Deepseek V4

[DiegoCao]

I was wondering if this DSA V4 Flash, can this be applied to the Deepseek V4 as well?

[fooSynaptic]

I was wondering if this DSA V4 Flash, can this be applied to the Deepseek V4 as well?

The idea may extend to V4 because its CSA path still uses a DSA-style indexer and top-k selection. However, V4 indexes 4:1 compressed entries and uses a heterogeneous attention stack, so this requires a new implementation and validation.

[DiegoCao]

Yes, I was wondering if this still has benefit towards DSV4, due to the FP4 indexer. For noit using fp4 indexer, will it be beneficial?

Meanwhile, have we validated DSA Indexer Cache performance in the high throughput case?