pub fn read_k_lr_slot(
ctx: &CudaCtx,
pred: &Predictor,
layer: usize,
block: usize,
) -> Result<Vec<bf16>>Expand description
2026-09-25: Copy the A_g slot of (layer, block) to the host, as BF16
[num_kv_heads, block_size, r], and synchronise the stream.