/OLMoE

OLMoE: Open Mixture-of-Experts Language Models

Primary LanguageJupyter NotebookApache License 2.0Apache-2.0

OLMoE: Open Mixture-of-Experts Language Models

Fully open, state-of-the-art Mixture of Expert model with 1.3 billion active and 6.9 billion total parameters. All data, code, and logs released.


This repository provides an overview of all resources for the paper "OLMoE: Open Mixture-of-Experts Language Models".

Artifacts

Inference

Install the transformers & torch libraries and run (Transformers must be from source for this PR or until the next release):

from transformers import OlmoeForCausalLM, AutoTokenizer
import torch

DEVICE = "cuda" if torch.cuda.is_available() else "cpu"

# Load different ckpts via passing e.g. `revision=step10000-tokens41B`
# also check allenai/OLMoE-1B-7B-0924-SFT & allenai/OLMoE-1B-7B-0924-Instruct
model = OlmoeForCausalLM.from_pretrained("allenai/OLMoE-1B-7B-0924").to(DEVICE)
tokenizer = AutoTokenizer.from_pretrained("allenai/OLMoE-1B-7B-0924")
inputs = tokenizer("Bitcoin is", return_tensors="pt")
inputs = {k: v.to(DEVICE) for k, v in inputs.items()}
out = model.generate(**inputs, max_length=64)
print(tokenizer.decode(out[0]))
# > # Bitcoin is a digital currency that is created and held electronically. No one controls it. Bitcoins aren’t printed, like dollars or euros – they’re produced by people and businesses running computers all around the world, using software that solves mathematical

You can list all revisions/branches by installing huggingface-hub & running:

from huggingface_hub import list_repo_refs
out = list_repo_refs("allenai/OLMoE-1B-7B-0924")
branches = [b.name for b in out.branches]

Pretraining

  1. Clone this OLMo branch & create an environment with its dependencies via cd OLMo; pip install -e .. If you want to use new features in OLMo clone from the main branch instead.
  2. Run pip install git+https://github.com/Muennighoff/megablocks.git@olmoe
  3. Setup a config file. configs/OLMoE-1B-7B-0924.yml was used for the pretraining of OLMoE-1B-7B-0924. You can find configs from various ablations in configs/ablations.
  4. Download the data from https://hf.co/datasets/allenai/OLMoE-mix-0924, tokenize it via the command below and adapt the paths in your training config to point to it.
dolma tokens \
--documents ${PATH_TO_DOWNLOADED_DATA} \
--destination ${PATH_WHERE_TO_SAVE_TOKENIZED_DATA} \
--tokenizer.name_or_path 'allenai/gpt-neox-olmo-dolma-v1_5' \
--max_size '2_147_483_648' \
--seed 0 \
--tokenizer.eos_token_id 50279 \
--tokenizer.pad_token_id 1 \
--processes ${NUMBER_OF_CPU_CORES_TO_USE}
  1. Submit your job. We used bash scripts/olmoe-gantry.sh which invokes https://github.com/allenai/OLMo/blob/Muennighoff/MoE/scripts/train.py and uses beaker gantry but you will likely need to change the script to work with your setup.
  2. To convert you pretraining checkpoint to Hugging Face transformers after training, you can use the script & instructions here.

Other design choices

For most of our experiments on other design choices, you can simply set them in the config file (e.g. change the respective hyperparam), except for:

  1. Sparse upcycling: To sparse upcycle your model, train it dense first using e.g. this config, then convert any of its checkpoints into an MoE using this script & its instructions at the top while making sure to modify the hardcoded values (num experts etc) as you'd like your model to be, then place the newly created model (model_sparse.safetensors) into a new folder with a name that ends in -unsharded and place the model file inside of it with the name model.safetensors, then launch a job that loads this model similar to our sparse upcycling job (note the settings --load_path=path_to_upcycled_ckpt --reset_optimizer_state=True --reset_trainer_state=True and --fast_forward_batches=XXX if you also want to continue on the same dataset with the same order). Also make sure to have the changes from this PR in your code: allenai/OLMo#573. Finally, if you want to reproduce upcycling from OLMo-1B (0724) as in the paper, the OLMo 1B checkpoint turned into an MoE with 8 experts to start from is here: https://huggingface.co/allenai/OLMo-1B-0724-954000steps-unsharded; download the files inside of it (e.g. wget https://huggingface.co/allenai/OLMo-1B-0724-954000steps-unsharded/resolve/main/model.safetensors), then use a config similar to this one to train the upcycled MoE from it.
  2. Expert choice: To run experiments with our expert choice implementation, you need to instead use the olmo branch Muennighoff/OLMoSE or simply copy over the small config changes that enable expert choice (i.e. here) to the Muennighoff/MoE branch. You can then run expert choice by activating it in your config (it will use this code: https://github.com/Muennighoff/megablocks/blob/4a25bc7b5665bcb9da93d72d5ad0c14d41e1a351/megablocks/layers/moe.py#L462 or https://github.com/Muennighoff/megablocks/blob/4a25bc7b5665bcb9da93d72d5ad0c14d41e1a351/megablocks/layers/moe.py#L477 depending on your selection; both should be ~equivalent implementations of expert choice; neither was better than dropless token choice in our experiments)
  3. Shared MoE layers (Appendix): For these experiments you need to use the olmo branch Muennighoff/OLMoSE and create your own config e.g. like the one used in this run.

