pub fn prefill_attention(
gpu: &dyn GpuBackend,
kernel: KernelHandle,
q: DevicePtr,
k: DevicePtr,
v: DevicePtr,
output: DevicePtr,
seq_len: u32,
batch: u32,
num_q_heads: u32,
num_kv_heads: u32,
head_dim: u32,
inv_sqrt_d: f32,
causal: bool,
sliding_window: u32,
stream: u64,
) -> Result<()>Expand description
2026-09-25: Flash-attention prefill on contiguous BF16 tensors: Q and O
[batch, seq_len, num_q_heads, head_dim], K and V [batch, seq_len, num_kv_heads, head_dim]. sliding_window == 0 disables the window;
otherwise keys with q - k >= sliding_window are masked.