npm - @fugood/llama.node - Versions diffs - 1.1.8 → 1.1.10 - Mend

@fugood/llama.node 1.1.8 → 1.1.10

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.

Files changed (47) hide show

package/lib/binding.ts +9 -0
package/lib/index.js +9 -2
package/lib/index.ts +57 -30
package/lib/version.js +2 -2
package/lib/version.ts +2 -2
package/package.json +14 -14
package/scripts/llama.cpp.patch +15 -5
package/src/LlamaCompletionWorker.cpp +12 -3
package/src/LlamaCompletionWorker.h +3 -1
package/src/LlamaContext.cpp +14 -1
package/src/llama.cpp/common/arg.cpp +6 -4
package/src/llama.cpp/common/chat.cpp +34 -3
package/src/llama.cpp/common/common.cpp +0 -15
package/src/llama.cpp/common/common.h +1 -2
package/src/llama.cpp/ggml/CMakeLists.txt +0 -1
package/src/llama.cpp/ggml/include/ggml.h +25 -0
package/src/llama.cpp/ggml/src/ggml-cpu/arch/s390/quants.c +316 -0
package/src/llama.cpp/ggml/src/ggml-cpu/arch-fallback.h +0 -2
package/src/llama.cpp/ggml/src/ggml-cpu/ggml-cpu-impl.h +8 -0
package/src/llama.cpp/ggml/src/ggml-cpu/ggml-cpu.c +6 -0
package/src/llama.cpp/ggml/src/ggml-cpu/ops.cpp +142 -0
package/src/llama.cpp/ggml/src/ggml-cpu/ops.h +1 -0
package/src/llama.cpp/include/llama.h +1 -110
package/src/llama.cpp/src/CMakeLists.txt +2 -2
package/src/llama.cpp/src/llama-arch.cpp +19 -0
package/src/llama.cpp/src/llama-arch.h +1 -0
package/src/llama.cpp/src/llama-chat.cpp +13 -2
package/src/llama.cpp/src/llama-chat.h +1 -0
package/src/llama.cpp/src/llama-context.cpp +5 -192
package/src/llama.cpp/src/llama-context.h +2 -7
package/src/llama.cpp/src/llama-cparams.h +0 -1
package/src/llama.cpp/src/llama-graph.cpp +35 -57
package/src/llama.cpp/src/llama-graph.h +36 -46
package/src/llama.cpp/src/llama-hparams.cpp +25 -0
package/src/llama.cpp/src/llama-hparams.h +6 -0
package/src/llama.cpp/src/{llama-kv-cache-unified-iswa.cpp → llama-kv-cache-iswa.cpp} +69 -52
package/src/llama.cpp/src/{llama-kv-cache-unified-iswa.h → llama-kv-cache-iswa.h} +28 -26
package/src/llama.cpp/src/{llama-kv-cache-unified.cpp → llama-kv-cache.cpp} +123 -474
package/src/llama.cpp/src/{llama-kv-cache-unified.h → llama-kv-cache.h} +34 -59
package/src/llama.cpp/src/llama-kv-cells.h +21 -21
package/src/llama.cpp/src/llama-memory-hybrid.cpp +34 -33
package/src/llama.cpp/src/llama-memory-hybrid.h +24 -28
package/src/llama.cpp/src/llama-memory-recurrent.cpp +7 -7
package/src/llama.cpp/src/llama-memory-recurrent.h +8 -12
package/src/llama.cpp/src/llama-memory.h +11 -8
package/src/llama.cpp/src/llama-model.cpp +396 -187
package/src/llama.cpp/src/llama-model.h +1 -0

package/src/llama.cpp/src/llama-model.cpp CHANGED Viewed

@@ -6,8 +6,8 @@
 #include "llama-cparams.h"
 #include "llama-model-loader.h"
-#include "llama-kv-cache-unified.h"
-#include "llama-kv-cache-unified-iswa.h"
+#include "llama-kv-cache.h"
+#include "llama-kv-cache-iswa.h"
 #include "llama-memory-hybrid.h"
 #include "llama-memory-recurrent.h"
@@ -83,6 +83,7 @@ const char * llm_type_name(llm_type type) {
         case LLM_TYPE_32B:           return "32B";
         case LLM_TYPE_34B:           return "34B";
         case LLM_TYPE_35B:           return "35B";
+        case LLM_TYPE_36B:           return "36B";
         case LLM_TYPE_40B:           return "40B";
         case LLM_TYPE_65B:           return "65B";
         case LLM_TYPE_70B:           return "70B";
@@ -1114,6 +1115,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                 hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;
                 hparams.set_swa_pattern(5);
+                hparams.n_layer_kv_from_start     = 20;
                 hparams.rope_freq_base_train_swa  = 10000.0f;
                 hparams.rope_freq_scale_train_swa = 1.0f;
                 hparams.f_attention_scale         = 1.0f;
@@ -1288,6 +1290,14 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                     default: type = LLM_TYPE_UNKNOWN;
                 }
             } break;
+        case LLM_ARCH_SEED_OSS:
+            {
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+                switch (hparams.n_layer) {
+                    case 64: type = LLM_TYPE_36B; break;
+                    default: type = LLM_TYPE_UNKNOWN;
+                }
+            } break;
         case LLM_ARCH_OLMOE:
             {
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
@@ -1465,12 +1475,15 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                 // Expert gating function (GLM-4.5 uses sigmoid)
                 ml.get_key(LLM_KV_EXPERT_GATING_FUNC,          hparams.expert_gating_func, false);
                 if (hparams.expert_gating_func == LLAMA_EXPERT_GATING_FUNC_TYPE_NONE) {
-                    hparams.expert_gating_func = LLAMA_EXPERT_GATING_FUNC_TYPE_SIGMOID;
+                    hparams.expert_gating_func =  LLAMA_EXPERT_GATING_FUNC_TYPE_SIGMOID;
                 }
                 // NextN/MTP parameters
                 ml.get_key(LLM_KV_NEXTN_PREDICT_LAYERS,        hparams.nextn_predict_layers, false);
+                // TODO: when MTP is implemented, this should probably be updated if needed
+                hparams.n_layer_kv_from_start = hparams.n_layer - hparams.nextn_predict_layers;
                 switch (hparams.n_layer) {
                     case 47: type = LLM_TYPE_106B_A12B; break; // GLM-4.5-Air (46 layers + 1 NextN layer)
                     case 93: type = LLM_TYPE_355B_A32B; break; // GLM-4.5 (92 layers + 1 NextN layer)
@@ -3967,6 +3980,43 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         layer.ffn_post_norm = create_tensor(tn(LLM_TENSOR_FFN_POST_NORM, "weight", i), {n_embd}, 0);
                     }
                 } break;
+            case LLM_ARCH_SEED_OSS:
+                {
+                    const uint32_t head_dim             = hparams.n_embd_head_k;
+                    const int64_t n_qo_dim              = n_head * head_dim;
+                    const int64_t n_kv_dim              = n_head_kv * head_dim;
+                    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+                    // output
+                    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+                    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+                    // if output is NULL, init from the input tok embed
+                    if (output == NULL) {
+                        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+                    }
+                    for (int i = 0; i < n_layer; ++i) {
+                        auto & layer = layers[i];
+                        layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_qo_dim}, 0);
+                        layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_kv_dim}, 0);
+                        layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_kv_dim}, 0);
+                        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_qo_dim, n_embd}, 0);
+                        layer.bq = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "bias", i), {n_qo_dim},   TENSOR_NOT_REQUIRED);
+                        layer.bk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "bias", i), {n_kv_dim},   TENSOR_NOT_REQUIRED);
