You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Right now we don't have support for these models which will be a great addition so as a first step we need to investigate if we can even export them using executorch.
Sounds like figuring out if executorch can handle exporting PrismML Bonsai models is the first hurdle. I'd start by checking executorch's current export capabilities.
Yeah, that's exactly the scope and I think that 1bit quantization is not an option right now in ExexuTorch. The best we can do right now is what Torchao gives, which are 4bit quantization.
I checked and torchao itself has more agressive quantizations schemas like even 1-bit, but not supported in ExecuTorch. So I guess a natural follow-up for this one would be adding more aggresive ones to ExecuTorch ;)
Activity
[-]Try export PrismML Bonsai models[/-][+]Try to export PrismML Bonsai models[/+]PushpakAg commentedon Sep 16, 2026
Sounds like figuring out if executorch can handle exporting PrismML Bonsai models is the first hurdle. I'd start by checking executorch's current export capabilities.
msluszniak commentedon Sep 17, 2026
Yeah, that's exactly the scope and I think that 1bit quantization is not an option right now in ExexuTorch. The best we can do right now is what Torchao gives, which are 4bit quantization.
PushpakAg commentedon Sep 17, 2026
Thanks for the insight! I'll dive into what Torchao offers with 4bit quantization and see how it aligns with our needs for exporting these models.
msluszniak commentedon Sep 17, 2026
I checked and torchao itself has more agressive quantizations schemas like even 1-bit, but not supported in ExecuTorch. So I guess a natural follow-up for this one would be adding more aggresive ones to ExecuTorch ;)