Converting Compressed LLaMA2 Model to Hugging Face-Compatible Format

[pavancshekar-dev]

Issue: Converting Compressed LLaMA2 Model to Hugging Face-Compatible Format

Description

We have successfully compressed a LLaMA2 model with 4.4 billion parameters. However, I am encountering issues when trying to convert the compressed model to a Hugging Face-compatible format. Specifically, when I use the model.save_pretrained(output_dir) and tokenizer.save_pretrained(output_dir) methods, the model parameters revert to the original 6.7 billion, and the output becomes worse and incoherent.

Steps to Reproduce

1. Compress a LLaMA2 model to 4.4 billion parameters.

2. Use the following code to save the model:

   import torch
   from transformers import AutoModelForCausalLM, AutoTokenizer
   
   def save_compressed_model(model, tokenizer, output_dir):
       # Save the model and tokenizer using Hugging Face's save_pretrained method
       model.save_pretrained(output_dir, safe_serialization=True)
       tokenizer.save_pretrained(output_dir)
   
   # Load your compressed model
   model_path = "path_to_your_compressed_model"
   tokenizer_path = "path_to_your_tokenizer"
   output_dir = "path_to_output_directory"
   
   model = torch.load(model_path)
   tokenizer = AutoTokenizer.from_pretrained(tokenizer_path)
   
   # Save the model and tokenizer
   save_compressed_model(model, tokenizer, output_dir)
   

3. Attempt to use the model from the output directory.

Observed Behavior

• The model parameters revert to the original 6.7 billion.
• The model output becomes worse and generates random gibberish.

Expected Behavior

• The model should retain its compressed state with 4.4 billion parameters.
• The model output should remain coherent and consistent with the compressed model's performance.

Additional Context

I have also attempted to convert the model to GGUF format, but encountered similar issues. Any guidance on correctly converting and saving the compressed model for Hugging Face would be greatly appreciated.

Thank you for your assistance!

[pavancshekar-dev]

Just to simplify, we are able compress and use the svdllm models. However, we are unable to convert them to Hugging Face formats like safetensors or GGUF. All our conversion attempts have resulted in the models getting distorted or modified. Can you please help us figure this out?

[hsb1995]

Just to simplify, we are able compress and use the svdllm models. However, we are unable to convert them to Hugging Face formats like safetensors or GGUF. All our conversion attempts have resulted in the models getting distorted or modified. Can you please help us figure this out?

hello, this problem is done? i also have met this promblem

[dellixx]

hello I also encountered this problem, compressed after conversion, the model output garbled

Bot: cilechas ERR littTH avéessefopacitychasillessefilles Insidehardtées DamcilePortailitto ERR suivanteesteouléesaset ERRPortailcilexesunciVIDVIDпадаDelegèmesCommand av programmeekenèmesuncisef CivilhardtvierunciracссаDeleg suivante Civiléesées ERR Civilèmes littVIDhardtmaste Daminea ERRпадаDelegHellovidchasxes ERR suivanteèmesvidxeseed suivanteasetéesasetitto ERR ERRссаVID suivanteittenimesaset Dam ERRittoèmes Dam Aufxes ERRпадаvid curr

[dellixx]

Just to simplify, we are able compress and use the svdllm models. However, we are unable to convert them to Hugging Face formats like safetensors or GGUF. All our conversion attempts have resulted in the models getting distorted or modified. Can you please help us figure this out?

hello, this problem is done? i also have met this promblem

Hi, Have you solved it? I have the same problem now

[NamburiSrinath]

Hi folks,

Does anyone solved this problem? Converting to Huggingface supported format should be super useful for community.

cc @mi-zhang , @tuidan for any potential updates on this.