Preprocessing and postprocessing parameter description
  • 8.6. proto configuration definition
• 9. Instructions for using model evaluation tools
  • 9.1. Parameter Description
  • 9.2. Usage example
• 10. Functional Safety

Other Tools

• 1. Tool usage instructions of caffe2onnx
  • 1.1. Convert Caffe model to ONNX model
  • 1.2. Validate the converted ONNX model
• 2. QAT 4W4F support
• 3. Polygraphy tool overview
  • 3.1. Common parameters
  • 3.2. Quick start
• 4. NPU trace tool instructions
  • 4.1. NPU trace overview
  • 4.2. Get trace.json
  • 4.3. Trace features
  • 4.4. Common issue analysis

appendix

• 1. NPU Operator Support Lists
  • 1.1. NPU Operators support list (AX650)
  • 1.2. NPU Operators support list(AX620E)
  • 1.3. NPU Operators support list (AX615)
  • 1.4. NPU Operators support list(M57)
  • 1.5. NPU Operators support list (AX637)
• 2. Accuracy Tuning Suggestions
  • 2.1. Causes of loss of precision
  • 2.2. Deployment accuracy loss
  • 2.3. Model conversion accuracy loss
  • 2.4. Quantitative chemical single template
  • 2.5. Q&A
• 3. LLM Build (Compile)
  • 3.1. Command reference
  • 3.2. Download the ax-llm-build project
  • 3.3. Download Qwen3-0.6B
  • 3.4. Build (compile)
  • 3.5. Run on the development board
• 4. Model conversion examples
  • 4.1. YOLOv5s (object detection)
  • 4.2. YOLO11s (object detection)
  • 4.3. Depth-Anything-V2 (monocular depth estimation)
  • 4.4. CN-CLIP (Chinese multimodal text encoder)
• 5. Common configuration examples
  • 5.1. RGB input
  • 5.2. YUV input
  • 5.3. Static batch configuration
  • 5.4. Dynamic batch configuration
  • 5.5. Multi-input configuration
  • 5.6. Skip onnxslim