+                        layer.bv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "bias", i), {n_kv_dim},   TENSOR_NOT_REQUIRED);
+                        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+                        layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), {n_embd}, 0);
+                        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
+                        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
+                        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
+                    }
+                } break;
             case LLM_ARCH_OLMOE:
                 {
                     tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
@@ -5474,8 +5524,13 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                 } break;
             case LLM_ARCH_LFM2:
                 {
-                    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+                    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD,      "weight"), {n_embd, n_vocab}, 0);
                     tok_norm = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "weight"), {n_embd}, 0);
+                    output   = create_tensor(tn(LLM_TENSOR_OUTPUT,          "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+                    if (output == NULL) {
+                        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+                    }
                     for (int i = 0; i < n_layer; ++i) {
                         auto & layer = layers[i];
@@ -5986,7 +6041,7 @@ struct llm_build_llama : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
         const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
@@ -6050,7 +6105,7 @@ struct llm_build_llama : public llm_graph_context {
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, kq_scale, il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
                 cb(cur, "attn_out", il);
             }
@@ -6146,7 +6201,7 @@ struct llm_build_llama_iswa : public llm_graph_context {
         ggml_tensor * inp_attn_scale = nullptr;
         inp_attn_scale = build_inp_attn_scale();
-        auto * inp_attn = build_attn_inp_kv_unified_iswa();
+        auto * inp_attn = build_attn_inp_kv_iswa();
         const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
@@ -6224,7 +6279,7 @@ struct llm_build_llama_iswa : public llm_graph_context {
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, kq_scale, il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
                 cb(cur, "attn_out", il);
             }
@@ -6325,7 +6380,7 @@ struct llm_build_deci : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
         const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
@@ -6401,7 +6456,7 @@ struct llm_build_deci : public llm_graph_context {
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, kq_scale, il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
             }
             if (il == n_layer - 1 && inp_out_ids) {
@@ -6481,7 +6536,7 @@ struct llm_build_baichuan : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = model.type == LLM_TYPE_7B ? build_inp_pos() : nullptr;
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
         ggml_tensor * inp_out_ids = build_inp_out_ids();
@@ -6533,7 +6588,7 @@ struct llm_build_baichuan : public llm_graph_context {
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
             if (il == n_layer - 1 && inp_out_ids) {
@@ -6603,7 +6658,7 @@ struct llm_build_xverse : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
         ggml_tensor * inp_out_ids = build_inp_out_ids();
@@ -6648,7 +6703,7 @@ struct llm_build_xverse : public llm_graph_context {
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
             if (il == n_layer - 1 && inp_out_ids) {
@@ -6717,7 +6772,7 @@ struct llm_build_falcon : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
         ggml_tensor * inp_out_ids = build_inp_out_ids();
@@ -6771,7 +6826,7 @@ struct llm_build_falcon : public llm_graph_context {
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
             if (il == n_layer - 1 && inp_out_ids) {
@@ -6841,7 +6896,7 @@ struct llm_build_grok : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
         ggml_tensor * inp_out_ids = build_inp_out_ids();
@@ -6901,7 +6956,7 @@ struct llm_build_grok : public llm_graph_context {
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f, il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f, il);
             }
             if (il == n_layer - 1 && inp_out_ids) {
@@ -7001,7 +7056,7 @@ struct llm_build_dbrx : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
         ggml_tensor * inp_out_ids = build_inp_out_ids();
@@ -7050,7 +7105,7 @@ struct llm_build_dbrx : public llm_graph_context {
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
             if (il == n_layer - 1 && inp_out_ids) {
@@ -7125,7 +7180,7 @@ struct llm_build_starcoder : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
         ggml_tensor * pos = ggml_get_rows(ctx0, model.pos_embd, inp_pos);
         cb(pos, "pos_embd", -1);
@@ -7164,7 +7219,7 @@ struct llm_build_starcoder : public llm_graph_context {
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
             if (il == n_layer - 1 && inp_out_ids) {
@@ -7230,7 +7285,7 @@ struct llm_build_refact : public llm_graph_context {
         inpL = build_inp_embd(model.tok_embd);
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
         ggml_tensor * inp_out_ids = build_inp_out_ids();
@@ -7263,7 +7318,7 @@ struct llm_build_refact : public llm_graph_context {
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
             if (il == n_layer - 1 && inp_out_ids) {
@@ -7426,7 +7481,7 @@ struct llm_build_bert : public llm_graph_context {
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
                 cb(cur, "kqv_out", il);
             }
@@ -7571,7 +7626,7 @@ struct llm_build_neo_bert : public llm_graph_context {
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, nullptr,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
                 cb(cur, "kqv_out", il);
             }
@@ -7632,7 +7687,7 @@ struct llm_build_bloom : public llm_graph_context {
         inpL = build_inp_embd(model.tok_embd);
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
         inpL = build_norm(inpL,
                 model.tok_norm,
@@ -7671,7 +7726,7 @@ struct llm_build_bloom : public llm_graph_context {
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
             if (il == n_layer - 1 && inp_out_ids) {
@@ -7739,7 +7794,7 @@ struct llm_build_mpt : public llm_graph_context {
         inpL = build_inp_embd(model.tok_embd);
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
         if (model.pos_embd) {
             // inp_pos - contains the positions
@@ -7819,7 +7874,7 @@ struct llm_build_mpt : public llm_graph_context {
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
             if (il == n_layer - 1 && inp_out_ids) {
@@ -7889,7 +7944,7 @@ struct llm_build_stablelm : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
         ggml_tensor * inp_out_ids = build_inp_out_ids();
