Many Tools

Developpers write many lines of code, many are part of a package, many are used to investigate what the first line produces. This section gathers some tools occasionally needed to write converters in sklearn-onnx, to implement kernels in onnxruntime, to add new operators in onnx. The first series is used to play with onnx files. A couple of the helpers described below are available through command lines.

onnx

• External Data and Big Models
  • Save a big model
  • Load a big model
  • Example with Llama
• Onnx Manipulations
  • Extract a subgraph
  • Analyze
• Quantization

The second series is used to investigate C++ implementations in onnxruntime.

onnxruntime

• Profiling onnxruntime
  • Time per type of node
  • Time per instance
• Debug Intermediate Results
• Compare multiple versions of onnxruntime
  • Step 1: save a test
  • Step 2: evaluate multiple versions of onnxruntime
• Trees
  • Parallelization implementation