Struct DflashTapSink 

pub struct DflashTapSink {
    pub layer_ids: Vec<usize>,
    pub buf: CudaSlice<f32>,
    pub hidden: usize,
    pub t: usize,
    pub base: usize,
}

See Cache::dflash_taps. Armed per forward by the dflash round (t = that forward’s row count); the trunk writes tap slot s of row r at buf[rn_tapshidden + s*hidden ..].

Fields

layer_ids: Vec<usize>

buf: CudaSlice<f32>

hidden: usize

t: usize

base: usize

Row offset for writers that walk the buffer in windows (the qwen chunked prime): tap rows land at [base..base+t_chunk). Whole-buffer writers leave it 0.