@@ -7965,7 +8020,7 @@ struct llm_build_stablelm : public llm_graph_context {
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
             if (il == n_layer - 1 && inp_out_ids) {
@@ -8041,7 +8096,7 @@ struct llm_build_qwen : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
         ggml_tensor * inp_out_ids = build_inp_out_ids();
@@ -8086,7 +8141,7 @@ struct llm_build_qwen : public llm_graph_context {
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
             if (il == n_layer - 1 && inp_out_ids) {
@@ -8156,7 +8211,7 @@ struct llm_build_qwen2 : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
         ggml_tensor * inp_out_ids = build_inp_out_ids();
@@ -8206,7 +8261,7 @@ struct llm_build_qwen2 : public llm_graph_context {
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
             if (il == n_layer - 1 && inp_out_ids) {
@@ -8320,8 +8375,9 @@ struct llm_build_dream : public llm_graph_context {
                 cb(Kcur, "Kcur", il);
                 cb(Vcur, "Vcur", il);
-                cur = build_attn(inp_attn, model.layers[il].wo, model.layers[il].bo, Qcur, Kcur, Vcur, nullptr,
-                                 nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
+                cur = build_attn(inp_attn,
+                        model.layers[il].wo, model.layers[il].bo,
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
             }
             if (il == n_layer - 1 && inp_out_ids) {
@@ -8420,8 +8476,9 @@ struct llm_build_llada : public llm_graph_context {
                 cb(Kcur, "Kcur", il);
                 cb(Vcur, "Vcur", il);
-                cur = build_attn(inp_attn, model.layers[il].wo, NULL, Qcur, Kcur, Vcur, nullptr, nullptr,
-                                 1.0f / sqrtf(float(n_embd_head)), il);
+                cur = build_attn(inp_attn,
+                        model.layers[il].wo, NULL,
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
             }
             if (il == n_layer - 1 && inp_out_ids) {
@@ -8481,7 +8538,7 @@ struct llm_build_qwen2vl : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
         int sections[4];
         std::copy(std::begin(hparams.rope_sections), std::begin(hparams.rope_sections) + 4, sections);
@@ -8534,7 +8591,7 @@ struct llm_build_qwen2vl : public llm_graph_context {
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
             if (il == n_layer - 1 && inp_out_ids) {
@@ -8602,7 +8659,7 @@ struct llm_build_qwen2moe : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
         ggml_tensor * inp_out_ids = build_inp_out_ids();
@@ -8661,7 +8718,7 @@ struct llm_build_qwen2moe : public llm_graph_context {
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
             if (il == n_layer - 1 && inp_out_ids) {
@@ -8761,7 +8818,7 @@ struct llm_build_qwen3 : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
         ggml_tensor * inp_out_ids = build_inp_out_ids();
@@ -8814,7 +8871,7 @@ struct llm_build_qwen3 : public llm_graph_context {
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
             if (il == n_layer - 1 && inp_out_ids) {
@@ -8882,7 +8939,7 @@ struct llm_build_qwen3moe : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
         ggml_tensor * inp_out_ids = build_inp_out_ids();
@@ -8935,7 +8992,7 @@ struct llm_build_qwen3moe : public llm_graph_context {
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
             if (il == n_layer - 1 && inp_out_ids) {
@@ -9012,7 +9069,7 @@ struct llm_build_phi2 : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
         ggml_tensor * inp_out_ids = build_inp_out_ids();
@@ -9075,7 +9132,7 @@ struct llm_build_phi2 : public llm_graph_context {
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f, il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f, il);
             }
             if (il == n_layer - 1 && inp_out_ids) {
@@ -9141,13 +9198,13 @@ struct llm_build_phi3 : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
-        using inp_attn_type = std::conditional_t<iswa, llm_graph_input_attn_kv_unified_iswa, llm_graph_input_attn_kv_unified>;
+        using inp_attn_type = std::conditional_t<iswa, llm_graph_input_attn_kv_iswa, llm_graph_input_attn_kv>;
         inp_attn_type * inp_attn = nullptr;
         if constexpr (iswa) {
-            inp_attn = build_attn_inp_kv_unified_iswa();
+            inp_attn = build_attn_inp_kv_iswa();
         } else {
-            inp_attn = build_attn_inp_kv_unified();
+            inp_attn = build_attn_inp_kv();
         }
         ggml_tensor * inp_out_ids = build_inp_out_ids();
@@ -9212,7 +9269,7 @@ struct llm_build_phi3 : public llm_graph_context {
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f, il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f, il);
             }
             if (il == n_layer - 1 && inp_out_ids) {
@@ -9299,7 +9356,7 @@ struct llm_build_plamo : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
         ggml_tensor * inp_out_ids = build_inp_out_ids();
@@ -9346,7 +9403,7 @@ struct llm_build_plamo : public llm_graph_context {
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
             if (il == n_layer - 1 && inp_out_ids) {
@@ -9415,7 +9472,7 @@ struct llm_build_gpt2 : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
         pos = ggml_get_rows(ctx0, model.pos_embd, inp_pos);
         cb(pos, "pos_embd", -1);
@@ -9454,7 +9511,7 @@ struct llm_build_gpt2 : public llm_graph_context {
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
             if (il == n_layer - 1 && inp_out_ids) {
@@ -9525,7 +9582,7 @@ struct llm_build_codeshell : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
         ggml_tensor * inp_out_ids = build_inp_out_ids();
@@ -9568,7 +9625,7 @@ struct llm_build_codeshell : public llm_graph_context {
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
             if (il == n_layer - 1 && inp_out_ids) {
@@ -9638,7 +9695,7 @@ struct llm_build_orion : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
         ggml_tensor * inp_out_ids = build_inp_out_ids();
@@ -9697,7 +9754,7 @@ struct llm_build_orion : public llm_graph_context {
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
             if (il == n_layer - 1 && inp_out_ids) {
@@ -9765,7 +9822,7 @@ struct llm_build_internlm2 : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
         ggml_tensor * inp_out_ids = build_inp_out_ids();
@@ -9824,7 +9881,7 @@ struct llm_build_internlm2 : public llm_graph_context {
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
             if (il == n_layer - 1 && inp_out_ids) {
@@ -9901,7 +9958,7 @@ struct llm_build_minicpm3 : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
         ggml_tensor * inp_out_ids = build_inp_out_ids();
@@ -10012,7 +10069,7 @@ struct llm_build_minicpm3 : public llm_graph_context {
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, NULL,
-                        q_states, k_states, v_states, nullptr, nullptr, kq_scale, il);
+                        q_states, k_states, v_states, nullptr, nullptr, nullptr, kq_scale, il);
             }
             if (il == n_layer - 1 && inp_out_ids) {
@@ -10096,7 +10153,7 @@ struct llm_build_gemma : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
