| 13096 | layer_kv_projection_normed_one(model, layer, attn_norm, kv); |
| 13097 | |
| 13098 | rope_tail_layer_inplace(q, DS4_N_HEAD, DS4_N_HEAD_DIM, DS4_N_ROT, pos, il, false); |
| 13099 | rope_tail_layer_inplace(kv, DS4_N_HEAD_KV, DS4_N_HEAD_DIM, DS4_N_ROT, pos, il, false); |
| 13100 | dsv4_fp8_kv_quantize_row_inplace_cpu(kv, DS4_N_HEAD_DIM, DS4_N_ROT); |
| 13101 | |
| 13102 | kv_cache_push_raw(cache, kv); |
| 13103 | |
| 13104 | const uint32_t ratio = cache->compress_ratio; |
| 13105 | if (ratio != 0) { |
| 13106 | float *comp = xmalloc((size_t)DS4_N_HEAD_DIM * sizeof(comp[0])); |
| 13107 | if (compressor_decode_one(comp, model, |
| 13108 | layer->attn_compressor_kv, |
| 13109 | layer->attn_compressor_gate, |
| 13110 | layer->attn_compressor_ape, |
| 13111 | layer->attn_compressor_norm, |
| 13112 | attn_norm, |
| 13113 | cache->attn_state_kv, |
no test coverage detected