microsoft / microsoft/onnxruntime
[Feature Request] Fp6 datatype support
- Dominant language
- C++
- Stars
- 21.9k
- Forks
- 4.2k
- Avg merge
- 4d 11h
- Merged PRs (30d)
- 184
Description
### Describe the feature request
Allow for Onnxruntime to parse and support the FP6 datatype for Onnx models.
Currently in the Onnx spec this doesn't exist yet but I've gotten a feature request started on ONNX as well
https://github.com/onnx/onnx/issues/7048
Lookng to leverage this within MIGraphX as well as the MIGraphX Execution Provider for prequantized models
### Describe scenario use case
Having execution providers take in fp6 (any bit scheme) quantized models and support inference
Contributor guide
Research direction
Start by reviewing the linked ONNX issue 7048 and the existing MIGraphX and MIGraphX Execution Provider integration points. Determine how an FP6 datatype should be represented while the ONNX specification is still missing it, including the requested bit schemes. Done means ONNX Runtime execution providers can accept prequantized FP6 ONNX models and run inference.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- cpp, machine-learning
- Domain
- backend, machine-learning
- Issue type
- Feature
- Difficulty
- 5/5
- Estimated time
- Over a week
- Activity status
- Stale
- Clarity
- Needs clarification
- Newbie friendliness
- 25/100