Guide creation of custom transformation passes for ONNX IR graphs.
npx skills add https://github.com/onnx/ir-py --skill pass-writing
Please help me install this Agent Skill. Skill: pass-writing Source: https://github.com/onnx/ir-py/tree/main/.github/skills/pass-writing Command: npx skills add https://github.com/onnx/ir-py --skill pass-writing
This Skill provides comprehensive guidance and best practices for developing transformation passes within the ONNX IR project, enabling users to effectively modify and optimize ONNX graphs.
InPlacePass, FunctionalPass) and pass lifecycle.Use the pass-writing skill to learn how to implement a new InPlacePass for ONNX IR.
High-intent search queries and answers about installing and using this skill.
To write a custom ONNX IR optimization pass, you implement a transformation by extending base classes like InPlacePass or FunctionalPass, then manipulate nodes and values to modify the graph structure safely.
InPlacePass modifies the ONNX graph directly during transformation, whereas FunctionalPass returns a new modified graph, offering different lifecycle patterns for ONNX graph optimization and analysis development.
Safe ONNX graph manipulation requires using the pass infrastructure's specific patterns for node replacement and removal, ensuring value references and initializer handling are correctly updated during the transformation pass.
Handling subgraphs and opset version management in ONNX IR passes involves specific graph traversal techniques to process nested structures and ensuring opset compatibility when modifying operations within the ONNX graph.
Yes, ONNX pass development requires explicit initializer handling, as initializers represent constant tensors that must be managed separately from dynamic graph values during node and value manipulation.
The best way to debug a custom ONNX IR transformation pass involves utilizing the pass infrastructure's built-in debugging features to inspect graph traversal, verify node manipulation, and validate subgraph processing outcomes.
amdExtracts ONNX model metadata needed for Quark PTQ planning and compatibility assessment.
JNZader-VaultDeploys ML models across Python, Rust, and Go using ONNX Runtime with quantization and graph transformations for edge scenarios.
kinnerkarmanishTraverse the knowledge graph to discover connections and synthesis opportunities.
mohitmishra786Generate, optimize, and inspect LLVM Intermediate Representation using Clang, opt, and llc.