Mixtral 7b 8expert

🚀 Mixtral 7b 8 Expert

This is a preliminary HuggingFace implementation of the newly released MoE model by MistralAi, offering high - performance inference capabilities.

🚀 Quick Start

This is a preliminary HuggingFace implementation of the newly released MoE model by MistralAi. Make sure to load with trust_remote_code=True.

Thanks to @dzhulgakov for his early implementation (https://github.com/dzhulgakov/llama-mistral) that helped me find a working setup.

Also many thanks to our friends at LAION and HessianAI for the compute used for these projects!

💻 Usage Examples

Basic Usage

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("DiscoResearch/mixtral-7b-8expert", low_cpu_mem_usage=True, device_map="auto", trust_remote_code=True)
tok = AutoTokenizer.from_pretrained("DiscoResearch/mixtral-7b-8expert")
x = tok.encode("The mistral wind in is a phenomenon ", return_tensors="pt").cuda()
x = model.generate(x, max_new_tokens=128).cpu()
print(tok.batch_decode(x))

📚 Documentation

Benchmark scores

hella swag: 0.8661
winogrande: 0.824
truthfulqa_mc2: 0.4855
arc_challenge:  0.6638
gsm8k: 0.5709
MMLU: 0.7173

Conversion

Use convert_mistral_moe_weights_to_hf.py --input_dir ./input_dir --model_size 7B --output_dir ./output to convert the original consolidated weights to this HF setup.

📄 License

This project is licensed under the Apache - 2.0 license.

Come chat about this in our Disco(rd)! :)