Mistral Transformer

This repository contains minimal code to run our 7B model and to finetune it.
Blog: https://mistral.ai/news/announcing-mistral-7b/
Discord: https://discord.com/invite/mistralai

Getting started

Download the model

wget https://models.mistralcdn.com/mistral-7b-v0-1/mistral-7B-v0.1.tar
tar -xf mistral-7B-v0.1.tar

Note: The unzipped folder can be used as initial_model_path: in the training config.

Installation Hackathon

Upon running the Docker container, all necessary dependencies can be installed with:

pip install -r requirements_hackathon.txt

Using the trained model

Deployment

The deploy folder contains code to build a vLLM image with the required dependencies to serve the Mistral AI model. In the image, the transformers library is used instead of the reference implementation. To build it:

docker build deploy --build-arg MAX_JOBS=8

Instructions to run the image can be found in the official documentation.

Run the model

python -m main demo /path/to/mistral-7B-v0.1/
# To give your own prompts
python -m main interactive /path/to/mistral-7B-v0.1/

Change temperature or max_tokens using:

python -m main interactive /path/to/mistral-7B-v0.1/ --max_tokens 256 --temperature 1.0

If you want a self-contained implementation, look at one_file_ref.py, or run it with

python -m one_file_ref /path/to/mistral-7B-v0.1/

This is a test of the emergency broadcast system. This is only a test.

If this were a real emergency, you would be told what to do.

This is a test
=====================
This is another test of the new blogging software. I’m not sure if I’m going to keep it or not. I’m not sure if I’m going to keep
=====================
This is a third test, mistral AI is very good at testing. 🙂

This is a third test, mistral AI is very good at testing. 🙂

This
=====================

To run logits equivalence through chunking and sliding window, launch

python -m test_generate

Fine-tune the model

Data

Data must be stored in jsonl format files.

You can build two types of data files:

Pretrain: plain text data stored in the "text" key. E.g:

{"text": "Text contained in document n°1"}
{"text": "Text contained in document n°2"}

Instruct: conversational data stored in the "interactions" key in the form of a list. Each list item is a dictionary containing the "text" and "is_user" keys. is_user is a boolean, if it is equal to True the loss will not be calculated on these tokens. E.g.:

{"interactions": [{"is_user": true, "text": "User interaction n°1 contained in document n°1"}, {"is_user": false, "text": "Bot interaction n°1 contained in document n°1"}, {"is_user": true, "text": "User interaction n°2 contained in document n°1"}, {"is_user": false, "text": "Bot interaction n°2 contained in document n°1"}]}
{"interactions": [{"is_user": true, "text": "User interaction n°1 contained in document n°2"}, {"is_user": false, "text": "Bot interaction n°1 contained in document n°2"}, {"is_user": true, "text": "User interaction n°2 contained in document n°2"}, {"is_user": false, "text": "Bot interaction n°2 contained in document n°2"}, {"is_user": true, "text": "User interaction n°3 contained in document n°2"}, {"is_user": false, "text": "Bot interaction n°3 contained in document n°2"}]}

LoRA Finetuning

To benefit from a memory-efficient and performant finetuning, we recommend to use LoRA. The idea is to freeze weights and to only learn 1-2% additional weights in the form of low-rank matrix perturbations.

With proper tuning (carefully calibrated learning rate, rank, LoRA dropout, learning the LoRA weights as well as the normalization layers), LoRA finetuning effectively recovers the performance of full finetuning. We support DDP on top of that, meaning that training speed can be increased on multiple GPUs.

After the training, we merge the LoRA weights: hence, the saved checkpoint is exacly in the same format as one would get with full finetuning. To run a LoRA finetuning on a single GPU, use:

torchrun --nproc-per-node 1 --master_port $RANDOM -m train reference/7B_lora.yaml

phanxuanphucnd / hackathon