         ggml_tensor * inp_out_ids = build_inp_out_ids();
@@ -10142,7 +10199,7 @@ struct llm_build_gemma : public llm_graph_context {
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f, il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f, il);
             }
             if (il == n_layer - 1 && inp_out_ids) {
@@ -10212,7 +10269,7 @@ struct llm_build_gemma2_iswa : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
-        auto * inp_attn = build_attn_inp_kv_unified_iswa();
+        auto * inp_attn = build_attn_inp_kv_iswa();
         ggml_tensor * inp_out_ids = build_inp_out_ids();
@@ -10257,7 +10314,7 @@ struct llm_build_gemma2_iswa : public llm_graph_context {
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f, il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f, il);
             }
             if (il == n_layer - 1 && inp_out_ids) {
@@ -10346,7 +10403,7 @@ struct llm_build_gemma3_iswa : public llm_graph_context {
         ggml_tensor * inp_pos = build_inp_pos();
         // TODO: is causal == true correct? might need some changes
-        auto * inp_attn = build_attn_inp_kv_unified_iswa();
+        auto * inp_attn = build_attn_inp_kv_iswa();
         ggml_tensor * inp_out_ids = build_inp_out_ids();
@@ -10399,7 +10456,7 @@ struct llm_build_gemma3_iswa : public llm_graph_context {
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f, il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f, il);
             }
             if (il == n_layer - 1 && inp_out_ids) {
@@ -10471,7 +10528,6 @@ struct llm_build_gemma3n_iswa : public llm_graph_context {
     const int64_t n_embd_altup;
     const int64_t n_altup;
     const int     i_altup_act;
-    const int     n_layer_kv = 20; // number of layers having KV [KV_REUSE]
     const int     n_layer_sparsity = 10; // number of layers using activation sparsity
     const float   f_sparsity_std_mul = 1.6448533535003662f; // std_multiplier = normal_dist.icdf(0.95)
@@ -10497,7 +10553,7 @@ struct llm_build_gemma3n_iswa : public llm_graph_context {
         ggml_tensor * inp_pos = build_inp_pos();
         // TODO: is causal == true correct? might need some changes
-        auto * inp_attn = build_attn_inp_kv_unified_iswa();
+        auto * inp_attn = build_attn_inp_kv_iswa();
         // inp_per_layer shape: [n_embd_altup, n_tokens, n_layer]
         ggml_tensor * inp_per_layer = project_per_layer_inputs(inpL, get_per_layer_inputs());
@@ -10521,8 +10577,6 @@ struct llm_build_gemma3n_iswa : public llm_graph_context {
         for (int il = 0; il < n_layer; ++il) {
             // this block is made to be closely resemble Gemma3p5DecoderLayer on python code
-            const bool has_kv = (il < n_layer_kv);
             const float freq_base_l  = model.get_rope_freq_base (cparams, il);
             const float freq_scale_l = model.get_rope_freq_scale(cparams, il);
@@ -10542,7 +10596,7 @@ struct llm_build_gemma3n_iswa : public llm_graph_context {
             ggml_tensor * laurel_out = laurel(cur, il); // [n_embd, n_tokens]
             // self-attention
-            if (has_kv) {
+            if (hparams.has_kv(il)) {
                 // compute Q and K and RoPE them
                 ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
                 cb(Qcur, "Qcur", il);
@@ -10580,9 +10634,9 @@ struct llm_build_gemma3n_iswa : public llm_graph_context {
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, hparams.f_attention_scale, il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, hparams.f_attention_scale, il);
             } else {
-                // no KV layers
+                // reuse KV cache of earlier layers
                 ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
                 cb(Qcur, "Qcur", il);
                 Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
@@ -10598,7 +10652,7 @@ struct llm_build_gemma3n_iswa : public llm_graph_context {
                 cur = build_attn(inp_attn,
                     model.layers[il].wo, NULL,
-                    Qcur, nullptr, nullptr, nullptr, nullptr, hparams.f_attention_scale, il);
+                    Qcur, nullptr, nullptr, nullptr, nullptr, nullptr, hparams.f_attention_scale, il);
             }
             cur = build_norm(cur,
@@ -10904,7 +10958,7 @@ struct llm_build_starcoder2 : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
         ggml_tensor * inp_out_ids = build_inp_out_ids();
@@ -10963,7 +11017,7 @@ struct llm_build_starcoder2 : public llm_graph_context {
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
             if (il == n_layer - 1 && inp_out_ids) {
@@ -11390,7 +11444,9 @@ struct llm_build_jamba : public llm_graph_context_mamba {
                 cb(Vcur, "Vcur", il);
                 // No RoPE :)
-                cur = build_attn(inp_hybrid->get_attn(), model.layers[il].wo, NULL, Qcur, Kcur, Vcur, NULL, NULL, 1.0f/sqrtf(float(n_embd_head)), il);
+                cur = build_attn(inp_hybrid->get_attn(),
+                        model.layers[il].wo, NULL,
+                        Qcur, Kcur, Vcur, NULL, NULL, NULL, 1.0f/sqrtf(float(n_embd_head)), il);
             }
             if (il == n_layer - 1 && inp_out_ids) {
@@ -11473,7 +11529,7 @@ struct llm_build_command_r : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
         ggml_tensor * inp_out_ids = build_inp_out_ids();
@@ -11548,7 +11604,7 @@ struct llm_build_command_r : public llm_graph_context {
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
             if (il == n_layer - 1 && inp_out_ids) {
@@ -11620,7 +11676,7 @@ struct llm_build_cohere2_iswa : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
-        auto * inp_attn = build_attn_inp_kv_unified_iswa();
+        auto * inp_attn = build_attn_inp_kv_iswa();
         ggml_tensor * inp_out_ids = build_inp_out_ids();
@@ -11683,7 +11739,7 @@ struct llm_build_cohere2_iswa : public llm_graph_context {
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
             if (il == n_layer - 1 && inp_out_ids) {
@@ -11755,7 +11811,7 @@ struct llm_build_olmo : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
         ggml_tensor * inp_out_ids = build_inp_out_ids();
@@ -11814,7 +11870,7 @@ struct llm_build_olmo : public llm_graph_context {
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, nullptr,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
             if (il == n_layer - 1 && inp_out_ids) {
@@ -11883,7 +11939,7 @@ struct llm_build_olmo2 : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
         ggml_tensor * inp_out_ids = build_inp_out_ids();
@@ -11934,7 +11990,7 @@ struct llm_build_olmo2 : public llm_graph_context {
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
             if (il == n_layer - 1 && inp_out_ids) {
@@ -12012,7 +12068,7 @@ struct llm_build_olmoe : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
         ggml_tensor * inp_out_ids = build_inp_out_ids();
@@ -12067,7 +12123,7 @@ struct llm_build_olmoe : public llm_graph_context {
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
             if (il == n_layer - 1 && inp_out_ids) {
@@ -12138,7 +12194,7 @@ struct llm_build_openelm : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
