https://github.com/MachineLearningSystem/23MLSYS-pipe-fisher

https://github.com/MachineLearningSystem/23MLSYS-pipe-fisher

Last synced: 9 days ago
JSON representation

• Host: GitHub
• URL: https://github.com/MachineLearningSystem/23MLSYS-pipe-fisher
• Owner: MachineLearningSystem
• Fork: true (kazukiosawa/pipe-fisher)
• Created: 2023-06-07T05:42:57.000Z (over 1 year ago)
• Default Branch: main
• Last Pushed: 2023-04-29T15:20:48.000Z (over 1 year ago)
• Last Synced: 2024-08-02T19:33:16.768Z (4 months ago)
• Size: 464 KB
• Stars: 0
• Watchers: 0
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

• awesome-AI-system - PipeFisher: Efficient Training of Large Language Models Using Pipelining and Fisher Information Matrices MLSYS'23

README

        
# PipeFisher

The implementation of pipeline-parallel training with K-FAC optimizer (PipeFisher) in PyTorch used in [PipeFisher: Efficient Training of Large Language Models Using Pipelining and Fisher Information Matrices](https://arxiv.org/abs/2211.14133) (to appear at MLSys 2023).

## Setup

### Data preparation

https://github.com/microsoft/AzureML-BERT/blob/master/docs/dataprep.md

Please store `wikipedia.segmented.nltk.txt` file under the `bert_data/` directory.

### Installation

```

pip install -r requirements.txt

pip install asdfghjkl/

```

For training, we use `apex.optimizers.FusedLAMB` of [NVIDIA's Apex library](https://github.com/NVIDIA/apex). Please follow the [instruction](https://github.com/NVIDIA/apex#installation) for installing `apex`. 

For profiling, we use [NVIDIA Nsight Systems](https://developer.nvidia.com/nsight-systems). Please make sure you can execute `nsys` command.

Our scripts are intended to run through the SLURM workload manager on a GPU cluster with 1 GPU per node.

## Training

Phase 1 pretraining of BERT-Base on the English Wikipedia by NVLAMB on 32 GPUs

```

sbatch scripts/train.sh

```

Phase 1 pretraining of BERT-Base on the English Wikipedia by K-FAC on 32 GPUs

```

sbatch scripts/train_kfac.sh

```



## Profiling

### Step 0. Profiling **Chimera** with 8 stages for BERT-Large on 8 GPUs 

```

sbatch scripts/prof_steps.sh

```

```

sh scripts/plot_cuda_timeline.sh

```

output: `bert_prof/bert-large_chimera_8stages_8gpus_microbs32_acc1.pdf`

### Step 1. Profiling **Chimera with K-FAC** with 8 stages for BERT-Large on 8 GPUs 

```

sbatch scripts/prof_kfac_steps.sh

```

```

sh scripts/plot_cuda_timeline_kfac.sh

```

output: `bert_prof/bert-large_chimera_8stages_8gpus_microbs32_acc1_kfac.pdf`

### Step 2. Automatic work assignments

```

sh scripts/auto_schedule.sh

```

output: `bert-large_chimera_8stages_8gpus_microbs32_acc1_kfac_schedule.pickle`

### Step 3. Profiling **Chimera with PipeFisher** with 8 stages for BERT-Large on 8 GPUs 

```

sbatch scripts/prof_pipefisher_steps.sh

```

```

sh scripts/plot_cuda_timeline_pipefisher.sh

```

output: `bert_prof/bert-large_chimera_8stages_8gpus_microbs32_acc1_pipefisher.pdf`



By changing the settings of each script, you can run training/profiling on other BERT models, pipeline methods, number of pipeline stages, number of GPUs, etc.