---
title: "distributed training | Skills We Assign For | Azendo"
description: "Distributed training — data versus model parallelism, communication cost, plus the Azendo roles assigned for it."
url: "https://azendo.co/skills/distributed-training/"
---

[Skills](https://azendo.co/skills/) MLOps and model delivery 

# distributed training.

Distributed training spreads model training across multiple GPUs or machines, through data parallelism where each device holds the model and processes different batches, or model parallelism where the model itself is split.

## Where distributed training fits on a long engagement.

Scaling is not linear and the gap is communication. Gradients must be synchronised across devices every step, and that traffic grows with device count, so eight GPUs do not deliver eight times the throughput. Interconnect bandwidth frequently becomes the limit rather than compute.

Batch size interacts with learning rate in ways that surprise people. Distributing across devices multiplies the effective batch, and keeping the original learning rate produces worse convergence. Scaling rules and warmup exist for this and are routinely omitted.

## What an assigned team does with distributed training.

Distributed training is expensive in both compute and engineering time. Before committing, it is worth establishing whether mixed precision, gradient accumulation or a smaller model reach an acceptable result on one device.

Making that assessment honestly rather than reaching for the cluster is the judgement scoped under [ai engineering services](https://azendo.co/services/ai-engineering/).

## What we use distributed training for.

* Training that will not fit one device Model or data parallelism where a single GPU genuinely cannot hold the work.
* Learning rate scaled with batch size Warmup and scaling rules applied, because the effective batch has changed.
* Single-device options exhausted first Mixed precision and accumulation tried before accepting the cluster cost.

## How distributed training capacity is assigned.

Training infrastructure is assigned inside AI capacity, with single-device alternatives evaluated before distribution is committed to.

## Roles we assign distributed training for

* [ML Platform Engineer MLOps engineering](https://azendo.co/services/mlops-engineering/ml-platform-engineer/)

## Service lines it sits in

* [MLOps engineering](https://azendo.co/services/mlops-engineering/)

Capacity is agreed as a committed monthly capacity across a discipline, not per skill.

## Related in mlops and model delivery

* [MLOps handoff — skill we assign for](https://azendo.co/skills/mlops-handoff/)
* [ONNX — skill we assign for](https://azendo.co/skills/onnx/)
* [TensorRT — skill we assign for](https://azendo.co/skills/tensorrt/)
* [Model registry — skill we assign for](https://azendo.co/skills/model-registry/)
* [Weights & Biases — skill we assign for](https://azendo.co/skills/weights-and-biases/)
* [Comet — skill we assign for](https://azendo.co/skills/comet/)
* [Neptune.ai — skill we assign for](https://azendo.co/skills/neptune-ai/)
* [DVC — skill we assign for](https://azendo.co/skills/dvc/)

## Tell us what your roadmap needs distributed training for.

A service delivery manager replies with the disciplines we would assign, the monthly capacity and what the first month looks like.

[All skills we assign for](https://azendo.co/skills/)