         ggml_tensor * inp_out_ids = build_inp_out_ids();
@@ -12200,7 +12256,7 @@ struct llm_build_openelm : public llm_graph_context {
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
             if (il == n_layer - 1 && inp_out_ids) {
@@ -12269,7 +12325,7 @@ struct llm_build_gptneox : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
         ggml_tensor * inp_out_ids = build_inp_out_ids();
@@ -12312,7 +12368,7 @@ struct llm_build_gptneox : public llm_graph_context {
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
             if (il == n_layer - 1 && inp_out_ids) {
@@ -12415,7 +12471,7 @@ struct llm_build_arctic : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
         ggml_tensor * inp_out_ids = build_inp_out_ids();
@@ -12462,7 +12518,7 @@ struct llm_build_arctic : public llm_graph_context {
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
             if (il == n_layer - 1 && inp_out_ids) {
@@ -12553,7 +12609,7 @@ struct llm_build_deepseek : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
         const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
@@ -12617,7 +12673,7 @@ struct llm_build_deepseek : public llm_graph_context {
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, kq_scale, il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
             }
             if (il == n_layer - 1 && inp_out_ids) {
@@ -12730,7 +12786,7 @@ struct llm_build_deepseek2 : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
         ggml_tensor * inp_out_ids = build_inp_out_ids();
@@ -12845,7 +12901,7 @@ struct llm_build_deepseek2 : public llm_graph_context {
                     // note: MLA with the absorption optimzation converts into MQA (ie: GQA with 1 group)
                     cur = build_attn(inp_attn,
                             model.layers[il].wo, NULL,
-                            Qcur, Kcur, Vcur, nullptr, model.layers[il].wv_b, kq_scale, il);
+                            Qcur, Kcur, Vcur, nullptr, nullptr, model.layers[il].wv_b, kq_scale, il);
                 } else {
                     ggml_tensor * kv = ggml_mul_mat(ctx0, model.layers[il].wkv_b, kv_cmpr);
                     cb(kv, "kv", il);
@@ -12879,7 +12935,7 @@ struct llm_build_deepseek2 : public llm_graph_context {
                     // note: MLA without the absorption optimization converts into MHA (ie: GQA with full n_head groups)
                     cur = build_attn(inp_attn,
                             model.layers[il].wo, NULL,
-                            Qcur, Kcur, Vcur, nullptr, nullptr, kq_scale, il);
+                            Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
                 }
             }
@@ -12977,7 +13033,7 @@ struct llm_build_bitnet : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
         ggml_tensor * inp_out_ids = build_inp_out_ids();
@@ -13046,7 +13102,7 @@ struct llm_build_bitnet : public llm_graph_context {
                 cur = build_attn(inp_attn,
                         NULL, NULL,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
                 cur = build_norm(cur,
                         model.layers[il].attn_sub_norm, NULL,
@@ -13169,7 +13225,7 @@ struct llm_build_t5_enc : public llm_graph_context {
                 cur = build_attn(inp_attn,
                         model.layers[il].wo_enc, nullptr,
-                        Qcur, Kcur, Vcur, kq_b, nullptr, 1.0f, il);
+                        Qcur, Kcur, Vcur, kq_b, nullptr, nullptr, 1.0f, il);
                 cb(cur, "kqv_out", il);
             }
@@ -13241,7 +13297,7 @@ struct llm_build_t5_dec : public llm_graph_context {
         const int64_t n_outputs_enc = embd_enc->ne[1];
-        auto * inp_attn_self  = build_attn_inp_kv_unified();
+        auto * inp_attn_self  = build_attn_inp_kv();
         auto * inp_attn_cross = build_attn_inp_cross();
         ggml_tensor * inp_out_ids = build_inp_out_ids();
@@ -13275,7 +13331,7 @@ struct llm_build_t5_dec : public llm_graph_context {
                 cur = build_attn(inp_attn_self,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, kq_b, nullptr, 1.0f, il);
+                        Qcur, Kcur, Vcur, kq_b, nullptr, nullptr, 1.0f, il);
                 cb(cur, "kqv_out", il);
             }
@@ -13307,7 +13363,7 @@ struct llm_build_t5_dec : public llm_graph_context {
                 cur = build_attn(inp_attn_cross,
                         model.layers[il].wo_cross, nullptr,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f, il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f, il);
                 cb(cur, "kqv_out", il);
                 //ggml_tensor * q =                 ggml_permute(ctx0, Qcur, 0, 2, 1, 3);
@@ -13406,7 +13462,7 @@ struct llm_build_jais : public llm_graph_context {
         inpL = build_inp_embd(model.tok_embd);
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
         ggml_tensor * inp_out_ids = build_inp_out_ids();
@@ -13439,7 +13495,7 @@ struct llm_build_jais : public llm_graph_context {
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/float(n_embd_head), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/float(n_embd_head), il);
             }
             if (il == n_layer - 1 && inp_out_ids) {
@@ -13504,7 +13560,7 @@ struct llm_build_chatglm : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
         ggml_tensor * inp_out_ids = build_inp_out_ids();
@@ -13571,7 +13627,7 @@ struct llm_build_chatglm : public llm_graph_context {
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
             if (il == n_layer - 1 && inp_out_ids) {
@@ -13637,7 +13693,7 @@ struct llm_build_glm4 : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
         ggml_tensor * inp_out_ids = build_inp_out_ids();
@@ -13704,7 +13760,7 @@ struct llm_build_glm4 : public llm_graph_context {
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
             if (il == n_layer - 1 && inp_out_ids) {
@@ -13787,7 +13843,7 @@ struct llm_build_glm4_moe : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
         ggml_tensor * inp_out_ids = build_inp_out_ids();
@@ -13853,7 +13909,7 @@ struct llm_build_glm4_moe : public llm_graph_context {
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
             if (il == n_transformer_layers - 1 && inp_out_ids) {
@@ -13947,7 +14003,7 @@ struct llm_build_nemotron : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
         ggml_tensor * inp_out_ids = build_inp_out_ids();
@@ -14007,7 +14063,7 @@ struct llm_build_nemotron : public llm_graph_context {
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
             if (il == n_layer - 1 && inp_out_ids) {
@@ -14076,7 +14132,7 @@ struct llm_build_exaone : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
         ggml_tensor * inp_out_ids = build_inp_out_ids();
@@ -14138,7 +14194,7 @@ struct llm_build_exaone : public llm_graph_context {
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
             if (il == n_layer - 1 && inp_out_ids) {
@@ -14208,13 +14264,13 @@ struct llm_build_exaone4 : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
-        using inp_attn_type = std::conditional_t<iswa, llm_graph_input_attn_kv_unified_iswa, llm_graph_input_attn_kv_unified>;