Adaptation

  1. Clone this open-instruct branch & follow its setup instructions. If you want to use new features in open-instruct clone from the main branch instead.
  2. SFT: After adapting as needed, run:
accelerate launch \
--mixed_precision bf16 \
--num_machines 1 \
--num_processes 8 \
--use_deepspeed \
--deepspeed_config_file configs/ds_configs/stage3_no_offloading_accelerate.conf \
open_instruct/finetune.py \
--model_name_or_path allenai/OLMoE-1B-7B-0924 \
--tokenizer_name allenai/OLMoE-1B-7B-0924 \
--use_flash_attn \
--max_seq_length 4096 \
--preprocessing_num_workers 128 \
--per_device_train_batch_size 2 \
--gradient_accumulation_steps 8 \
--learning_rate 2e-05 \
--lr_scheduler_type linear \
--warmup_ratio 0.03 \
--weight_decay 0.0 \
--num_train_epochs 2 \
--output_dir output/ \
--with_tracking \
--report_to wandb \
--logging_steps 1 \
--reduce_loss sum \
--model_revision main \
--dataset_mixer_list allenai/tulu-v3-mix-preview-4096-OLMoE 1.0 ai2-adapt-dev/daring-anteater-specialized 1.0 \
--checkpointing_steps epoch \
--add_bos
  1. DPO: Run
accelerate launch \
--mixed_precision bf16 \
--num_machines 1 \
--num_processes 8 \
--use_deepspeed \
--deepspeed_config_file configs/ds_configs/stage3_no_offloading_accelerate.conf \
open_instruct/dpo_tune.py \
--model_name_or_path allenai/OLMoE-1B-7B-0924-SFT \
--tokenizer_name allenai/OLMoE-1B-7B-0924-SFT \
--use_flash_attn \
--gradient_checkpointing \
--dataset_name argilla/ultrafeedback-binarized-preferences-cleaned \
--max_seq_length 4096 \
--preprocessing_num_workers 16 \
--per_device_train_batch_size 1 \
--gradient_accumulation_steps 4 \
--learning_rate 5e-7 \
--lr_scheduler_type linear \
--warmup_ratio 0.1 \
--weight_decay 0. \
--num_train_epochs 3 \
--output_dir output/ \
--report_to tensorboard \
--logging_steps 1 \
--reduce_loss sum \
--add_bos \
--checkpointing_steps epoch \
--dpo_beta 0.1
  1. KTO: Install trl and run https://github.com/Muennighoff/kto/blob/master/kto.py via WANDB_PROJECT=olmoe accelerate launch --config_file=config_8gpusdsz2_m7.yml kto.py --model_name_or_path allenai/OLMoE-1B-7B-0924-SFT --output_dir OLMoE-1B-7B-0924-SFT-KTO-3EP --report_to "wandb" --per_device_train_batch_size 4 --gradient_accumulation_steps 1 --optim rmsprop --learning_rate 5e-07 --beta 0.1 --logging_steps 1 --bf16 --sanity_check False --num_train_epochs 3 (if you want to run the Adam optimizer change to --optim adamw_torch). We used trl==0.9.6.

Evaluation

During pretraining

Evaluation during pretraining is done automatically and configured in the config file. It uses the code here: https://github.com/allenai/OLMo/tree/Muennighoff/MoE/olmo/eval.

After pretraining

OLMES Evals: Follow the instructions at https://github.com/allenai/OLMo-Eval/blob/51c5ba579e75ef4ce7e9b29936eaa72c1a0e99eb/olmo_eval/tasks/olmes_v0_1/README.md

DCLM Evals: Run scripts/run_dclm_evals* and refer to instructions from https://github.com/mlfoundations/dclm

After adaptation

Visuals

Citation

@misc{muennighoff2024olmoeopenmixtureofexpertslanguage,
      title={OLMoE: Open Mixture-of-Experts Language Models}, 
      author={Niklas Muennighoff and Luca Soldaini and Dirk Groeneveld and Kyle Lo and Jacob Morrison and Sewon Min and Weijia Shi and Pete Walsh and Oyvind Tafjord and Nathan Lambert and Yuling Gu and Shane Arora and Akshita Bhagia and Dustin Schwenk and David Wadden and Alexander Wettig and Binyuan Hui and Tim Dettmers and Douwe Kiela and Ali Farhadi and Noah A. Smith and Pang Wei Koh and Amanpreet Singh and Hannaneh Hajishirzi},
      year={2024},
      eprint={2409.02060},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2409.02060}, 
}