o Justin Chu, Microsoft – ONNX Converters
o Freddy Chiu, Intel – ONNX Optimizations
o Haihao Shen, Intel – Neural Compressor architect
o Huang Tai, Intel – Neural Compressor lead
o Feng Tian, Intel – Neural Compressor tech lead
o Ramalingam G, Microsoft – ONNX/ONNX Runtime
Documented charter: sigs/optimizations at main · onnx/sigs (github.com)
- Scope of interest so far
- Graph optimizations
- As different execution back-ends have preference for certain graph patterns, optimizing models becomes a non-trivial problem for developers
- Model converters today (e.g. PyTorch => ONNX) may produce different graphs that may not match patterns preferred by back-end to be deployed
- Need: Systematic way to get graph transformations to benefit ONNX community for their respective backend
- Next steps: Review/discuss existing solutions, and define goals to solve this need in the WG
- Model compression
- Goal is to have Neural Compressor in ONNX repo in ‘24
- Neural Compressor v2.4 released: Release Intel® Neural Compressor v2.4 Release · intel/neural-compressor (github.com)
- Graph optimizations
- Tentative monthly meeting time: 5PM Tuesdays
- Action Item: As tentative time conflicts with compiler SIG, Freddy to send out proposed alternatives to avoid conflict if possible
- Optimization challenges and how do we address them?
- Haihao: Previously had encountered challenges converting low-precision PyTorch models to ONNX.
- Justin: Support is on the way – Dynamo does not yet support this
- Haihao: Packing (e.g. INT4) is still a problem due to different ways to represent.
- Rama: there’s no consensus yet in the community on how to pack. Recommendation is to engage in PR and provide feedback. onnx/onnx#5811
- Haihao: How do we resolve opens regarding data types and flavors not yet available?
- Rama: Recommendation is to participate in Operators SIG and arch/infrastructure SIG for helping with PR proposals for defining new data types
- Haihao: Previously had encountered challenges converting low-precision PyTorch models to ONNX.
- Co-Chair – Neta Zmora stepped down. Open position for co-chair. Drop me a note if you’re interested!
- Ramakrishna unable to join because of zoom issues. Can we socialize zoom instructions.