cui-llama.rn 1.5.0 → 1.6.1

package/cpp/llama-model.cpp

@@ -40,14 +40,17 @@ const char * llm_type_name(llm_type type) {
         case LLM_TYPE_335M:          return "335M";
         case LLM_TYPE_410M:          return "410M";
         case LLM_TYPE_450M:          return "450M";
+        case LLM_TYPE_475M:          return "475M";
         case LLM_TYPE_770M:          return "770M";
         case LLM_TYPE_780M:          return "780M";
         case LLM_TYPE_0_5B:          return "0.5B";
+        case LLM_TYPE_0_6B:          return "0.6B";
         case LLM_TYPE_1B:            return "1B";
         case LLM_TYPE_1_3B:          return "1.3B";
         case LLM_TYPE_1_4B:          return "1.4B";
         case LLM_TYPE_1_5B:          return "1.5B";
         case LLM_TYPE_1_6B:          return "1.6B";
+        case LLM_TYPE_1_7B:          return "1.7B";
         case LLM_TYPE_1_8B:          return "1.8B";
         case LLM_TYPE_2B:            return "2B";
         case LLM_TYPE_2_8B:          return "2.8B";
@@ -66,6 +69,7 @@ const char * llm_type_name(llm_type type) {
         case LLM_TYPE_15B:           return "15B";
         case LLM_TYPE_16B:           return "16B";
         case LLM_TYPE_20B:           return "20B";
+        case LLM_TYPE_27B:           return "27B";
         case LLM_TYPE_30B:           return "30B";
         case LLM_TYPE_32B:           return "32B";
         case LLM_TYPE_34B:           return "34B";
@@ -74,6 +78,7 @@ const char * llm_type_name(llm_type type) {
         case LLM_TYPE_65B:           return "65B";
         case LLM_TYPE_70B:           return "70B";
         case LLM_TYPE_236B:          return "236B";
+        case LLM_TYPE_290B:          return "290B";
         case LLM_TYPE_314B:          return "314B";
         case LLM_TYPE_671B:          return "671B";
         case LLM_TYPE_SMALL:         return "0.1B";
@@ -88,10 +93,10 @@ const char * llm_type_name(llm_type type) {
         case LLM_TYPE_16x3_8B:       return "16x3.8B";
         case LLM_TYPE_10B_128x3_66B: return "10B+128x3.66B";
         case LLM_TYPE_57B_A14B:      return "57B.A14B";
-        case LLM_TYPE_27B:           return "27B";
-        case LLM_TYPE_290B:          return "290B";
         case LLM_TYPE_17B_16E:       return "17Bx16E (Scout)";
         case LLM_TYPE_17B_128E:      return "17Bx128E (Maverick)";
+        case LLM_TYPE_30B_A3B:       return "30B.A3B";
+        case LLM_TYPE_235B_A22B:     return "235B.A22B";
         default:                     return "?B";
     }
 }
@@ -695,13 +700,19 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                 }
             } break;
         case LLM_ARCH_NOMIC_BERT:
+        case LLM_ARCH_NOMIC_BERT_MOE:
             {
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS,    hparams.f_norm_eps);
                 ml.get_key(LLM_KV_ATTENTION_CAUSAL,           hparams.causal_attn);
                 ml.get_key(LLM_KV_POOLING_TYPE,               hparams.pooling_type);
+                ml.get_key(LLM_KV_MOE_EVERY_N_LAYERS,         hparams.moe_every_n_layers, 0);
 
                 if (hparams.n_layer == 12 && hparams.n_embd == 768) {
-                    type = LLM_TYPE_137M;
+                    if (arch == LLM_ARCH_NOMIC_BERT) {
+                        type = LLM_TYPE_137M;
+                    } else if (arch == LLM_ARCH_NOMIC_BERT_MOE && hparams.moe_every_n_layers == 2) {
+                        type = LLM_TYPE_475M;
+                    }
                 }
             } break;
         case LLM_ARCH_BLOOM:
@@ -762,6 +773,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
             // fall through
         case LLM_ARCH_QWEN2:
             {
+                ml.get_key(LLM_KV_POOLING_TYPE, hparams.pooling_type, false);
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
                 switch (hparams.n_layer) {
                     case 24: type = hparams.n_embd == 1024 ? LLM_TYPE_0_5B : LLM_TYPE_1B; break;
@@ -791,6 +803,10 @@ void llama_model::load_hparams(llama_model_loader & ml) {
             {
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
                 switch (hparams.n_layer) {
+                    case 28: type = hparams.n_embd == 1024 ? LLM_TYPE_0_6B : LLM_TYPE_1_7B; break;
+                    case 36: type = hparams.n_embd == 2560 ? LLM_TYPE_4B : LLM_TYPE_8B; break;
+                    case 40: type = LLM_TYPE_14B; break;
+                    case 64: type = LLM_TYPE_32B; break;
                     default: type = LLM_TYPE_UNKNOWN;
                 }
             } break;
@@ -800,6 +816,8 @@ void llama_model::load_hparams(llama_model_loader & ml) {
 
