Skip to content

Latest commit

 

History

History
82 lines (62 loc) · 2.37 KB

LLaMA-MoE-v2.md

File metadata and controls

82 lines (62 loc) · 2.37 KB
Raw

🌴 Setup

dependencies:

cuda: 11.8 python: 3.11.10

Just follow the installation guide in README.md, which can be simplified as:

conda create -n smoe python=3.11
pip install torch==2.3.1 torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118
pip install -r requirements.txt
pip install flash-attn --no-build-isolation
git [email protected]:LLaMA-MoE/LLaMA-MoE-v2.git
cd LLaMA-MoE-v2
pip install -e .[dev]

Finally, ensure that you environments satisfy:

deepspeed==0.14.4
flash-attn==2.6.1
torch==2.3.1
triton==2.3.1
transformers==4.42.4

For disabling wandb during training, you may run the following command:

$ wandb disabled

🗃️ Data Preparation

Run the following commands, and the data would be prepared in resources/OpenHermes-2.5/openhermes2_5.jsonl .

$ mkdir resources
$ huggingface-cli download teknium/OpenHermes-2.5 --repo-type dataset --local-dir resources/OpenHermes-2.5 --local-dir-use-symlinks False

Create a file named json2jsonl.py in the resources/OpenHermes-2.5/ directory with the following content:

from smoe.utils.io import load_json, dump_jsonlines

data = load_json("resources/OpenHermes-2.5/openhermes2_5.json")
dump_jsonlines(data, "resources/OpenHermes-2.5/openhermes2_5.jsonl")

Use the following command to run the script:

$ srun -p MoE python resources/OpenHermes-2.5/json2jsonl.py

🧃 Model Preparation (Converting dense models to MoE)

  • Vanilla LLaMA-MoE-v2: sbatch scripts/expert_construction/convert/convert_mixtral_v2.sh
  • Residual LLaMA-MoE-v2: sbatch scripts/expert_construction/convert/convert_mixtral_residual_v2.sh

For more information, please refer to Expert Construction docs.

🚀 Training

  • packedpad Check the settings in scripts/sft/sft_8e_top2_base_res_pad.sh and run sbatch scripts/sft/sft_8e_top2_base_res_pad.sh .
  • nopad Check the settings in scripts/sft/sft_8e_top2_base_res.sh and run sbatch scripts/sft/sft_8e_top2_base_res.sh .

##SBATCH means the slurm setting is not activated.

🛫 Evaluation

$ git clone https://github.com/EleutherAI/lm-evaluation-harness
$ cd lm-evaluation-harness
$ git checkout d14b36e81aea4cef
$ pip install -e .
# copy the scripts in smoe - `scripts/sft/eval` to here
# change the model dir and evaluation settings
$ bash run.sh