search

ROCm / AMDMIGraphX

AMD's graph optimization engine.
https://rocm.docs.amd.com/projects/AMDMIGraphX/en/latest/
MIT License
185 stars 85 forks source link

Allow parsing of E4M3FN models using scales manipulation #2533

Open umangyadav opened 11 months ago

umangyadav commented 11 months ago

If the E4M3FN model is quantized using QDQ pairs then, it can be converted into E4M3FNUZ types by multiplying scales with some constants.

This would allow using E4M3FN models inside MIGraphX.

fxmarty commented 6 months ago

@umangyadav I am curious whether there is a conversion E4M3FN + scale <-> E4M3FNUZ + scale implemented anywhere?

umangyadav commented 6 months ago

@fxmarty We don't have that logic implemented in MIGraphX yet. I have an issue open that describes the logic https://github.com/ROCm/AMDMIGraphX/issues/2717 It's pretty simple to convert, but haven't had chance to work on it.