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
                 switch (hparams.n_layer) {
+                    case 48: type = LLM_TYPE_30B_A3B; break;
+                    case 94: type = LLM_TYPE_235B_A22B; break;
                     default: type = LLM_TYPE_UNKNOWN;
                 }
             } break;
@@ -1156,6 +1174,8 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                     ml.get_key(LLM_KV_ATTENTION_Q_LORA_RANK, hparams.n_lora_q);
                 }
                 ml.get_key(LLM_KV_ATTENTION_KV_LORA_RANK,     hparams.n_lora_kv);
+                ml.get_key(LLM_KV_ATTENTION_KEY_LENGTH_MLA,   hparams.n_embd_head_k_mla, false);
+                ml.get_key(LLM_KV_ATTENTION_VALUE_LENGTH_MLA, hparams.n_embd_head_v_mla, false);
                 ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp);
                 ml.get_key(LLM_KV_EXPERT_SHARED_COUNT,        hparams.n_expert_shared);
                 ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE,       hparams.expert_weights_scale);
@@ -1205,6 +1225,15 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                     default: type = LLM_TYPE_UNKNOWN;
                 }
             } break;
+        case LLM_ARCH_GLM4:
+            {
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+                switch (hparams.n_layer) {
+                    case 40: type = LLM_TYPE_9B; break;
+                    case 61: type = LLM_TYPE_32B; break;
+                    default: type = LLM_TYPE_UNKNOWN;
+                }
+            } break;
         case LLM_ARCH_BITNET:
             {
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
@@ -2046,6 +2075,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                 } break;
             case LLM_ARCH_BERT:
             case LLM_ARCH_NOMIC_BERT:
+            case LLM_ARCH_NOMIC_BERT_MOE:
                 {
                     tok_embd     = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD,  "weight"), {n_embd, n_vocab}, 0);
                     type_embd    = create_tensor(tn(LLM_TENSOR_TOKEN_TYPES, "weight"), {n_embd, n_token_types}, 0);
@@ -2079,20 +2109,31 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                             layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, 0);
                         }
 
+                        if (arch == LLM_ARCH_NOMIC_BERT_MOE) {
+                            layer.bqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "bias", i), {n_embd + 2*n_embd_gqa}, 0);
+                        }
+
                         layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT,      "weight", i), {n_embd, n_embd}, 0);
 
                         layer.attn_out_norm   = create_tensor(tn(LLM_TENSOR_ATTN_OUT_NORM, "weight", i), {n_embd}, 0);
                         layer.attn_out_norm_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT_NORM, "bias", i),   {n_embd}, 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);
-
-                        if (arch == LLM_ARCH_BERT) {
+                        if (hparams.moe_every_n_layers > 0 && i % hparams.moe_every_n_layers == 1) {
                             layer.bo         = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, 0);
-                            layer.ffn_up_b   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "bias", i), {n_ff}, 0);
-                            layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}, 0);
+                            layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {  n_embd, n_ff,   n_expert}, 0);
+                            layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {  n_ff,   n_embd, n_expert}, 0);
+                            layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,   "weight", i), {n_embd, n_expert}, 0);
                         } else {
-                            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);
+
+                            if (arch == LLM_ARCH_BERT || arch == LLM_ARCH_NOMIC_BERT_MOE) {
+                                layer.bo         = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, 0);
+                                layer.ffn_up_b   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "bias", i), {n_ff}, 0);
+                                layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}, 0);
+                            } else {
+                                layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0);
+                            }
                         }
 
                         layer.layer_out_norm   = create_tensor(tn(LLM_TENSOR_LAYER_OUT_NORM, "weight", i), {n_embd}, 0);
@@ -3196,8 +3237,14 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                 {
                     const bool is_lite = (hparams.n_layer == 27);
 
+                    const bool is_mla = (hparams.n_embd_head_k_mla != 0 && hparams.n_embd_head_v_mla != 0);
+
+                    // note: these are the actual head sizes you get when treating as MHA or after "decompression" using wv_b for MLA
+                    const int64_t n_embd_head_k_mla = is_mla ? hparams.n_embd_head_k_mla : hparams.n_embd_head_k;
+                    const int64_t n_embd_head_v_mla = is_mla ? hparams.n_embd_head_v_mla : hparams.n_embd_head_v;
+
                     const int64_t n_embd_head_qk_rope = hparams.n_rot;
-                    const int64_t n_embd_head_qk_nope = hparams.n_embd_head_k - hparams.n_rot;
+                    const int64_t n_embd_head_qk_nope = n_embd_head_k_mla - n_embd_head_qk_rope;
 
                     const int64_t q_lora_rank  = hparams.n_lora_q;
                     const int64_t kv_lora_rank = hparams.n_lora_kv;
@@ -3223,14 +3270,22 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
 
