[object Object]

[object Object][object Object]undefined[object Object]

• API function: Calculates the variable step-size Recurrent Gated Delta Rule (RGDR).

• Formula:

  RGDR is an operator designed for recurrent neural networks (RNNs) and serves as a key component in linear attention mechanisms. At each time step $t$, the network processes the current inputs query $q_t$, key $k_t$, and value $v_t$ along with the previous hidden state $S_{t-1}$ to produce the attention output $o_t$ and new hidden state $S_t$. The gating mechanism controls what proportion of new information enters the hidden state and what proportion of existing information is discarded.

  $$S_t := S_{t-1}(\alpha_t Diag(\alpha_{kt})(I - \beta_t k_t k_t^T)) + \beta_t v_t k_t^T = \alpha_t Diag(\alpha_{kt})S_{t-1} + \beta_t (v_t - \alpha_t Diag(\alpha_{kt})S_{t-1}k_t)k_t^T$$ $$o := \frac{S_t q_t}{\sqrt{d_k}}$$

  In the preceding information, $S_{t-1},S_t \in R^{d_v \times d_k}$, $q_t, k_t \in R^{d_k}$, $v_t \in R^{d_v}$, $\alpha_t \in R$, $\alpha_k \in R^{d_k}$, $\beta_t \in R$, $o \in R^{d_v}$

[object Object]

Each operator has calls. First, [object Object] is called to obtain the workspace size required for computation and the executor that contains the operator computation process. Then, [object Object] is called to perform computation.

[object Object]

• Parameters

  [object Object]

  $B$ indicates the batch size. Let $L_i$ be the length of the ith sequence, then $T=\sum_i^B L_i$ represents the cumulative sequence length. $N_k$ indicates the number of key heads, $N_v$ indicates the number of value heads, $D_k$ indicates the dimension of the key vector, and $D_v$ indicates the dimension of the value vector.

• Returns

  [object Object]: status code. For details, see .

  The first-phase API implements input parameter verification. The following errors may be thrown:

  [object Object]

[object Object]

• Parameters

  [object Object]

• Returns

  [object Object]: status code. For details, see .

[object Object]

• Deterministic computing:
  • [object Object] defaults to a deterministic implementation.
• The input shape size must meet the following constraints: $L_i \le 8$, $N_k \le 256$, $N_v \le 256$, $D_k \le 256$, $D_v \le 256$.

[object Object]

The following example is for reference only. For details, see .