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.

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)

# set num_experts_per_tok 1 or 2 ?
model.config.num_experts_per_tok = 2

# message
messages = [
    {"role": "user", "content": "Tell me what's for dinner tonight."},
]

with torch.no_grad():
    token_ids = tokenizer.apply_chat_template(messages, return_tensors="pt")
    output_ids = model.generate(
        token_ids.to(model.device),
        temperature=0.5,
        do_sample=True,
        top_p=0.95,
        top_k=40,
        max_new_tokens=128,
        repetition_penalty=1.5
    )
output = tokenizer.decode(output_ids[0][token_ids.size(1) :])
print(output)