                         if (!is_lite) {
                             layer.wq_a = create_tensor(tn(LLM_TENSOR_ATTN_Q_A, "weight", i), {n_embd, q_lora_rank}, 0);
-                            layer.wq_b = create_tensor(tn(LLM_TENSOR_ATTN_Q_B, "weight", i), {q_lora_rank, n_head * n_embd_head_k}, 0);
+                            layer.wq_b = create_tensor(tn(LLM_TENSOR_ATTN_Q_B, "weight", i), {q_lora_rank, n_head * n_embd_head_k_mla}, 0);
                         } else {
-                            layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd_k_gqa}, 0);
+                            layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_head * n_embd_head_k_mla}, 0);
                         }
 
-                        layer.wkv_a_mqa = create_tensor(tn(LLM_TENSOR_ATTN_KV_A_MQA, "weight", i), {n_embd, kv_lora_rank + (n_embd_head_qk_rope)}, 0);
-                        layer.wkv_b     = create_tensor(tn(LLM_TENSOR_ATTN_KV_B,     "weight", i), {kv_lora_rank, n_head * (n_embd_head_qk_nope + n_embd_head_v)}, 0);
-                        layer.wo        = create_tensor(tn(LLM_TENSOR_ATTN_OUT,      "weight", i), {              n_head * (                      n_embd_head_v), n_embd}, 0);
+                        layer.wkv_a_mqa = create_tensor(tn(LLM_TENSOR_ATTN_KV_A_MQA, "weight", i), {n_embd, kv_lora_rank + n_embd_head_qk_rope}, 0);
+
+                        // note: only old legacy GGUF files will have the unsplit wkv_b tensor in
+                        if (is_mla) {
+                            layer.wk_b = create_tensor(tn(LLM_TENSOR_ATTN_K_B, "weight", i), {n_embd_head_qk_nope, kv_lora_rank, n_head}, 0);
+                            layer.wv_b = create_tensor(tn(LLM_TENSOR_ATTN_V_B, "weight", i), {kv_lora_rank, n_embd_head_v_mla, n_head}, 0);
+                        } else {
+                            layer.wkv_b = create_tensor(tn(LLM_TENSOR_ATTN_KV_B, "weight", i), {kv_lora_rank, n_head * (n_embd_head_qk_nope + n_embd_head_v_mla)}, 0);
+                        }
+
+                        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_head * n_embd_head_v_mla, n_embd}, 0);
 
                         layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
 
@@ -3476,6 +3531,45 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         layer.ffn_down   = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, 0);
                     }
                 } break;
+            case LLM_ARCH_GLM4:
+                {
+                    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.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+                        layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, TENSOR_NOT_REQUIRED);
+                        layer.bqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "bias", i),   {n_embd + 2*n_embd_gqa}, TENSOR_NOT_REQUIRED);
+
+                        if (layer.wqkv == nullptr) {
+                            layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd_head_k * n_head}, 0);
+                            layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_k_gqa}, 0);
+                            layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_v_gqa}, 0);
+                            layer.bq = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
+                            layer.bk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "bias", i), {n_embd_gqa}, TENSOR_NOT_REQUIRED);
+                            layer.bv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "bias", i), {n_embd_gqa}, TENSOR_NOT_REQUIRED);
+                        }
+
+                        layer.wo   = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
+
+                        layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), {n_embd}, 0);
+
+                        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+                        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
+                        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_ff * 2}, 0);
+
+                        layer.ffn_post_norm  = create_tensor(tn(LLM_TENSOR_FFN_POST_NORM, "weight", i), {n_embd}, 0);
+                    }
+                } break;
             case LLM_ARCH_NEMOTRON:
                 {
                     tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
@@ -4242,6 +4336,8 @@ void llama_model::print_info() const {
         LLAMA_LOG_INFO("%s: n_layer_dense_lead   = %d\n",     __func__, hparams.n_layer_dense_lead);
         LLAMA_LOG_INFO("%s: n_lora_q             = %d\n",     __func__, hparams.n_lora_q);
         LLAMA_LOG_INFO("%s: n_lora_kv            = %d\n",     __func__, hparams.n_lora_kv);
+        LLAMA_LOG_INFO("%s: n_embd_head_k_mla    = %d\n",     __func__, hparams.n_embd_head_k_mla);
+        LLAMA_LOG_INFO("%s: n_embd_head_v_mla    = %d\n",     __func__, hparams.n_embd_head_v_mla);
         LLAMA_LOG_INFO("%s: n_ff_exp             = %d\n",     __func__, hparams.n_ff_exp);
         LLAMA_LOG_INFO("%s: n_expert_shared      = %d\n",     __func__, hparams.n_expert_shared);
         LLAMA_LOG_INFO("%s: expert_weights_scale = %.1f\n",   __func__, hparams.expert_weights_scale);
@@ -4350,6 +4446,19 @@ const lm_ggml_tensor * llama_model::get_tensor(const char * name) const {
     return it->second;
 }
 
