Open Models Almanac
Methodsinference & servingserving parallelism

specific method · filed under inference & serving

Attention data parallelism

DP-attention is a serving parallelism scheme that runs attention under data parallelism across ranks.1 Its purpose is to avoid copying KV cache between ranks while provisioning distributed KV-cache capacity.2

How this definition was written

Summarised by glm-5.3-flash from 1 source; every sentence is checked against the quote it cites.

  1. multi-node inference with DP-attention for MLA — GLM-5: from Vibe Coding to Agentic Engineering
  2. DP-attention is primarily introduced to prevent copying KV across different ranks — GLM-5: from Vibe Coding to Agentic Engineering

Also called Attention Data Parallelism (DP), Data Parallel attention, --enable-dp-attention, DP-attention.

sources
5
models
4
labs adopt it
3
strongest
used

Documented in

Evidence

6 quotes from 5 sources, one card per model, each opening with its strongest.

How sources treat it
mentioned 2optional 2used 2

Step-5-Preview

StepFundiscussed2 sources

For more optimized configurations, including MegaMoE, attention data parallelism, and HiCache, see the [SGLang Cookbook for Step 5 Preview]

mentionedinference servingin Step 5 PreviewStepFun
1 more from this model

For more optimized configurations, including MegaMoE, attention data parallelism and HiCache, see the SGLang Cookbook for Step 5 Preview.

mentionedinference servingin Step 5 PreviewStepFun

Further reading

Picked by hand, not extracted: where to read more, not evidence for anything on this page.

↑