Method process_kv_inner

cake-core/src/models/common/cache.rs:184–210

(
        &mut self,
        block_idx: usize,
        mut k: Tensor,
        mut v: Tensor,
        limit: usize,
    )

    }

    fn process_kv_inner(
        &mut self,
        block_idx: usize,
        mut k: Tensor,
        mut v: Tensor,
        limit: usize,
    ) -> Result<(Tensor, Tensor)> {
        if self.use_kv_cache {
            if let Some((cache_k, cache_v)) = &self.kvs[block_idx] {
                // tensor shape is (batch, num_heads, seq_len, head_dim)
                // cat() already returns a contiguous tensor — no need for .contiguous()
                k = Tensor::cat(&[cache_k, &k], 2)?;
                v = Tensor::cat(&[cache_v, &v], 2)?;

                let k_seq_len = k.dims()[2];
                if k_seq_len > limit {
                    k = k.narrow(2, k_seq_len - limit, limit)?.contiguous()?;
                }
                let v_seq_len = v.dims()[2];
                if v_seq_len > limit {
                    v = v.narrow(2, v_seq_len - limit, limit)?.contiguous()?;
                }
            }
            self.kvs[block_idx] = Some((k.clone(), v.clone()))
        }
        Ok((k, v))
    }

    /// Directly set the KV cache for a layer (used for voice prompt injection).
    pub fn set_kv(&mut self, block_idx: usize, k: Tensor, v: Tensor) {