+lm_ggml_tensor * llama_model::get_rope_factors(uint32_t n_ctx_per_seq, int il) const {
+    // choose long/short freq factors based on the context size
+    if (layers[il].rope_freqs != nullptr) {
+        return layers[il].rope_freqs;
+    }
+
+    if (n_ctx_per_seq > hparams.n_ctx_orig_yarn) {
+        return layers[il].rope_long;
+    }
+
+    return layers[il].rope_short;
+}
+
 struct llm_build_llama : public llm_graph_context {
     llm_build_llama(const llama_model & model, const llm_graph_params & params, lm_ggml_cgraph * gf) : llm_graph_context(params) {
         const int64_t n_embd_head = hparams.n_embd_head_v;
@@ -4390,7 +4499,7 @@ struct llm_build_llama : public llm_graph_context {
             // self-attention
             {
                 // rope freq factors for llama3; may return nullptr for llama2 and other models
-                lm_ggml_tensor * rope_factors = static_cast<const llama_kv_cache_unified *>(memory)->cbs.get_rope_factors(n_ctx_per_seq, il);
+                lm_ggml_tensor * rope_factors = model.get_rope_factors(n_ctx_per_seq, il);
 
                 // compute Q and K and RoPE them
                 lm_ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
@@ -4440,15 +4549,15 @@ struct llm_build_llama : public llm_graph_context {
 
                 if (arch == LLM_ARCH_LLAMA4 && use_rope && hparams.use_kq_norm) {
                     // Llama4TextL2Norm
-                    Qcur = lm_ggml_rms_norm(ctx0, Qcur, 1e-6);
-                    Kcur = lm_ggml_rms_norm(ctx0, Kcur, 1e-6);
+                    Qcur = lm_ggml_rms_norm(ctx0, Qcur, hparams.f_norm_rms_eps);
+                    Kcur = lm_ggml_rms_norm(ctx0, Kcur, hparams.f_norm_rms_eps);
                     cb(Qcur, "Qcur_normed", il);
                     cb(Kcur, "Kcur_normed", il);
                 }
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, kq_scale, il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, kq_scale, il);
                 cb(cur, "attn_out", il);
             }
 
@@ -4615,7 +4724,7 @@ struct llm_build_deci : public llm_graph_context {
             } else if (n_head > 0) {
                 // self-attention
                 // rope freq factors for llama3; may return nullptr for llama2 and other models
-                lm_ggml_tensor * rope_factors = static_cast<const llama_kv_cache_unified *>(memory)->cbs.get_rope_factors(n_ctx_per_seq, il);
+                lm_ggml_tensor * rope_factors = model.get_rope_factors(n_ctx_per_seq, il);
 
                 // compute Q and K and RoPE them
                 lm_ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
@@ -4661,7 +4770,7 @@ struct llm_build_deci : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, kq_scale, il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, kq_scale, il);
             }
 
             if (il == n_layer - 1) {
@@ -4803,7 +4912,7 @@ struct llm_build_baichuan : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -4918,7 +5027,7 @@ struct llm_build_xverse : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -5043,7 +5152,7 @@ struct llm_build_falcon : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -5173,7 +5282,7 @@ struct llm_build_grok : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f, il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f, il);
             }
 
             if (il == n_layer - 1) {
@@ -5324,7 +5433,7 @@ struct llm_build_dbrx : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -5438,7 +5547,7 @@ struct llm_build_starcoder : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -5537,7 +5646,7 @@ struct llm_build_refact : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -5664,6 +5773,11 @@ struct llm_build_bert : public llm_graph_context {
                 cur = build_lora_mm(model.layers[il].wqkv, cur);
                 cb(cur, "wqkv", il);
 
+                if (model.arch == LLM_ARCH_NOMIC_BERT_MOE) {
+                    cur = lm_ggml_add(ctx0, cur, model.layers[il].bqkv);
+                    cb(cur, "bqkv", il);
+                }
+
                 Qcur = lm_ggml_cont(ctx0, lm_ggml_view_2d(ctx0, cur, n_embd,     n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd)));
                 Kcur = lm_ggml_cont(ctx0, lm_ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd)));
                 Vcur = lm_ggml_cont(ctx0, lm_ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)));
@@ -5691,7 +5805,7 @@ struct llm_build_bert : public llm_graph_context {
 
             cur = build_attn(inp_attn, gf,
                     model.layers[il].wo, model.layers[il].bo,
-                    Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                    Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             cb(cur, "kqv_out", il);
 
             if (il == n_layer - 1 && pooling_type == LLAMA_POOLING_TYPE_NONE) {
@@ -5716,13 +5830,29 @@ struct llm_build_bert : public llm_graph_context {
             cb(ffn_inp, "ffn_inp", il);
 
