Model techniques map
Techniquesinference & servingserving parallelism

specific method · filed under inference & serving

Attention data parallelism

Applies data parallelism to attention, with the evidence noting it avoids copying KV across ranks.

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

sources
3
models
3
labs adopt it
3
strongest
used

How sources treat it

One count per evidence span, weakest treatment to strongest.

optional 2used 2

Documented in

Further reading

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

Evidence

4 spans quoted from the sources, strongest treatment first.

DP-attention is primarily introduced to prevent copying KV across different ranks

usedmodel architecturein GLM-5Z.ai

We utilized a hybrid approach combining Attention Data Parallelism (DP) and MoE Expert Parallelism (EP)

usedinference servingin GLM-5Z.ai

enable_attention_dp: true

optionalinference servingin Nemotron 3 UltraNVIDIA

--enable-dp-attention

optionalinference servingin MiMo-V2.5Xiaomi

Filed alongside

Other methods under inference & serving :: serving parallelism.