Qeravio
Canonical AI event

CUDA sparse-fa for dsv4 prefill re-enabled, query loop unrolled.

The CUDA update for dsv4 prefill reduces scan time from 46 to 17 microseconds at 49k columns.

24 Sept 20261 verified claims1 sources1 observations
What happened

The official source reports this update: b11140: CUDA: enable sparse-fa for dsv4 prefill (again). CUDA: enable sparse-fa for dsv4 prefill (again) ( #29298 ) CUDA: enable sparse-fa for dsv4 prefill (again) CUDA: unroll the query loop of the sparse mask scan The query loop of flash_attn_mask_to_sparse_indices has a runtime trip count, which keeps the unrolled scan over the values of a lane from issuing its loads together. Template the kernel on ncols1 so the loop is bounded at compile time: batch one decodes compile to straight line code and the scan drops from 46 to 17 us at 49k columns on sparse decode shapes.

Why it matters

This official update documents a development concerning b11140: CUDA: enable sparse-fa for dsv4 prefill (again). Its practical significance depends on the scope and evidence stated by the source.

What to watch next

Read the official source update and verify its stated scope, evidence and timing before acting on it.

Connected knowledge

Entities affected by this event

Continue this topic
Evidence trail

Sources behind the event