             // feed-forward network
-            if (model.arch == LLM_ARCH_BERT) {
+            if (hparams.moe_every_n_layers > 0 && il % hparams.moe_every_n_layers == 1) {
+                // MoE branch
+                cur = build_moe_ffn(cur,
+                        model.layers[il].ffn_gate_inp,
+                        model.layers[il].ffn_up_exps,
+                        nullptr,
+                        model.layers[il].ffn_down_exps,
+                        nullptr,
+                        hparams.n_expert,
+                        hparams.n_expert_used,
+                        LLM_FFN_GELU,
+                        false, false,
+                        0.0f,
+                        LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, il);
+                cb(cur, "ffn_moe_out", il);
+            } else if (model.arch == LLM_ARCH_BERT || model.arch == LLM_ARCH_NOMIC_BERT_MOE) {
                 cur = build_ffn(cur,
                         model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
                         NULL,                      NULL,                        NULL,
                         model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
                         NULL,
                         LLM_FFN_GELU, LLM_FFN_SEQ, il);
+                cb(cur, "ffn_out", il);
             } else if (model.arch == LLM_ARCH_JINA_BERT_V2) {
                 cur = build_ffn(cur,
                         model.layers[il].ffn_up,   NULL,                        NULL,
@@ -5730,6 +5860,7 @@ struct llm_build_bert : public llm_graph_context {
                         model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
                         NULL,
                         LLM_FFN_GELU, LLM_FFN_PAR, il);
+                cb(cur, "ffn_out", il);
             } else {
                 cur = build_ffn(cur,
                         model.layers[il].ffn_up,   NULL, NULL,
@@ -5737,8 +5868,8 @@ struct llm_build_bert : public llm_graph_context {
                         model.layers[il].ffn_down, NULL, NULL,
                         NULL,
                         LLM_FFN_SILU, LLM_FFN_PAR, il);
+                cb(cur, "ffn_out", il);
             }
-            cb(cur, "ffn_out", il);
 
             // attentions bypass the intermediate layer
             cur = lm_ggml_add(ctx0, cur, ffn_inp);
@@ -5808,7 +5939,7 @@ struct llm_build_bloom : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -5949,7 +6080,7 @@ struct llm_build_mpt : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -6095,7 +6226,7 @@ struct llm_build_stablelm : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -6218,7 +6349,7 @@ struct llm_build_qwen : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -6338,7 +6469,7 @@ struct llm_build_qwen2 : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -6459,7 +6590,7 @@ struct llm_build_qwen2vl : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -6586,7 +6717,7 @@ struct llm_build_qwen2moe : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -6739,7 +6870,7 @@ struct llm_build_qwen3 : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -6860,7 +6991,7 @@ struct llm_build_qwen3moe : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -7000,7 +7131,7 @@ struct llm_build_phi2 : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f, il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f, il);
             }
 
             if (il == n_layer - 1) {
@@ -7075,7 +7206,7 @@ struct llm_build_phi3 : public llm_graph_context {
             // self-attention
             {
                 // rope freq factors for 128k context
-                lm_ggml_tensor * rope_factors = static_cast<const llama_kv_cache_unified *>(memory)->cbs.get_rope_factors(n_ctx_per_seq, il);
+                lm_ggml_tensor * rope_factors = model.get_rope_factors(n_ctx_per_seq, il);
 
                 lm_ggml_tensor* attn_norm_output = build_norm(inpL,
                         model.layers[il].attn_norm,
@@ -7129,7 +7260,7 @@ struct llm_build_phi3 : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f, il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f, il);
             }
 
             if (il == n_layer - 1) {
@@ -7264,7 +7395,7 @@ struct llm_build_plamo : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
             lm_ggml_tensor * sa_out = cur;
 
@@ -7371,7 +7502,7 @@ struct llm_build_gpt2 : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -7487,7 +7618,7 @@ struct llm_build_codeshell : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -7616,7 +7747,7 @@ struct llm_build_orion : public llm_graph_context {
 
             cur = build_attn(inp_attn, gf,
                     model.layers[il].wo, NULL,
-                    Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                    Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
         }
 
         if (il == n_layer - 1) {
@@ -7743,7 +7874,7 @@ struct llm_build_internlm2 : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -7827,7 +7958,7 @@ struct llm_build_minicpm3 : public llm_graph_context {
         for (int il = 0; il < n_layer; ++il) {
             lm_ggml_tensor * inpSA = inpL;
 
-            lm_ggml_tensor * rope_factors = static_cast<const llama_kv_cache_unified *>(memory)->cbs.get_rope_factors(n_ctx_per_seq, il);
+            lm_ggml_tensor * rope_factors = model.get_rope_factors(n_ctx_per_seq, il);
 
