metadata

license: apache-2.0
language:
  - fr
  - it
  - de
  - es
  - en
inference: false

Model Card for Mixtral-Fusion-4x7B-Instruct-v0.1

This model is an experimental model created by merging mistralai/Mixtral-8x7B-Instruct-v0.1 experts.

How we merged experts

We simply take the average of every two experts.weight.
The same goes for gate.weight.
Unfortunately, this model has a large hallucination. Look extraction version. -> mmnga/Mixtral-Extraction-4x7B-Instruct-v0.1

How To Convert

use colab cpu-high-memory.
convert_mixtral_8x7b_to_4x7b.ipynb

Usage

pip install git+https://github.com/huggingface/transformers --upgrade
pip install torch accelerate bitsandbytes flash_attn

from transformers import AutoTokenizer, AutoModelForCausalLM, MixtralForCausalLM
import torch

model_name_or_path = "mmnga/Mixtral-Fusion-4x7B-Instruct-v0.1"

tokenizer = AutoTokenizer.from_pretrained(model_name_or_path)
model = MixtralForCausalLM.from_pretrained(model_name_or_path, load_in_8bit=True)

text = "Tell me what's for dinner tonight. "
inputs = tokenizer(text, return_tensors="pt")

outputs = model.generate(**inputs, max_new_tokens=128)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))