metaflow-deepspeed 0.0.7

An EXPERIMENTAL Deepspeed decorator for Metaflow

Introduction

Deepspeed is a highly scalable framework from Microsoft for distributed training and model serving. The Metaflow @deepspeed decorator helps you run these workflows inside of Metaflow tasks.

Features

• Automatic SSH configuration: Multi-node Deepspeed jobs are built around OpenMPI or Horovod. Like Metaflow's @mpi decorator, the @deepspeed decorator automatically configures SSH requirements between nodes, so you can focus on research code.
• Seamless Python interface: Metaflow's @deepspeed exposes a method current.deepspeed.run to make it easy to run Deepspeed commands on your transient MPI cluster, in the same way you'd launch Deepspeed from the terminal independent of Metaflow. A major design goal is to get the orchestration and other benefits of Metaflow, without requiring modification to research code.

Installation

Install this experimental module:

pip install metaflow-deepspeed

Getting Started

After installing the module, you can import the deepspeed decorator and use it in your Metaflow steps. This exposes the current.deepspeed.run method, which you can map your terminal commands for running Deepspeed.

from metaflow import FlowSpec, step, deepspeed, current, batch, environment

class HelloDeepspeed(FlowSpec):

    @step
    def start(self):
        self.next(self.train, num_parallel=2)

    @environment(vars={"NCCL_SOCKET_IFNAME": "eth0"})
    @batch(gpu=8, cpu=64, memory=256000)
    @deepspeed
    @step
    def train(self):
        current.deepspeed.run(
            entrypoint="my_torch_dist_script.py"
        )
        self.next(self.join)

    @step
    def join(self, inputs):
        self.next(self.end)

    @step
    def end(self):
        pass
        
if __name__ == "__main__":
    HelloDeepspeed()

Examples

Directory	MPI program description
CPU Check	The easiest way to check your Deepspeed infrastructure on CPUs.
Hello Deepspeed	The easiest way to check your Deepspeed infrastructure on GPUs.
BERT	Train your BERT model using Deepspeed!
Dolly	A multi-node implementation of Databricks' Dolly.

Cloud-specific use cases

Directory	MPI program description
Automatically upload a directory on AWS	Push a checkpoint of any directory to S3 after the Deepspeed process completes.
Automatically upload a directory on Azure	Push a checkpoint of any directory to Azure Blob storage after the Deepspeed process completes.
Use Metaflow S3 client from the Deepspeed process	Upload arbitrary bytes to S3 storage from the Deepspeed process.
Use Metaflow Azure Blob client from the Deepspeed process	Upload arbitrary bytes to Azure Blob storage from the Deepspeed process.
Use a Metaflow Huggingface checkpoint on S3	Push a checkpoint to S3 at the end of each epoch using a customizable Huggingface callback. See the implementation here to build your own.
Use a Metaflow Huggingface checkpoint on Azure	Push a checkpoint to Azure Blob storage at the end of each epoch using a customizable Huggingface callback. See the implementation here to build your own.

License

metaflow-deepspeed is distributed under the Apache License.