             // norm
             cur = build_norm(inpL,
@@ -7940,7 +8071,7 @@ struct llm_build_minicpm3 : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, NULL,
-                        q_states, k_states, v_states, nullptr, kq_scale, il);
+                        q_states, k_states, v_states, nullptr, nullptr, kq_scale, il);
             }
 
             if (il == n_layer - 1) {
@@ -8070,7 +8201,7 @@ struct llm_build_gemma : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f, il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f, il);
             }
 
             if (il == n_layer - 1) {
@@ -8192,7 +8323,7 @@ struct llm_build_gemma2 : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f, il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f, il);
             }
 
             cur = build_norm(cur,
@@ -8333,7 +8464,7 @@ struct llm_build_gemma3 : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, hparams.f_attention_scale, il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, hparams.f_attention_scale, il);
             }
 
             cur = build_norm(cur,
@@ -8473,7 +8604,7 @@ struct llm_build_starcoder2 : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -8594,7 +8725,7 @@ struct llm_build_mamba : public llm_graph_context {
              lm_ggml_tensor * state_mask,
       const llama_ubatch & ubatch,
                      int   il) const {
-        const llama_kv_cache_unified * kv_self = static_cast<const llama_kv_cache_unified *>(memory);
+        const llama_kv_cache_recurrent * kv_self = static_cast<const llama_kv_cache_recurrent *>(memory);
 
         const auto kv_head = kv_self->head;
 
@@ -8808,7 +8939,7 @@ struct llm_build_command_r : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -8895,7 +9026,7 @@ struct llm_build_cohere2 : public llm_graph_context {
             // self-attention
             {
                 // rope freq factors for 128k context
-                lm_ggml_tensor * rope_factors = static_cast<const llama_kv_cache_unified *>(memory)->cbs.get_rope_factors(n_ctx_per_seq, il);
+                lm_ggml_tensor * rope_factors = model.get_rope_factors(n_ctx_per_seq, il);
 
                 // compute Q and K and RoPE them
                 lm_ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
@@ -8943,7 +9074,7 @@ struct llm_build_cohere2 : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -9074,7 +9205,7 @@ struct llm_build_olmo : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, nullptr,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -9194,7 +9325,7 @@ struct llm_build_olmo2 : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             cur = build_norm(cur,
@@ -9327,7 +9458,7 @@ struct llm_build_olmoe : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -9460,7 +9591,7 @@ struct llm_build_openelm : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -9574,7 +9705,7 @@ struct llm_build_gptneox : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -9724,7 +9855,7 @@ struct llm_build_arctic : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -9833,7 +9964,7 @@ struct llm_build_deepseek : public llm_graph_context {
             // self-attention
             {
                 // rope freq factors for llama3; may return nullptr for llama2 and other models
-                lm_ggml_tensor * rope_factors = static_cast<const llama_kv_cache_unified *>(memory)->cbs.get_rope_factors(n_ctx_per_seq, il);
+                lm_ggml_tensor * rope_factors = model.get_rope_factors(n_ctx_per_seq, il);
 
                 // compute Q and K and RoPE them
                 lm_ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
@@ -9879,7 +10010,7 @@ struct llm_build_deepseek : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, kq_scale, il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, kq_scale, il);
             }
 
             if (il == n_layer - 1) {
@@ -9969,15 +10100,22 @@ struct llm_build_deepseek2 : public llm_graph_context {
     llm_build_deepseek2(const llama_model & model, const llm_graph_params & params, lm_ggml_cgraph * gf) : llm_graph_context(params) {
         bool is_lite = (hparams.n_layer == 27);
 
+        const bool is_mla = (hparams.n_embd_head_k_mla != 0 && hparams.n_embd_head_v_mla != 0);
+
+        // note: these are the actual head sizes you get when treating as MHA or after "decompression" using wv_b for MLA
+        const int64_t n_embd_head_k = is_mla ? hparams.n_embd_head_k_mla : hparams.n_embd_head_k;
+        const int64_t n_embd_head_v = is_mla ? hparams.n_embd_head_v_mla : hparams.n_embd_head_v;
+
+        const int64_t n_embd_head_qk_rope = hparams.n_rot;
+        const int64_t n_embd_head_qk_nope = n_embd_head_k - n_embd_head_qk_rope;
+
+        const uint32_t kv_lora_rank = hparams.n_lora_kv;
+
         // We have to pre-scale kq_scale and attn_factor to make the YaRN RoPE work correctly.
         // See https://github.com/ggerganov/llama.cpp/discussions/7416 for detailed explanation.
         const float mscale = attn_factor * (1.0f + hparams.rope_yarn_log_mul * logf(1.0f / freq_scale));
-        const float kq_scale = 1.0f*mscale*mscale/sqrtf(float(hparams.n_embd_head_k));
-        const float attn_factor_scaled = 1.0f / (1.0f + 0.1f * logf(1.0f / freq_scale));
-
-        const uint32_t n_embd_head_qk_rope = hparams.n_rot;
-        const uint32_t n_embd_head_qk_nope = hparams.n_embd_head_k - hparams.n_rot;
-        const uint32_t kv_lora_rank = hparams.n_lora_kv;
+        const float kq_scale = 1.0f*mscale*mscale/sqrtf(float(n_embd_head_k));
+        const float attn_factor = 1.0f / (1.0f + 0.1f * logf(1.0f / freq_scale));
 
         lm_ggml_tensor * cur;
         lm_ggml_tensor * inpL;
@@ -10003,16 +10141,14 @@ struct llm_build_deepseek2 : public llm_graph_context {
             {
                 lm_ggml_tensor * q = NULL;
                 if (!is_lite) {
-                    // {n_embd, q_lora_rank} * {n_embd, n_tokens} -> {q_lora_rank, n_tokens}
                     q = lm_ggml_mul_mat(ctx0, model.layers[il].wq_a, cur);
                     cb(q, "q", il);
 
                     q = build_norm(q,
-                            model.layers[il].attn_q_a_norm, NULL,
+                            model.layers[il].attn_q_a_norm, nullptr,
                             LLM_NORM_RMS, il);
                     cb(q, "q", il);
 
