Edit model card

AMD-OLMo

AMD-OLMo are a series of 1B language models trained from scratch by AMD on AMD Instinct™ MI250 GPUs. The training code used is based on OLMo. We release the pre-trained model, supervised fine-tuned model, and DPO aligned model as follows:

Description:

  • Hardware: Each compute node consists of 4 AMD Instinct™ MI250 GPUs. We use 16 nodes for pretraining AMD-OLMo-1B

  • Training throughput: 12,200 tokens/sec/gpu

  • Model architecture: AMD-OLMo-1B is based on the model architecture and training set up of fully open source 1 billion version of OLMo-1B with the details below:

    Parameter size Number of layers Number of heads Hidden size Context length Vocabulary Size
    1.2B 16 16 2048 2048 50,280
  • Hyper-parameters:

    Stage LR schedule Peak LR Warmup steps Epochs Batch size (tokens)
    Pretraining Cosine 4.0e-4 2000 1 4M
    SFT Phase 1 Linear 2.0e-5 200 3 262K
    SFT Phase 2 Linear 2.0e-5 200 3 1024K
    DPO Cosine 4.0e-6 47 1 64K

For more details, please refer to our blog.

Usage

PyTorch on AMD GPUs

For running pytorch on AMD GPUs you can use the following rocm docker as in docker hub

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

Use Example

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])

Main Results

Pretraining Results

Standard Benchmarks TinyLLaMA-v1.1 (1.1B) MobiLLaMA-1B (1.2B) OLMo-1B (1.2B) OpenELM-1_1B (1.1B) OLMo-1B-0724-hf (1.2B) AMD-OLMo-1B (1.2B)
arc_easy 55.47 56.65 57.28 55.43 56.65 63.64
arc_challenge 32.68 32.00 31.06 32.34 32.34 33.70
hellaswag 61.47 61.80 62.92 64.81 66.12 63.61
piqa 73.56 75.30 75.14 75.57 75.08 75.57
boolq 55.99 60.83 61.74 63.58 66.18 60.58
sciq 89.30 88.20 87.00 90.60 92.70 93.20
winogrande 59.43 59.27 59.98 61.72 61.72 61.64
openbookqa 36.80 35.40 36.20 36.20 35.60 35.80
mmlu (0-shot) 25.02 24.81 24.23 25.26 25.45 24.88
gsm8k (8-shot) 1.82 0.00 2.50 2.81 8.95 2.88
bbh (3-shot) 25.63 0.00 25.63 16.77 21.67 20.95
Average 47.02 44.93 47.61 47.73 49.31 48.77

Instruction Tuning Results

Standard Benchmarks TinyLlama-1.1B-Chat-v1.0 (1.1B) MobiLlama-1B-Chat (1.2B) OpenELM-1_1B-Instruct (1.1B) AMD-OLMo-1B-SFT (1.2B) AMD-OLMo-1B-SFT-DPO (1.2B)
arc_easy 54.42 57.41 52.44 63.68 64.31
arc_challenge 32.85 34.56 37.80 37.12 37.37
hellaswag 60.40 62.51 71.29 61.63 61.91
piqa 74.48 75.73 75.03 74.43 74.16
boolq 61.04 55.66 70.28 68.53 70.24
sciq 88.40 87.10 89.50 91.20 92.10
winogrande 60.54 60.77 62.19 60.22 60.62
openbookqa 37.20 36.80 39.20 37.40 40.20
mmlu 24.61 25.25 25.54 29.97 30.52
gsm8k (8-shot) 2.81 0.23 1.82 18.20 15.77
bbh (3-shot) 26.83 0.00 13.40 25.17 25.45
Average 47.60 45.09 48.95 51.60 52.06
Chat Benchmarks TinyLlama-1.1B-Chat-v1.0 (1.1B) MobiLlama-1B-Chat (1.2B) OpenELM-1_1B-Instruct (1.1B) AMD-OLMo-1B-SFT (1.2B) AMD-OLMo-1B-SFT-DPO (1.2B)
AlpacaEval 1 (Win Rate) 50.81 34.90 37.72 50.12 54.22
AlpacaEval 2 (LC Win Rate) 1.54 1.59 0.49 3.88 2.37
MTBench 3.38 2.89 - 4.35 4.10
Responsible AI Benchmarks TinyLlama-1.1B-Chat-v1.0 (1.1B) MobiLlama-1B-Chat (1.2B) OpenELM-1_1B-Instruct (1.1B) AMD-OLMo-1B-SFT (1.2B) AMD-OLMo-1B-SFT-DPO (1.2B)
ToxiGen 41.70 37.23 42.34 39.04 39.68
crows_pairs 60.35 58.50 59.93 60.29 61.00
TruthfulQA-mc2 37.92 38.46 45.84 37.45 40.06

