Figure 5 · DCP8 attention module
The exact one-round merge makes DCP8 sparse attention 1.3–1.5× faster
Time of one DSA attention module at DCP8 on 8× GB300, 5 query tokens per request, in µs; lower is better. Kimi K3's dense MLA is the reference.
Figure 5 · DCP8 attention module
Time of one DSA attention module at DCP8 on 8× GB300, 5 query tokens per request, in µs; lower is better. Kimi K3's dense MLA is the reference.