pub fn mamba2_ssd_cumsum(
gpu: &dyn GpuBackend,
kernel: KernelHandle,
dt_raw: DevicePtr,
a_log: DevicePtr,
dt_bias: DevicePtr,
dt_out: DevicePtr,
da_cs: DevicePtr,
seq_len: u32,
num_heads: u32,
nchunks: u32,
batch_size: u32,
dt_stride: u32,
dt_min: f32,
dt_max: f32,
stream: u64,
) -> Result<()>Expand description
K1: per-chunk dt (softplus+clamp) and inclusive cumsum of the log-decay.