-                    // {q_lora_rank, n_head * hparams.n_embd_head_k} * {q_lora_rank, n_tokens} -> {n_head * hparams.n_embd_head_k, n_tokens}
                     q = lm_ggml_mul_mat(ctx0, model.layers[il].wq_b, q);
                     cb(q, "q", il);
                 } else {
@@ -10020,96 +10156,125 @@ struct llm_build_deepseek2 : public llm_graph_context {
                     cb(q, "q", il);
                 }
 
-                // split into {n_head * n_embd_head_qk_nope, n_tokens}
-                lm_ggml_tensor * q_nope = lm_ggml_view_3d(ctx0, q, n_embd_head_qk_nope, n_head, n_tokens,
-                        lm_ggml_row_size(q->type, hparams.n_embd_head_k),
-                        lm_ggml_row_size(q->type, hparams.n_embd_head_k * n_head),
+                // split into {n_embd_head_qk_nope, n_head, n_tokens}
+                lm_ggml_tensor * q_nope = lm_ggml_view_3d(ctx0, q,
+                        n_embd_head_qk_nope, n_head, n_tokens,
+                        lm_ggml_row_size(q->type, n_embd_head_k),
+                        lm_ggml_row_size(q->type, n_embd_head_k) * n_head,
                         0);
                 cb(q_nope, "q_nope", il);
 
-                // and {n_head * n_embd_head_qk_rope, n_tokens}
-                lm_ggml_tensor * q_pe = lm_ggml_view_3d(ctx0, q, n_embd_head_qk_rope, n_head, n_tokens,
-                        lm_ggml_row_size(q->type, hparams.n_embd_head_k),
-                        lm_ggml_row_size(q->type, hparams.n_embd_head_k * n_head),
+                // and {n_embd_head_qk_rope, n_head, n_tokens}
+                lm_ggml_tensor * q_pe = lm_ggml_view_3d(ctx0, q,
+                        n_embd_head_qk_rope, n_head, n_tokens,
+                        lm_ggml_row_size(q->type, n_embd_head_k),
+                        lm_ggml_row_size(q->type, n_embd_head_k) * n_head,
                         lm_ggml_row_size(q->type, n_embd_head_qk_nope));
                 cb(q_pe, "q_pe", il);
 
-                // {n_embd, kv_lora_rank + n_embd_head_qk_rope} * {n_embd, n_tokens} -> {kv_lora_rank + n_embd_head_qk_rope, n_tokens}
-                lm_ggml_tensor * kv_pe_compresseed = lm_ggml_mul_mat(ctx0, model.layers[il].wkv_a_mqa, cur);
-                cb(kv_pe_compresseed, "kv_pe_compresseed", il);
+                lm_ggml_tensor * kv_cmpr_pe = lm_ggml_mul_mat(ctx0, model.layers[il].wkv_a_mqa, cur);
+                cb(kv_cmpr_pe, "kv_cmpr_pe", il);
 
                 // split into {kv_lora_rank, n_tokens}
-                lm_ggml_tensor * kv_compressed = lm_ggml_view_2d(ctx0, kv_pe_compresseed, kv_lora_rank, n_tokens,
-                        kv_pe_compresseed->nb[1],
+                lm_ggml_tensor * kv_cmpr = lm_ggml_view_2d(ctx0, kv_cmpr_pe,
+                        kv_lora_rank, n_tokens,
+                        lm_ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope),
                         0);
-                cb(kv_compressed, "kv_compressed", il);
+                cb(kv_cmpr, "kv_cmpr", il);
+
+                // and {n_embd_head_qk_rope, 1, n_tokens}
+                lm_ggml_tensor * k_pe = lm_ggml_view_3d(ctx0, kv_cmpr_pe,
+                        n_embd_head_qk_rope, 1, n_tokens,
+                        lm_ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope),
+                        lm_ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope),
+                        lm_ggml_row_size(kv_cmpr_pe->type, kv_lora_rank));
+                cb(k_pe, "k_pe", il);
 
-                // and {n_embd_head_qk_rope, n_tokens}
-                lm_ggml_tensor * k_pe = lm_ggml_view_3d(ctx0, kv_pe_compresseed, n_embd_head_qk_rope, 1, n_tokens,
-                        kv_pe_compresseed->nb[1],
-                        kv_pe_compresseed->nb[1],
-                        lm_ggml_row_size(kv_pe_compresseed->type, kv_lora_rank));
+                q_pe = lm_ggml_rope_ext(ctx0, q_pe, inp_pos, nullptr,
+                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                        ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(q_pe, "q_pe", il);
+
+                k_pe = lm_ggml_rope_ext(ctx0, k_pe, inp_pos, nullptr,
+                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                        ext_factor, attn_factor, beta_fast, beta_slow
+                );
                 cb(k_pe, "k_pe", il);
 
