Skip to content

Latest commit

 

History

History
338 lines (249 loc) · 9.72 KB

STEP-BY-STEP-GUIDE.md

File metadata and controls

338 lines (249 loc) · 9.72 KB
Raw

How to train 🌿 Shurale7B-v1: Narrative based chit-chat model

Developed by @BobaZooba | CV | LinkedIn | bobazooba@gmail.com

Powered by X—LLM

GitHub Repo | Model based on Mistral-7B-v0.1

Project overview

All the code is located here: shurale/cli/*. Everything needed, including the dataset, has been implemented in the xllm project, so to train Shurale7B-v1, all you need to do is run the scripts from the command line. So, the most important thing is found in the scripts folder. These are bash scripts for all necessary actions in the project.

Script name Purpose
prepare.sh Download and prepare data, download model.
train.sh Train the model (this is usually done using LoRA).
train-deepspeed.sh Train the model (this is usually done using LoRA) using multi-gpu and DeepSpeed. Specify the correct number of GPUs used in the script.
fuse.sh Fuse LoRA and upload fused model to Huggingface Hub
quantize.sh GPTQ quantization of the fused model. Optional step.

Please check these useful materials:

  • Docs: here, we go into detail about everything the library can do
  • Demo project: here's a step-by-step example of how to use X—LLM and fit it into your own project
  • Template project: here's a template, a kickoff point you can use for your projects

Steps to reproduce

0. Run the correct environment

Cuda version must be >= 11.8

Good docker image: huggingface/transformers-pytorch-gpu:latest

1. Install X—LLM and other requirements

Install requirements: latest transformers and xllm (including deepspeed, flash-attn and auto-gptq)

pip install -r requirements-train.txt

Or

make install-train

2. Prepare .env file

You can find a file .env.template with the content:

HUGGING_FACE_HUB_TOKEN=
WANDB_API_KEY=
WANDB_ENTITY=
WANDB_PROJECT=
TOKENIZERS_PARALLELISM=false

Make a .env file with filled variables. HUGGING_FACE_HUB_TOKEN is necessary, everything about wandb is only needed for training. These variables will be loaded into the environment when any of the scripts are called.

3. Set up correct config

You need to set the correct config variables in all the scripts used (prepare, train, etc.). For example, the correct name of your project in HuggigFace Hub and so on.

Useful materials

4. Give access permissions to the scripts/*

To run bash scripts you need to give them permissions.

chmod 755 -R ./scripts/

Or

make scripts-access

5. Prepare dataset and model

Before you begin the training process, it's essential that you first download the necessary data and the model. The training won't proceed without this as it's during this stage that your training data gets prepared. We've separated this as a distinct step for a few reasons, chief among them being: if you're utilizing distributed training across several GPUs - for instance via DeepSpeed - you would otherwise end up redundantly downloading the dataset and model on each individual GPU, even though you only need to do it once.

Run this command line:

make prepare

Or

./scripts/prepare.sh

Or directly

python3 shurale/cli/prepare.py \
  --dataset_key soda \
  --model_name_or_path mistralai/Mistral-7B-v0.1 \
  --path_to_env_file ./.env

6. Train model

For more information about configuring your setup, refer to the X—LLM documentation. Make sure to fill in the correct values, such as your specific hub_model_id and other relevant settings.

Single GPU

Run this command line:

make train

Or

./scripts/train.sh

Or directly

python3 shurale/cli/train.py \
  --use_gradient_checkpointing True \
  --deepspeed_stage 0 \
  --stabilize True \
  --model_name_or_path mistralai/Mistral-7B-v0.1 \
  --use_flash_attention_2 False \
  --load_in_4bit True \
  --apply_lora True \
  --raw_lora_target_modules all \
  --per_device_train_batch_size 2 \
  --gradient_accumulation_steps 4 \
  --warmup_steps 1000 \
  --save_total_limit 0 \
  --push_to_hub True \
  --hub_model_id BobaZooba/Shurale7B-v1-LoRA \
  --hub_private_repo True \
  --report_to_wandb True \
  --logging_steps 1 \
  --num_train_epochs 3 \
  --save_steps 1000 \
  --max_steps 10050 \
  --save_safetensors True \
  --use_gradient_checkpointing True \
  --stabilize True \
  --max_length 2048 \
  --use_flash_attention_2 False \
  --prepare_model_for_kbit_training True \
  --load_in_4bit True \
  --apply_lora True \
  --label_smoothing_factor 0.1 \
  --path_to_env_file ./.env

