[Feature request] make model combination possible (long shot!)

[skidd-level-100]

I would love to see a model combination tool! it would be awesome! we have loras which is awesome, but full fledged combining would be even better!

[SlyEcho]

How does model combination work?

Do you have some paper to link to so I could understand?

[FNsi]

How does model combination work?

Do you have some paper to link to so I could understand?

I think he means merge multi lora weights to the llama? For example merge both wizard-vicuña and starcode in the same time.

i do find it may work but take some time.
First convert_llama_weights_to_hf
Second merge hf with lora to pth
Third convert-pth-to-ggml.py

it's sad if you want to convert 30B, need 100G+ ram.

So there's only one way... copy the merged weights from llama.cpp runtime😂

[github-actions]

This issue was closed because it has been inactive for 14 days since being marked as stale.