AMD-OLMo-1B-SFT

docker pull rocm/pytorch:latest
# Inside docker
pip install transformers

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("amd/AMD-OLMo-1B-SFT").to("cuda") # remove .to("cuda") to load on cpu
tokenizer = AutoTokenizer.from_pretrained("amd/AMD-OLMo-1B-SFT")

prompt = "What is large language model?"
bos = tokenizer.eos_token
template = bos + "<|user|>\n{prompt}\n<|assistant|>\n"

input_text = template.format(prompt=prompt)
inputs = tokenizer([input_text], return_tensors='pt', return_token_type_ids=False).to("cuda")
outputs = model.generate(**inputs, max_new_tokens=1000, do_sample=True, top_k=50, top_p=0.95)
print(tokenizer.batch_decode(outputs, skip_special_tokens=True)[0])

# lm-eval-harness
git clone https://github.com/EleutherAI/lm-evaluation-harness
cd lm-evaluation-harness
pip install -e .

# AlpacaEval
pip install git+https://github.com/tatsu-lab/alpaca_eval
cd alpaca_eval
pip install -e .

# MT-Bench
git clone https://github.com/lm-sys/FastChat.git
cd FastChat
pip install -e ".[model_worker,llm_judge]"

# lm-eval-harness
HF_MODEL=amd/AMD-OLMo-1B-SFT-DPO
accelerate launch -m lm_eval --model hf \
    --model_args pretrained=$HF_MODEL,trust_remote_code=True \
    --tasks arc_easy,arc_challenge,hellaswag,piqa,boolq,sciq,winogrande,openbookqa,mmlu,gsm8k_cot,bbh_cot_fewshot,toxigen,truthfulqa,crows_pairs \
    --device cuda \
    --batch_size 32 \
    --output_path ./lm-eval-results/$HF_MODEL

WORK_DIR="<path_to_your_working_directory>"
cd $WORK_DIR
# Clone OLMo codebase:
git clone https://github.com/allenai/OLMo.git --branch v0.3.0
cd OLMo
# Clone AMD-OLMo that contains files to reproduce our model training
git clone https://huggingface.co/amd/AMD-OLMo

docker pull rocm/pytorch:latest
docker run -it --network=host --device=/dev/kfd --device=/dev/dri --group-add=video --ipc=host --cap-add=SYS_PTRACE --security-opt seccomp=unconfined --shm-size 8G -v $WORK_DIR/OLMo:/OLMo -w /OLMo rocm/pytorch:latest

# Remove Line 17 as the docker already has ROCm PyTorch installed
sed -i '17d' pyproject.toml
pip install -e .[all]

# Download
DATA_DIR=./datasets/dolma
mkdir -p $DATA_DIR

PARALLEL_DOWNLOADS="<number_of_parallel_downloads>"
cat "AMD-OLMo/dolma_v1_7_subset.txt" | xargs -n 1 -P $PARALLEL_DOWNLOADS wget -q -P $DATA_DIR

# Prepare
NUM_WORKERS="<number_of_workers>"
python scripts/prepare_memmap_dataset.py $DATA_DIR/*.json.gz -o $DATA_DIR/memmap_dataset --workers $NUM_WORKERS

# 1st phase SFT dataset
python AMD-OLMo/prepare_sft_data.py --output_dir ./datasets/tulu --tokenizer tokenizers/allenai_eleuther-ai-gpt-neox-20b-pii-special.json --dataset tulu

# 2nd phase SFT dataset
python AMD-OLMo/prepare_sft_data.py --output_dir ./datasets/OpenHermes_WebInstructSub_CodeFeedBack --tokenizer tokenizers/allenai_eleuther-ai-gpt-neox-20b-pii-special.json --dataset 2nd-phase

# Single node
HSA_FORCE_FINE_GRAIN_PCIE=1 OMP_NUM_THREADS=128 NCCL_DEBUG=INFO torchrun --nproc_per_node=8 ./scripts/train.py AMD-OLMo/AMD-OLMo-1B.yaml

# Multiple nodes
HSA_FORCE_FINE_GRAIN_PCIE=1 OMP_NUM_THREADS=128 NCCL_DEBUG=INFO torchrun --nnodes=$nnodes --node-rank=$node_rank --master_addr=$master_addr --master_port=$master_port --nproc_per_node=8 ./scripts/train.py AMD-OLMo/AMD-OLMo-1B.yaml

# install trl library
git clone https://github.com/huggingface/trl.git -b v0.8.6

# replace dpo_trainer.py
cp AMD-OLMo/dpo_trainer.py trl/trl/trainer

pip install -e ./trl

# install alignment-handbook
git clone https://github.com/huggingface/alignment-handbook.git hf-align
# 70769f9 is the main branch on 2024-04-11.
cd hf-align && git checkout 70769f9 && cd .. 

pip install -e ./hf-align

# Copy AMD OLMo DPO recipe to hf-align/recipes.
cp AMD-OLMo/AMD-OLMo-1B-dpo.yaml hf-align/recipes/

# Prepare the converted AMD-OLMo SFT Huggingface model to ckpt_dir.
ckpt_dir=amd/AMD-OLMo-1B-SFT
local_tokenizer_dir=${ckpt_dir}

# Set output checkpoint dir.
dpo_ckpt_dir=<your_output_checkpoint_dir>

accelerate launch --config_file hf-align/recipes/accelerate_configs/deepspeed_zero3.yaml \
hf-align/scripts/run_dpo.py hf-align/recipes/AMD-OLMo-1B-dpo.yaml \
--trust_remote_code=true \
--model_name_or_path=${ckpt_dir} \
--tokenizer_name_or_path=${local_tokenizer_dir} \
--output_dir=${dpo_ckpt_dir} \
--num_train_epochs=1 \
--learning_rate=4e-6 \
--beta=0.3 \
--loss_type=sigmoid

@article{instella,
  title={Instella: Fully Open Language Models with Stellar Performance},
  author={Liu, Jiang and Wu, Jialian and Yu, Xiaodong and Su, Yusheng and Mishra, Prakamya and Ramesh, Gowtham and Ranjan, Sudhanshu and Manem, Chaitanya and Sun, Ximeng and Wang, Ze and Brahma, Pratik Prabhanjan and Liu, Zicheng and Barsoum, Emad},
  journal={arXiv preprint arXiv:2511.10628},
  year={2025}
}