*In generating tokens for chat benchmark evaluations, we use max_length=2048 for AlpacaEval and max_new_tokens=2048 for MTBench.

*All numbers in above tables were obtained from our evaluations.

Evaluation

We use the following open source evaluation frameworks for evaluating our models:

  • Language Model Evaluation Harness: For evaluating on commonsense reasoning, multi-task understanding & responsible AI benchmarks
  • AlpacaEval: For evaluating instruction-following capabilities of chat models.
  • MT-Bench: For evaluating multi-turn capabilities of chat models.

Setup

# 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]"

Run evaluation

# 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

Training

Setup

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 and prepare pretraining datasets

# 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

Download and prepare SFT datasets

# 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

Run Training

Pretrainig config: AMD-OLMo-1B.yaml

SFT config: AMD-OLMo-1B-SFT-1st-phase.yaml and AMD-OLMo-1B-SFT-2nd-phase.yaml

# 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

Run DPO Training

DPO recipe: AMD-OLMo-1B-dpo.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

Bias, Risks, and Limitations

  • The models are being released for research purposes only and are not intended for use cases that require high levels of factuality, safety critical situations, health or medical applications, generating false information, facilitating toxic conversations.
  • Model checkpoints are made accessible without any safety guarantees. It is crucial for users to conduct comprehensive evaluations and implement safety filtering mechanisms as per their respective use cases.
  • It may be possible to prompt the model to generate content that may be factually inaccurate, harmful, violent, toxic, biased, or otherwise objectionable. Such content may also get generated by prompts that did not intend to produce output as such. Users are thus requested to be aware of this and exercise caution and responsible thinking when using the model.
  • Multi-lingual abilities of the models have not been tested and thus may misunderstand and generate erroneous responses across different languages.

Appendix

Evaluation Metrics

Benchmark Metric
arc_easy Normalized Accuracy
arc_challenge Normalized Accuracy
hellaswag Normalized Accuracy
piqa Accuracy
boolq Accuracy
sciq Accuracy
winogrande Accuracy
openbookqa Normalized Accuracy
mmlu Accuracy
gsm8k (8-shot) Exact Match (Flexible Extract)
bbh (3-shot) Exact Match
ToxiGen Accuracy
crows_pairs PCT Stereotype
TruthfulQA-mc2 Accuracy
AlpacaEval 1 (Win Rate) Win Rate (chatgpt_fn)
AlpacaEval 2 (LC Win Rate) Length Control Win Rate (weighted_alpaca_eval_gpt4_turbo)
MTBench Average score for single-answer grading (2 turns)

Feel free to cite our AMD-OLMo models:

@misc{AMD-OLMo,
    title = {AMD-OLMo: A series of 1B language models trained from scratch by AMD on AMD Instinct™ MI250 GPUs.},
    url = {https://huggingface.co/amd/AMD-OLMo},
    author = {Jiang Liu, Jialian Wu, Prakamya Mishra, Zicheng Liu, Sudhanshu Ranjan, Pratik Prabhanjan Brahma, Yusheng Su, Gowtham Ramesh, Peng Sun, Zhe Li, Dong Li, Lu Tian, Emad Barsoum},
    month = {October},
    year = {2024}
}

License

Copyright (c) 2018-2024 Advanced Micro Devices, Inc. All Rights Reserved.

Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at

http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License.

Downloads last month
236
Safetensors
Model size
1.18B params
Tensor type
F32
·
Inference Examples
Unable to determine this model's library. Check the docs .

Model tree for amd/AMD-OLMo-1B-SFT-DPO

Quantizations
5 models

Dataset used to train amd/AMD-OLMo-1B-SFT-DPO

Collection including amd/AMD-OLMo-1B-SFT-DPO