Paper | Usage | Citation | Presentation
This repository contains the official code for our NeurIPS 2022 paper AMP. AMP is an automatic approach to find fast model-parallel strategies to train large Deep Learning models. We design AMP to tackle real-world scnerios where users are training hetergeneous models with uneven layers and hetergeneous cluster with mixed generations of GPUs. Concretely, it contributes
- A valid representation of model-parallelism strategies.
- A cost model that accurately predicts the running time of a strategy without launching expensive real trials.
- An automatic optimization procedure that uses the cost model and a dynamic programming algorithm to efficiently find fast strategies.
AMP finds strategies that have similar performance to the state-of-the-art strategy finder[1] when no heterogeneity in the model and in the cluster. AMP fins strategies that are 1.54x better than the SOTA when heterogeneity exists in the cluster, and 1.77x better when heterogeneity exists in the model. In particular, the cost model in AMP can accurately predict low costs for top strategies.
We provide two settings: (1) use AMP to predict top strategies, (2) Additionally launch real trials with DeepSpeed to validate the ground truth runtime. Setting 1 requires a single CPU, while Setting 2 requires 16 GPUs in AWS EC2 (we provide the instance details in the paper). We have installed the environment and prepare necessary intermediate results for Setting 2 in an AMI for ease of setup.
cd ~
git clone https://github.com/MccRee17/AMP
conda create -n amp python=3.7.3
conda activate amp
conda install numpy ninja pyyaml mkl mkl-include setuptools cmake cffi typing_extensions future six requests
pip install tqdm spur torch==1.7.1+cu110 torchvision==0.8.2+cu110 -f https://download.pytorch.org/whl/torch_stable.html
Use our AWS AMI:
| AMI Name | AMI ID | Region | Instances in the paper |
|---|---|---|---|
| AMP-Oct-31 | ami-011d5dd7f6fe79d32 | us-west-2 | G4dn.12xlarge, P3.8xlarge |
Launch instances specified in the paper, e.g. 4 G4dn.12xlarge instances for the homogeneous experiment. Lauch them within the same placement group using the cluster strategy in EC2 so that they have the maximum bandwidth. Assume that AWS assigns 4 machines with IP address IP1, IP2, IP3, and IP4. Then do several steps on the master Machine IP1:
- ssh into IP[1-4] and exit to store the public ssh key of all machines in IP1, so the ssh verification does not prompt during trials.
- Add IP[1-4] to
~/hostfileand state the number of GPUs in each machine (DeepSpeed Tutorial). For instance, all 4x4 clusters in the paper are specified by:
IP1 slots=4
IP2 slots=4
IP3 slots=4
IP4 slots=4
- Activate our environment by
source anaconda3/bin/activate; conda activate amp.
Suggestions: (1) Warm up each AWS machine before running, otherwise trials may get terminated by timeout. A simple warmup is python; import torch; a = torch.randn(100,100).cuda() (2) If some trials hang, one can manually login to each machine and kill GPU processes. The optimization algorithms runs on CPU and will not be affected. A useful command to check processes on GPU: sudo fuser -v /dev/nvidia*. (3) If processes constantly get stuck, try removing all caches by rm -rf ~/amp_simulate; rm -rf ~/tmp. If there are other blockers in launching distributed experiments, please leave an issue here or send the author an email.
With Setting 1:
cd ~/AMP/src
python homogeneous.py
This will finish in around 500 seconds and store the result in ~/amp_main_logs/homogeneous_[time_stamp].txt.
With Setting 2:
cd ~/AMP/DeepSpeed/DeepSpeedExamples/Megatron-LM-v1.1.5-3D_parallelism
python homogeneous.py --full --budget 10
This will run the prediction and launch top 10 predicted strategies. It will finish in around 1500 seconds and store the result in ~/amp_main_logs/homogeneous_[time_stamp].txt. To run x numbers of real trials, use argument --budget x. The raw log from our modified DeepSpeed contains a lot of details such as the pipeline schedule, we recommend redirecting it into another log.txt for further interpretation.
Cached results with 53 real trials are in AMP/src/results/homogeneous_results.txt and logs in AMP/src/results/homogeneous_log.txt.
With Setting 1:
cd ~/AMP/src
python het_cluster.py
This will finish in around 500 seconds and store the result in ~/amp_main_logs/het_cluster_[time_stamp].txt.
With Setting 2:
cd ~/AMP/DeepSpeed/DeepSpeedExamples/Megatron-LM-v1.1.5-3D_parallelism
python het_cluster.py --full --budget 10
Predicting and 10 real trials takes around 1600 seconds. Cached results with 53 real trials are in AMP/src/results/het_cluster_results.txt and logs in AMP/src/results/het_cluster_log.txt.
With Setting 1:
cd ~/AMP/src
python het_model.py
This will finish in around 200 seconds and store the result in ~/amp_main_logs/het_model_[time_stamp].txt.
With Setting 2:
cd ~/AMP/DeepSpeed/DeepSpeedExamples/Megatron-LM-v1.1.5-3D_parallelism
python het_model.py --full --budget 10
Predicting and 10 real trials takes around 1300 seconds. Cached results with 65 real trials are in AMP/src/results/het_model_results.txt and logs in AMP/src/results/het_model_log.txt.
Basic logic of AMP is implemented in several files:
- The main function (homogeneous.py, het_cluster.py, het_model.py) iteratively applies the cost model and optimize.
- cost_xxx.py implements the cost model.
- pipe.py implements the dynamic programming algorithm.
- sa.py provides possible candidates for the main function.
- amp_utils.py implements other functions such as launching real trials with given configurations.
If you find this repository useful, please cite our paper using
@article{li2022amp,
title={AMP: Automatically Finding Model Parallel Strategies with Heterogeneity Awareness},
author={Li, Dacheng and Wang, Hongyi and Xing, Eric and Zhang, Hao},
journal={arXiv preprint arXiv:2210.07297},
year={2022}
}
[1] Narayanan, Deepak, et al. "Efficient large-scale language model training on gpu clusters using megatron-lm." Proceedings of the International Conference for High Performance Computing, Networking, Storage and Analysis. 2021.