| 26195 | layer->hc_ffn_base, |
| 26196 | cpu_after_attn_hc, cpu_ffn_cur, ffn_post, ffn_comb); |
| 26197 | rms_norm_weight(cpu_ffn_norm, cpu_ffn_cur, tensor_data(model, layer->ffn_norm), DS4_N_EMBD, DS4_RMS_EPS); |
| 26198 | quantize_q8_0_activation(cpu_ffn_norm, shared_xq, shared_xscale, shared_in_dim); |
| 26199 | matvec_q8_0_pair_prequant(cpu_shared_gate, |
| 26200 | cpu_shared_up, |
| 26201 | model, |
| 26202 | layer->ffn_gate_shexp, |
| 26203 | layer->ffn_up_shexp, |
| 26204 | shared_xq, |
| 26205 | shared_xscale); |
| 26206 | swiglu(cpu_shared_mid, cpu_shared_gate, cpu_shared_up, shared_dim, DS4_SWIGLU_CLAMP_EXP); |
| 26207 | matvec_q8_0(cpu_shared, model, layer->ffn_down_shexp, cpu_shared_mid); |
| 26208 | layer_routed_moe_one_prealloc(cpu_routed, |
| 26209 | model, |
| 26210 | layer, |
| 26211 | cpu_ffn_norm, |
| 26212 | il, |
| 26213 | token, |
| 26214 | DS4_SWIGLU_CLAMP_EXP, |
| 26215 | routed_mid_all, |
| 26216 | routed_xq, |
| 26217 | routed_midq, |
| 26218 | routed_q8_xq, |
| 26219 | routed_q8_xscale, |
| 26220 | routed_q8_midq, |
| 26221 | routed_q8_midscale); |
| 26222 | if (layer->ffn_gate_tid2eid) { |
no test coverage detected