ggml
ggml copied to clipboard
Implement Quick GELU
Closes #253
This is needed for CLIP-like models and I'm implementing clip.cpp here. It will also be a base for upcoming multimodal models that uses CLIP as an image embedder.
you let some formatting happen. please only commit code changes.
Oh sorry, missed that. I'll fix it.
This is ready for review now.
Hi @ggerganov Can ı have your attention here please? clip.cpp is almost ready, and I'm only crafting more examples and then I can announce it.
I can also add a link in the examples section in readme if you thing it deserves this.
p.s.: The next step will be using clip.cpp and llama.cpp to infer with LLaVA
Unfortunately that theoretically small divergence between GELU and Quick GELU lead to large differences at the end, I suppose it accumulates through 12 layers. So I couldn't get good results until implementing Quick GELU.