kutacc_af2_triangle_multiplication_calc_proj
Calculate the intermediate values generated during triangle_multiplication, that is, left and right projections.
Interface Definition
void kutacc_af2_triangle_multiplication_calc_proj(kutacc_af2_tm_act_inputs_t *tm_acts_ptr, kutacc_tensor_h mask, kutacc_af2_tm_proj_weights_t *tm_weights_ptr, bool input_prepack);
Parameters
Table 1 Input parameter definition
Parameter |
Type |
Description |
Input/Output |
|---|---|---|---|
tm_acts_ptr |
kutacc_af2_tm_act_inputs_t * |
Pointer of the kutacc_af2_tm_act_inputs_t type. For details about the data structure, see Table 2. |
Input |
mask |
kutacc_tensor_h |
Mask tensor |
Input |
tm_weights_ptr |
kutacc_af2_tm_proj_weights_t * |
Pointer of the kutacc_af2_tm_proj_weights_t type. For details about the data structure, see Table 3. |
Input |
input_prepack |
bool |
Whether to perform the prepack operation on the input tensor |
Input |
Table 2 kutacc_af2_tm_act_inputs_t structure definition
Parameter |
Type |
Description |
Input/Output |
|---|---|---|---|
n_res |
int64_t |
Number of residues |
Input |
n_res_gather |
int64_t |
Number of residues |
Input |
proj_act |
kutacc_tensor_h |
Left and right projections of act |
Output |
input_act |
kutacc_tensor_h |
Value generated after the input act is normalized by LayerNorm |
Input |
proj_act_gate |
kutacc_tensor_h |
Temporary gating tensor |
Input |
Table 3 kutacc_af2_tm_proj_weights_t structure definition
Parameter |
Type |
Description |
Input/Output |
|---|---|---|---|
c_o |
int64_t |
Input feature dimension |
Input |
c_i |
int64_t |
Output feature dimension |
Input |
proj_w |
kutacc_tensor_h |
Weight required for projection generation |
Input |
proj_b |
kutacc_tensor_h |
Bias required for projection generation |
Input |
gate_w |
kutacc_tensor_h |
Gating weight |
Input |
gate_b |
kutacc_tensor_h |
Gating bias |
Input |
Constraints on integer parameters of triangle_multiplication:
n_res, n_res_gather, c_i, c_o > 0
n_res * n_res_gather < INT64_MAX
c_i = c_o
In single-process mode, n_res must be equal to n_res_gather. In multi-process mode, this condition does not need to be met.