flashinfer.gdn2_prefill

Gated Delta-Rule 2 prefill. chunk_gated_delta_rule2 is the chunked GDN-2 scan: GDN’s per-head scalar forget and erase gates become per key channel, and a third per-value-channel write gate scales the incoming value.

FlashInfer carries no GDN-2 kernel of its own, so this runs on cuDNN’s fused SM100 linear-attention engine (flashinfer.cudnn.cudnn_chunk_gated_delta_rule2()). It needs an SM100-family device and cudnn-frontend 1.28+ with the cutedsl extra; the engine declines anything else it cannot serve.

chunk_gated_delta_rule2(q, k, v[, g, beta, ...])

Chunked Gated Delta Rule 2 (GDN-2) attention for prefill.