+        using inp_attn_type = std::conditional_t<iswa, llm_graph_input_attn_kv_iswa, llm_graph_input_attn_kv>;
         inp_attn_type * inp_attn = nullptr;
         if constexpr (iswa) {
-            inp_attn = build_attn_inp_kv_unified_iswa();
+            inp_attn = build_attn_inp_kv_iswa();
         } else {
-            inp_attn = build_attn_inp_kv_unified();
+            inp_attn = build_attn_inp_kv();
         }
         ggml_tensor * inp_out_ids = build_inp_out_ids();
@@ -14269,7 +14325,7 @@ struct llm_build_exaone4 : public llm_graph_context {
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
                 cb(cur, "attn_out", il);
             }
@@ -15097,7 +15153,7 @@ struct llm_build_granite : public llm_graph_context {
             inp_pos = build_inp_pos();
         }
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
         ggml_tensor * inp_out_ids = build_inp_out_ids();
@@ -15148,12 +15204,12 @@ struct llm_build_granite : public llm_graph_context {
     }
     ggml_tensor * build_attention_layer(
-              ggml_tensor                     * cur,
-              ggml_tensor                     * inp_pos,
-              llm_graph_input_attn_kv_unified * inp_attn,
-        const llama_model                     & model,
-        const int64_t                           n_embd_head,
-        const int                               il) {
+              ggml_tensor             * cur,
+              ggml_tensor             * inp_pos,
+              llm_graph_input_attn_kv * inp_attn,
+        const llama_model             & model,
+        const int64_t                 n_embd_head,
+        const int                     il) {
         // compute Q and K and (optionally) RoPE them
         ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
@@ -15204,7 +15260,7 @@ struct llm_build_granite : public llm_graph_context {
         const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
         cur = build_attn(inp_attn,
                 model.layers[il].wo, model.layers[il].bo,
-                Qcur, Kcur, Vcur, nullptr, nullptr, kq_scale, il);
+                Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
                 cb(cur, "attn_out", il);
         return cur;
     }
@@ -15367,12 +15423,12 @@ struct llm_build_granite_hybrid : public llm_graph_context_mamba {
     }
     ggml_tensor * build_attention_layer(
-              ggml_tensor                     * cur,
-              ggml_tensor                     * inp_pos,
-              llm_graph_input_attn_kv_unified * inp_attn,
-        const llama_model                     & model,
-        const int64_t                           n_embd_head,
-        const int                               il) {
+              ggml_tensor             * cur,
+              ggml_tensor             * inp_pos,
+              llm_graph_input_attn_kv * inp_attn,
+        const llama_model             & model,
+        const int64_t                 n_embd_head,
+        const int                     il) {
         // compute Q and K and (optionally) RoPE them
         ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
@@ -15423,7 +15479,7 @@ struct llm_build_granite_hybrid : public llm_graph_context_mamba {
         const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
         cur = build_attn(inp_attn,
                 model.layers[il].wo, model.layers[il].bo,
-                Qcur, Kcur, Vcur, nullptr, nullptr, kq_scale, il);
+                Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
                 cb(cur, "attn_out", il);
         return cur;
     }
@@ -15529,7 +15585,7 @@ struct llm_build_chameleon : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
         ggml_tensor * inp_out_ids = build_inp_out_ids();
@@ -15608,7 +15664,7 @@ struct llm_build_chameleon : public llm_graph_context {
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, nullptr,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
             if (il == n_layer - 1 && inp_out_ids) {
@@ -15860,7 +15916,7 @@ struct llm_build_plm : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
         ggml_tensor * inp_out_ids = build_inp_out_ids();
@@ -15964,7 +16020,7 @@ struct llm_build_plm : public llm_graph_context {
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, NULL,
-                        q_states, k_states, v_states, nullptr, nullptr, kq_scale, il);
+                        q_states, k_states, v_states, nullptr, nullptr, nullptr, kq_scale, il);
             }
             if (il == n_layer - 1 && inp_out_ids) {
@@ -16025,7 +16081,7 @@ struct llm_build_bailingmoe : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
         ggml_tensor * inp_out_ids = build_inp_out_ids();
@@ -16087,7 +16143,7 @@ struct llm_build_bailingmoe : public llm_graph_context {
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_rot)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_rot)), il);
             }
             if (il == n_layer - 1 && inp_out_ids) {
@@ -16174,7 +16230,7 @@ struct llm_build_dots1 : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
         ggml_tensor * inp_out_ids = build_inp_out_ids();
@@ -16227,7 +16283,7 @@ struct llm_build_dots1 : public llm_graph_context {
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
             if (il == n_layer - 1 && inp_out_ids) {
@@ -16324,7 +16380,7 @@ struct llm_build_ernie4_5 : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
         for (int il = 0; il < n_layer; ++il) {
             ggml_tensor * inpSA = inpL;
@@ -16382,7 +16438,7 @@ struct llm_build_ernie4_5 : public llm_graph_context {
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
             if (il == n_layer - 1) {
@@ -16454,7 +16510,7 @@ struct llm_build_ernie4_5_moe : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
         ggml_tensor * inp_out_ids = build_inp_out_ids();
@@ -16515,7 +16571,7 @@ struct llm_build_ernie4_5_moe : public llm_graph_context {
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
                 cb(cur, "attn_out", il);
             }
@@ -16668,7 +16724,7 @@ struct llm_build_falcon_h1 : public llm_graph_context_mamba {
             ggml_tensor * attn_out = build_attn(inp->get_attn(),
                     model.layers[il].wo, NULL,
-                    Qcur, Kcur, Vcur, nullptr, nullptr, kq_scale, il);
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
             cb(attn_out, "attn_out", il);
             cur = build_norm(inpL,
@@ -16828,7 +16884,7 @@ struct llm_build_plamo2 : public llm_graph_context_mamba {
 private:
     ggml_tensor * build_plamo2_attn_layer(
-            llm_graph_input_attn_kv_unified * inp,
+            llm_graph_input_attn_kv * inp,
             ggml_tensor * inp_pos,
             ggml_tensor * cur,
             const llama_model & model,
@@ -16878,7 +16934,9 @@ private:
                     ext_factor, attn_factor, beta_fast, beta_slow
                     );
-            cur = build_attn(inp, model.layers[il].wo, NULL, Qcur, Kcur, Vcur, NULL, NULL, 1.0f/sqrtf(float(n_embd_head_v)), il);
+            cur = build_attn(inp,
+                    model.layers[il].wo, NULL,