Pay attention to fields

  • hub_model_id

Multiple GPU

This project utilizes DeepSpeed's Stage 2 offloading to facilitate multi-GPU training. However, you have the flexibility to modify this setting via the config arguments. For instance, you can switch the setting to use fsdp if needed.

Also, please specify the correct number of GPUs used in the script.

Run this command line:

make deepspeed-train

Or

./scripts/train-deepspeed.sh

Or directly

deepspeed --num_gpus=8 shurale/cli/train.py \
  --use_gradient_checkpointing True \
  --deepspeed_stage 2 --stabilize True \
  --model_name_or_path mistralai/Mistral-7B-v0.1 \
  --use_flash_attention_2 False \
  --load_in_4bit True \
  --apply_lora True \
  --raw_lora_target_modules all \
  --per_device_train_batch_size 4 \
  --gradient_accumulation_steps 4 \
  --warmup_steps 1000 \
  --save_total_limit 0 \
  --push_to_hub True \
  --hub_model_id BobaZooba/Shurale7B-v1-LoRA \
  --hub_private_repo True \
  --report_to_wandb True \
  --logging_steps 1 \
  --num_train_epochs 3 \
  --save_steps 1000 \
  --max_steps 10050 \
  --save_safetensors True \
  --use_gradient_checkpointing True \
  --stabilize True \
  --max_length 2048 \
  --use_flash_attention_2 False \
  --prepare_model_for_kbit_training True \
  --load_in_4bit True \
  --apply_lora True \
  --label_smoothing_factor 0.1 \
  --path_to_env_file ./.env

Pay attention to fields

  • hub_model_id

7. Fuse LoRA

Если вы обучали модель с использованием LoRA, то вам следует

If you trained the model using LoRA

Run this command line:

make fuse

Or

./scripts/fuse.sh

Or directly

python3 shurale/cli/fuse.py \
  --model_name_or_path mistralai/Mistral-7B-v0.1 \
  --lora_hub_model_id BobaZooba/Shurale7B-v1-LoRA \
  --hub_model_id BobaZooba/Shurale7B-v1 \
  --hub_private_repo True \
  --force_fp16 True \
  --fused_model_local_path ./fused_model/ \
  --max_shard_size 1GB \
  --path_to_env_file ./.env

Pay attention to fields

  • hub_model_id
  • lora_hub_model_id

8. [Optional] GPTQ quantization

Run this command line:

make quantize

Or

./scripts/quantize.sh

Or directly

python3 shurale/cli/quantize.py \
  --model_name_or_path ./fused_model/ \
  --apply_lora False \
  --stabilize False \
  --quantization_max_samples 64 \
  --quantized_model_path ./quantized_model/ \
  --prepare_model_for_kbit_training False \
  --quantized_hub_model_id BobaZooba/Shurale7B-v1-GPTQ \
  --quantized_hub_private_repo True \
  --low_cpu_mem_usage True \
  --path_to_env_file ./.env

Pay attention to fields

  • hub_model_id
  • quantized_hub_model_id

9. 🎉 Done! You are awesome!

Now your model was trained, fused (and maybe quantized) and saved to HuggingFace Hub.

You can load the model using transformers

from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("BobaZooba/Shurale7B-v1")
model = AutoModelForCausalLM.from_pretrained("BobaZooba/Shurale7B-v1")

Tale Quest

Tale Quest is my personal project which was built using xllm and Shurale. It's an interactive text-based game in Telegram with dynamic AI characters, offering infinite scenarios

You will get into exciting journeys and complete fascinating quests. Chat with George Orwell, Tech Entrepreneur, Young Wizard, Noir Detective, Femme Fatale and many more

Try it now: https://t.me/talequestbot