pub fn qsa_block_pool(
gpu: &dyn GpuBackend,
kernel: KernelHandle,
raw_keys: DevicePtr,
k_norm_w: DevicePtr,
block_keys: DevicePtr,
first_block: u32,
n_new: u32,
ratio: u32,
hd: u32,
rot: u32,
theta: f32,
eps: f32,
stream: u64,
) -> Result<()>Expand description
2026-09-25: Pool n_new newly complete blocks starting at first_block: the
mean of ratio raw keys, then RMSNorm scaled by 1 + w, then rope at the
block’s first token position. Returns without launching when n_new == 0.