+                    Qcur, Kcur, Vcur, NULL, NULL, NULL, 1.0f/sqrtf(float(n_embd_head_v)), il);
         }
         cb(cur, "attn_out", il);
@@ -17061,7 +17119,7 @@ struct llm_build_arcee : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
         const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
@@ -17125,7 +17183,7 @@ struct llm_build_arcee : public llm_graph_context {
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, kq_scale, il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
                 cb(cur, "attn_out", il);
             }
@@ -17196,7 +17254,7 @@ struct llm_build_hunyuan_moe : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
         const float kq_scale = 1.0f / sqrtf(float(n_embd_head));
@@ -17270,7 +17328,7 @@ struct llm_build_hunyuan_moe : public llm_graph_context {
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, kq_scale, il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
                 cb(cur, "attn_out", il);
             }
@@ -17357,7 +17415,7 @@ struct llm_build_hunyuan_dense : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
         const float kq_scale = 1.0f / sqrtf(float(n_embd_head));
@@ -17430,7 +17488,7 @@ struct llm_build_hunyuan_dense : public llm_graph_context {
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, kq_scale, il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
                 cb(cur, "attn_out", il);
             }
@@ -17495,7 +17553,7 @@ struct llm_build_smollm3 : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
         const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
@@ -17560,7 +17618,7 @@ struct llm_build_smollm3 : public llm_graph_context {
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, kq_scale, il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
                 cb(cur, "attn_out", il);
             }
@@ -17627,7 +17685,7 @@ struct llm_build_openai_moe_iswa : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
-        auto * inp_attn = build_attn_inp_kv_unified_iswa();
+        auto * inp_attn = build_attn_inp_kv_iswa();
         for (int il = 0; il < n_layer; ++il) {
             ggml_tensor * inpSA = inpL;
@@ -17682,9 +17740,9 @@ struct llm_build_openai_moe_iswa : public llm_graph_context {
                 cb(Kcur, "Kcur", il);
                 cb(Vcur, "Vcur", il);
-                cur = build_attn_with_sinks(inp_attn,
+                cur = build_attn(inp_attn,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, model.layers[il].attn_sinks, 1.0f/sqrtf(float(n_rot)), il);
+                        Qcur, Kcur, Vcur, nullptr, model.layers[il].attn_sinks, nullptr, 1.0f/sqrtf(float(n_rot)), il);
                 cb(cur, "attn_out", il);
             }
@@ -17781,8 +17839,7 @@ struct llm_build_lfm2 : public llm_graph_context {
         cb(cur, "model.embedding_norm", -1);
         res->t_embd = cur;
-        // lm_head is tied with embeddings
-        cur = build_lora_mm(model.tok_embd, cur);
+        cur = build_lora_mm(model.output, cur);
         cb(cur, "lm_head", -1);
         res->t_logits = cur;
@@ -17809,10 +17866,10 @@ struct llm_build_lfm2 : public llm_graph_context {
         return cur;
     }
-    ggml_tensor * build_attn_block(ggml_tensor                     * cur,
-                                   ggml_tensor                     * inp_pos,
-                                   llm_graph_input_attn_kv_unified * inp_attn,
-                                   int                               il) const {
+    ggml_tensor * build_attn_block(ggml_tensor             * cur,
+                                   ggml_tensor             * inp_pos,
+                                   llm_graph_input_attn_kv * inp_attn,
+                                   int                     il) const {
         GGML_ASSERT(hparams.n_embd_v_gqa(il) == hparams.n_embd_k_gqa(il));
         auto const n_embd_head = hparams.n_embd_head_v;
         auto const n_head_kv = hparams.n_head_kv(il);
@@ -17847,7 +17904,7 @@ struct llm_build_lfm2 : public llm_graph_context {
                 );
         cur = build_attn(inp_attn, model.layers[il].wo, NULL,
-                q, k, v, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                q, k, v, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
         cb(cur, "model.layers.{}.self_attn.out_proj", il);
@@ -17924,6 +17981,137 @@ struct llm_build_lfm2 : public llm_graph_context {
     }
 };
+struct llm_build_seed_oss : public llm_graph_context {
+    llm_build_seed_oss(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+        const int64_t n_embd_head = hparams.n_embd_head_v;
+        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+        GGML_ASSERT(n_embd_head == hparams.n_rot);
+        ggml_tensor * cur;
+        ggml_tensor * inpL;
+        inpL = build_inp_embd(model.tok_embd);
+        // inp_pos - contains the positions
+        ggml_tensor * inp_pos = build_inp_pos();
+        auto * inp_attn = build_attn_inp_kv();
+        const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
+        ggml_tensor * inp_out_ids = build_inp_out_ids();
+        for (int il = 0; il < n_layer; ++il) {
+            ggml_tensor * inpSA = inpL;
+            // norm
+            cur = build_norm(inpL,
+                    model.layers[il].attn_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "attn_norm", il);
+            // self-attention
+            {
+                // compute Q and K and RoPE them
+                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+                cb(Qcur, "Qcur", il);
+                if (model.layers[il].bq) {
+                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                    cb(Qcur, "Qcur", il);
+                }
+                ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+                cb(Kcur, "Kcur", il);
+                if (model.layers[il].bk) {
+                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                    cb(Kcur, "Kcur", il);
+                }
+                ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+                cb(Vcur, "Vcur", il);
+                if (model.layers[il].bv) {
+                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                    cb(Vcur, "Vcur", il);
+                }
+                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+                Qcur = ggml_rope_ext(
+                        ctx0, Qcur, inp_pos, nullptr,
+                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                        ext_factor, attn_factor, beta_fast, beta_slow
+                        );
+                Kcur = ggml_rope_ext(
+                        ctx0, Kcur, inp_pos, nullptr,
+                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                        ext_factor, attn_factor, beta_fast, beta_slow
+                        );
+                cb(Qcur, "Qcur", il);
+                cb(Kcur, "Kcur", il);
+                cb(Vcur, "Vcur", il);
+                cur = build_attn(inp_attn,
+                        model.layers[il].wo, model.layers[il].bo,
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
