LGAI-EXAONE

/

EXAONE-3.5-2.4B-Instruct

like 79

Follow

LG AI Research 224

"tie_word_embeddings" is true! The output_weight is just the embedding matrix....

Strange then that the produced GGUF I get is different from the one they posted 🤔

Hello, thank you for your question.

We basically use tie_word_embeddings to be True on 2.4B models, for memory efficiency on inference stage.
However, we found that conversion steps for AWQ and GGUF do not fit to the model with tie_word_embeddings, so we need to convert "Tied" model into "Un-tied" model before quantization.

When you try making another GGUF model on yourself, please copy the model.transformer.wte to model.lm_head and save it for quantization/conversion.

@bartowski Would you proceed as per nuxlear's comment?"

I'll look into how to do that later today and get back to you

Do you have any reference material on how to do this? From quick searching online I can't find any kind of reference to converting a "tied" model to an "un-tied" model..

Load w/ huggingface transformers, and copy the tensor there?

@bartowski

It's a bit unclear to me too. But, I made this tiny script and it worked, make sure to replace the directories:

import torch
from transformers import AutoModelForCausalLM

model = AutoModelForCausalLM.from_pretrained('/model/exaone_folder/')

model.lm_head.weight = torch.nn.Parameter(model.transformer.wte.weight.clone().detach())

print(model.lm_head.weight.requires_grad) 

model.tie_word_embeddings = False

model.save_pretrained('/model/exaone_folder/')

Finally got around to this and @QuantPanda 's script worked perfectly, thanks for that !