File size: 17,154 Bytes
27651a9 32e007b 27651a9 6c2a460 27651a9 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 |
---
license: apache-2.0
datasets:
- allenai/dolma
---
# 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](https://github.com/allenai/OLMo).
We release the pre-trained model, supervised fine-tuned model, and DPO aligned model as follows:
- [AMD-OLMo-1B](https://huggingface.co/amd/AMD-OLMo-1B): Pre-trained on a subset of [Dolma v1.7](https://huggingface.co/datasets/allenai/dolma) that consists of 1.3 trillion tokens.
- [AMD-OLMo-1B-SFT](https://huggingface.co/amd/AMD-OLMo-1B-SFT): Supervised fine-tuned (SFT) on [Tulu V2](https://huggingface.co/datasets/allenai/tulu-v2-sft-mixture) dataset (1st phase) and then [OpenHermes-2.5](https://huggingface.co/datasets/teknium/OpenHermes-2.5), [WebInstructSub](https://huggingface.co/datasets/TIGER-Lab/WebInstructSub), and [Code-Feedback](https://huggingface.co/datasets/m-a-p/Code-Feedback) datasets (2nd phase).
- [AMD-OLMo-1B-SFT-DPO](https://huggingface.co/amd/AMD-OLMo-1B-SFT-DPO): Aligned with human preferences using Direct Preference Optimization (DPO) on [UltraFeedback](https://huggingface.co/datasets/argilla/ultrafeedback-binarized-preferences-cleaned) dataset.
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](https://github.com/allenai/OLMo) 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](https://www.amd.com/en/developer/resources/technical-articles/introducing-the-first-amd-1b-language-model.html).
## Usage
### PyTorch on AMD GPUs
For running pytorch on AMD GPUs you can use the following rocm docker as in [docker hub](https://hub.docker.com/r/rocm/pytorch)
```bash
docker pull rocm/pytorch:latest
# Inside docker
pip install transformers
```
### Use Example
```python
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](https://huggingface.co/TinyLlama/TinyLlama_v1.1) (1.1B) | [MobiLLaMA-1B](https://huggingface.co/MBZUAI/MobiLlama-1B) (1.2B) | [OLMo-1B](https://huggingface.co/allenai/OLMo-1B-hf) (1.2B) | [OpenELM-1_1B](https://huggingface.co/apple/OpenELM-1_1B) (1.1B) | [OLMo-1B-0724-hf](https://huggingface.co/allenai/OLMo-1B-0724-hf) (1.2B) | [AMD-OLMo-1B](https://huggingface.co/amd/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](https://huggingface.co/TinyLlama/TinyLlama-1.1B-Chat-v1.0) (1.1B)|[MobiLlama-1B-Chat](https://huggingface.co/MBZUAI/MobiLlama-1B-Chat) (1.2B)|[OpenELM-1_1B-Instruct](https://huggingface.co/apple/OpenELM-1_1B-Instruct) (1.1B)|[AMD-OLMo-1B-SFT](https://huggingface.co/amd/AMD-OLMo-1B-SFT) (1.2B)|[AMD-OLMo-1B-SFT-DPO](https://huggingface.co/amd/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](https://huggingface.co/TinyLlama/TinyLlama-1.1B-Chat-v1.0) (1.1B)|[MobiLlama-1B-Chat](https://huggingface.co/MBZUAI/MobiLlama-1B-Chat) (1.2B)|[OpenELM-1_1B-Instruct](https://huggingface.co/apple/OpenELM-1_1B-Instruct) (1.1B)|[AMD-OLMo-1B-SFT](https://huggingface.co/amd/AMD-OLMo-1B-SFT) (1.2B)|[AMD-OLMo-1B-SFT-DPO](https://huggingface.co/amd/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](https://huggingface.co/TinyLlama/TinyLlama-1.1B-Chat-v1.0) (1.1B)|[MobiLlama-1B-Chat](https://huggingface.co/MBZUAI/MobiLlama-1B-Chat) (1.2B)|[OpenELM-1_1B-Instruct](https://huggingface.co/apple/OpenELM-1_1B-Instruct) (1.1B)|[AMD-OLMo-1B-SFT](https://huggingface.co/amd/AMD-OLMo-1B-SFT) (1.2B)|[AMD-OLMo-1B-SFT-DPO](https://huggingface.co/amd/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](https://github.com/EleutherAI/lm-evaluation-harness): For evaluating on commonsense reasoning, multi-task understanding & responsible AI benchmarks
- [AlpacaEval](https://github.com/tatsu-lab/alpaca_eval): For evaluating instruction-following capabilities of chat models.
- [MT-Bench](https://github.com/lm-sys/FastChat/tree/main/fastchat/llm_judge): For evaluating multi-turn capabilities of chat models.
### Setup
```bash
# 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
```bash
# 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
```bash
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
```bash
# 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 $NUM_WORKERS
```
### Download and prepare SFT datasets
```bash
# 1st phase SFT dataset
python AMD-OLMo/prepare_sft_data.py --output_dir ./datasets/tulu --tokenizer tokenizers/allenai_eleuther-ai-gpt-neox-20b-pii-special.json --dataset tulu
# 2nd phase SFT dataset
python AMD-OLMo/prepare_sft_data.py --output_dir ./datasets/OpenHermes_WebInstructSub_CodeFeedBack --tokenizer tokenizers/allenai_eleuther-ai-gpt-neox-20b-pii-special.json --dataset 2nd-phase
```
### Run Training
Pretrainig config: [AMD-OLMo-1B.yaml](AMD-OLMo-1B.yaml)
SFT config: [AMD-OLMo-1B-SFT-1st-phase.yaml](AMD-OLMo-1B-SFT-1st-phase.yaml) and [AMD-OLMo-1B-SFT-2nd-phase.yaml](AMD-OLMo-1B-SFT-2nd-phase.yaml)
```bash
# Single node
HSA_FORCE_FINE_GRAIN_PCIE=1 OMP_NUM_THREADS=128 NCCL_DEBUG=INFO torchrun --nproc_per_node=8 ./scripts/train.py AMD-OLMo/AMD-OLMo-1B.yaml
# Multiple nodes
HSA_FORCE_FINE_GRAIN_PCIE=1 OMP_NUM_THREADS=128 NCCL_DEBUG=INFO torchrun --nnodes=$nnodes --node-rank=$node_rank --master_addr=$master_addr --master_port=$master_port --nproc_per_node=8 ./scripts/train.py AMD-OLMo/AMD-OLMo-1B.yaml
```
### Run DPO Training
DPO recipe: [AMD-OLMo-1B-dpo.yaml](AMD-OLMo-1B-dpo.yaml).
```bash
# install trl library
git clone https://github.com/huggingface/trl.git -b v0.8.6
# replace dpo_trainer.py
cp AMD-OLMo/dpo_trainer.py trl/trl/trainer
pip install -e ./trl
# install alignment-handbook
git clone https://github.com/huggingface/alignment-handbook.git hf-align
# 70769f9 is the main branch on 2024-04-11.
cd hf-align && git checkout 70769f9 && cd ..
pip install -e ./hf-align
# Copy AMD OLMo DPO recipe to hf-align/recipes.
cp AMD-OLMo/AMD-OLMo-1B-dpo.yaml hf-align/recipes/
# Prepare the converted AMD-OLMo SFT Huggingface model to ckpt_dir.
ckpt_dir=amd/AMD-OLMo-1B-SFT
local_tokenizer_dir=${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:
```bash
@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. |