+                cb(cur, "attn_out", il);
+            }
+            if (il == n_layer - 1 && inp_out_ids) {
+                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+            }
+            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+            cb(ffn_inp, "ffn_inp", il);
+            // feed-forward network
+            cur = build_norm(ffn_inp,
+                    model.layers[il].attn_post_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "attn_post_norm", il);
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up,   NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL,
+                    LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+            cur = ggml_add(ctx0, cur, ffn_inp);
+            cb(cur, "ffn_out", il);
+            cur = build_cvec(cur, il);
+            cb(cur, "l_out", il);
+            // input for next layer
+            inpL = cur;
+        }
+        cur = inpL;
+        cur = build_norm(cur,
+                model.output_norm, NULL,
+                LLM_NORM_RMS, -1);
+        cb(cur, "result_norm", -1);
+        res->t_embd = cur;
+        // lm_head
+        cur = build_lora_mm(model.output, cur);
+        cb(cur, "result_output", -1);
+        res->t_logits = cur;
+        ggml_build_forward_expand(gf, cur);
+    }
+};
 template <bool iswa>
 struct llm_build_smallthinker : public llm_graph_context{
     llm_build_smallthinker(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params){
@@ -17940,13 +18128,13 @@ struct llm_build_smallthinker : public llm_graph_context{
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
-        using inp_attn_type = std::conditional_t<iswa, llm_graph_input_attn_kv_unified_iswa, llm_graph_input_attn_kv_unified>;
+        using inp_attn_type = std::conditional_t<iswa, llm_graph_input_attn_kv_iswa, llm_graph_input_attn_kv>;
         inp_attn_type * inp_attn = nullptr;
         if constexpr (iswa) {
-            inp_attn = build_attn_inp_kv_unified_iswa();
+            inp_attn = build_attn_inp_kv_iswa();
         } else {
-            inp_attn = build_attn_inp_kv_unified();
+            inp_attn = build_attn_inp_kv();
         }
         ggml_tensor * inp_out_ids = build_inp_out_ids();
@@ -17991,7 +18179,7 @@ struct llm_build_smallthinker : public llm_graph_context{
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
             }
             if (il == n_layer - 1 && inp_out_ids) {
@@ -18069,14 +18257,14 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
                 if (llm_arch_is_recurrent(arch)) {
                     res = new llama_memory_recurrent(
                             *this,
-                            nullptr,
                             GGML_TYPE_F32,
                             GGML_TYPE_F32,
                             cparams.offload_kqv,
                             std::max((uint32_t) 1, cparams.n_seq_max),
-                            cparams.n_seq_max);
+                            cparams.n_seq_max,
+                            nullptr);
                 } else if (llm_arch_is_hybrid(arch)) {
-                    const auto padding = llama_kv_cache_unified::get_padding(cparams);
+                    const auto padding = llama_kv_cache::get_padding(cparams);
                     cparams.n_ctx = GGML_PAD(cparams.n_ctx, padding);
@@ -18098,7 +18286,7 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
                         /* filter_attn       */ (arch == LLM_ARCH_FALCON_H1) ? [&](int32_t) { return true; } : (llama_memory_hybrid::layer_filter_cb)nullptr,
                         /* filter_recr       */ (arch == LLM_ARCH_FALCON_H1) ? [&](int32_t) { return true; } : (llama_memory_hybrid::layer_filter_cb)nullptr);
                 } else {
-                    const auto padding = llama_kv_cache_unified::get_padding(cparams);
+                    const auto padding = llama_kv_cache::get_padding(cparams);
                     uint32_t n_ctx_per_stream = cparams.n_ctx;
@@ -18115,10 +18303,22 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
                     LLAMA_LOG_DEBUG("%s: n_ctx = %u (padded)\n", __func__, cparams.n_ctx);
+                    llama_memory_i::layer_reuse_cb reuse = nullptr;
+                    if (arch == LLM_ARCH_GEMMA3N) {
+                        reuse = [&](int32_t il) {
+                            if (il >= (int32_t) hparams.n_layer_kv_from_start) {
+                                return (int32_t) hparams.n_layer_kv_from_start - (hparams.is_swa(il) ? 2 : 1);
+                            }
+                            return -1;
+                        };
+                    }
                     if (hparams.swa_type != LLAMA_SWA_TYPE_NONE) {
                         GGML_ASSERT(hparams.is_swa_any());
-                        res = new llama_kv_cache_unified_iswa(
+                        res = new llama_kv_cache_iswa(
                                 *this,
                                 params.type_k,
                                 params.type_v,
@@ -18129,13 +18329,14 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
                                 n_ctx_per_stream,
                                 cparams.n_seq_max,
                                 cparams.n_ubatch,
-                                padding);
+                                padding,
+                                nullptr,
+                                reuse);
                     } else {
                         GGML_ASSERT(!hparams.is_swa_any());
-                        res = new llama_kv_cache_unified(
+                        res = new llama_kv_cache(
                                 *this,
-                                nullptr,
                                 params.type_k,
                                 params.type_v,
                                 !cparams.flash_attn,
@@ -18145,7 +18346,9 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
                                 cparams.n_seq_max,
                                 padding,
                                 hparams.n_swa,
-                                hparams.swa_type);
+                                hparams.swa_type,
+                                nullptr,
+                                nullptr);
                     }
                 }
             }
@@ -18462,6 +18665,10 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
             {
                 llm = std::make_unique<llm_build_bailingmoe>(*this, params);
             } break;
+        case LLM_ARCH_SEED_OSS:
+            {
+                llm = std::make_unique<llm_build_seed_oss>(*this, params);
+            } break;
         case LLM_ARCH_DOTS1:
             {
                 llm = std::make_unique<llm_build_dots1>(*this, params);
@@ -18520,6 +18727,7 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
     return llm->res->get_gf();
 }
 //
 // interface implementation
 //
@@ -18714,6 +18922,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
         case LLM_ARCH_LFM2:
         case LLM_ARCH_SMALLTHINKER:
         case LLM_ARCH_GLM4_MOE:
+        case LLM_ARCH_SEED_OSS:
             return LLAMA_ROPE_TYPE_NEOX;
         case LLM_ARCH_QWEN2VL: