# Runway — Why Distributed Training Is Hard: DTensor, Correctness and the Costs of Abstraction

- Company: Runway (runway.com)
- Announced: 2026-05-18
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://runway.com/news/engineering/dtensor-distributed-training
- Record: https://forck.live/items/7880-why-distributed-training-is-hard-dtensor-correctness-and-the-costs-of
- Subject: Gen / Aleph

Runway published a technical post explaining the challenges of distributed training with DTensor, including gradient correctness issues that arise when sharding tensors across process groups.

## Evidence

Verbatim from https://runway.com/news/engineering/dtensor-distributed-training:

> DTensor makes distributed training correct by attaching placement metadata to every tensor. At scale it can also introduce costs that quietly erode throughput unless you design around them.

---

Record: https://forck.live/items/7880-why-distributed-training-is-hard-dtensor-correctness-and-the-costs-of
Catalogue: https://forck.live/llms.txt
Current issue: https://forck.live/feed.md