flashinfer.gdp_prefill

Gated DeltaProduct prefill. chunk_gated_delta_product is the chunked GDP scan: GDN with num_householder beta-gated Householder updates per token instead of one, on an expanded k/v/beta sub-token timeline.

backend="auto" and backend="cudnn" run cuDNN’s fused SM100 linear-attention engine (flashinfer.cudnn.cudnn_chunk_gated_delta_product()); it needs an SM100-family device and cudnn-frontend 1.28+ with the cutedsl extra, and declines anything else it cannot serve. backend="cake" runs the Cake GDP prefill kernel (K=128, V=64, one to four Householder steps per token; SM90) with the same argument contract.

chunk_gated_delta_product(q, k, v[, g, ...])

Chunked Gated DeltaProduct (GDP) attention for prefill.

Decode and MTP verification (Cake, SM90 / SM100 / SM103)

Gated DeltaProduct decode and multi-token (MTP) verification: K=128, V=64, one to four Householder steps per token, over an FP32 pretranspose state pool with indexed read / write / per-token snapshot slots. One fused kernel per call; the docstring states the input contracts.

cake_gated_delta_product_mtp(q, k, v, ...[, ...])

Gated DeltaProduct decode / MTP verification over an FP32 state pool.