Thanks! Does this actually work? I guess I never considered just adding those matrices blindly and letting try_add sort out the layers. Can you post the results of the Heretic run?

Sure! here is the full log
heretic.log

model uploaded to Hugginface:
https://huggingface.co/Biogenic/granite-4.0-micro-heretic-uncensored

yeah same here, I wrote some code that more or less does the same as this PR

Abliteration parameters

Performance

give this a spin: https://huggingface.co/pszemraj/granite-4.0-h-1b-heretic

LFM2 models, like https://hf.co/LiquidAI/LFM2-2.6B are proving a bit trickier fwiw - it seems to need more than simply adding the conv layers (i even tried expanding the ranges too)

oh oops, my bad. I realized this PR is for the transformer MoE version of granite-4.0, I did the mamba hybrid one(s) (as.. that was the whole point) + LFM2

if you like, I can submit a separate PR?

Thanks, great stuff!

I made the same error as @pszemraj in assuming that this is the Mamba version. I actually wasn't aware that Granite also has a traditional transformer variant. Either way, it's great to have broader model support, the more the merrier.

@pszemraj Yes please.

Thank you guys, awesome project!)

Great! @p-e-w i'll take a look later today or tomorrow & fork/reconcile my changes vs latest main