-                // TODO: the CUDA backend used to not support non-cont. (RMS) norm, investigate removing lm_ggml_cont
-                kv_compressed = lm_ggml_cont(ctx0, kv_compressed);
-                kv_compressed = build_norm(kv_compressed,
-                        model.layers[il].attn_kv_a_norm, NULL,
+                kv_cmpr = build_norm(kv_cmpr,
+                        model.layers[il].attn_kv_a_norm, nullptr,
                         LLM_NORM_RMS, il);
-                cb(kv_compressed, "kv_compressed", il);
+                cb(kv_cmpr, "kv_cmpr", il);
 
-                // {kv_lora_rank, n_head * (n_embd_head_qk_nope + n_embd_head_v)} * {kv_lora_rank, n_tokens} -> {n_head * (n_embd_head_qk_nope + n_embd_head_v), n_tokens}
-                lm_ggml_tensor * kv = lm_ggml_mul_mat(ctx0, model.layers[il].wkv_b, kv_compressed);
-                cb(kv, "kv", il);
+                if (is_mla) {
+                    // {n_embd_head_qk_nope, n_tokens, n_head}
+                    q_nope = lm_ggml_permute(ctx0, q_nope, 0, 2, 1, 3);
+                    cb(q_nope, "q_nope_perm", il);
 
-                // split into {n_head * n_embd_head_qk_nope, n_tokens}
-                lm_ggml_tensor * k_nope = lm_ggml_view_3d(ctx0, kv, n_embd_head_qk_nope, n_head, n_tokens,
-                        lm_ggml_row_size(kv->type, n_embd_head_qk_nope + hparams.n_embd_head_v),
-                        lm_ggml_row_size(kv->type, n_head * (n_embd_head_qk_nope + hparams.n_embd_head_v)),
-                        0);
-                cb(k_nope, "k_nope", il);
+                    // {n_embd_head_qk_nope, kv_lora_rank, n_head} x {n_embd_head_qk_nope, n_tokens, n_head}
+                    lm_ggml_tensor * q_nope_absorbed = lm_ggml_mul_mat(ctx0, model.layers[il].wk_b, q_nope);
+                    cb(q_nope_absorbed, "q_nope_absorbed", il);
 
-                // and {n_head * n_embd_head_v, n_tokens}
-                lm_ggml_tensor * v_states = lm_ggml_view_3d(ctx0, kv, hparams.n_embd_head_v, n_head, n_tokens,
-                        lm_ggml_row_size(kv->type, (n_embd_head_qk_nope + hparams.n_embd_head_v)),
-                        lm_ggml_row_size(kv->type, (n_embd_head_qk_nope + hparams.n_embd_head_v)*n_head),
-                        lm_ggml_row_size(kv->type, (n_embd_head_qk_nope)));
-                cb(v_states, "v_states", il);
+                    // {kv_lora_rank, n_head, n_tokens}
+                    q_nope_absorbed = lm_ggml_permute(ctx0, q_nope_absorbed, 0, 2, 1, 3);
+                    cb(q_nope_absorbed, "q_nope_absorbed_perm", il);
 
-                v_states = lm_ggml_cont(ctx0, v_states);
-                cb(v_states, "v_states", il);
+                    // {n_embd_head_qk_rope + kv_lora_rank, n_head, n_tokens}
+                    // note: rope must go first for in-place context shifting in build_rope_shift()
+                    lm_ggml_tensor * Qcur = lm_ggml_concat(ctx0, q_pe, q_nope_absorbed, 0);
+                    cb(Qcur, "Qcur", il);
 
-                v_states = lm_ggml_view_2d(ctx0, v_states, hparams.n_embd_head_v * n_head, n_tokens,
-                        lm_ggml_row_size(kv->type, hparams.n_embd_head_v * n_head),
-                        0);
-                cb(v_states, "v_states", il);
+                    kv_cmpr = lm_ggml_reshape_3d(ctx0, kv_cmpr, kv_lora_rank, 1, n_tokens);
+                    cb(kv_cmpr, "kv_cmpr_reshape", il);
 
-                q_pe = lm_ggml_cont(ctx0, q_pe); // TODO: the CUDA backend used to not support non-cont. RoPE, investigate removing this
-                q_pe = lm_ggml_rope_ext(
-                        ctx0, q_pe, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor_scaled, beta_fast, beta_slow
-                        );
-                cb(q_pe, "q_pe", il);
+                    // {n_embd_head_qk_rope + kv_lora_rank, 1, n_tokens}
+                    lm_ggml_tensor * Kcur = lm_ggml_concat(ctx0, k_pe, kv_cmpr, 0);
+                    cb(Kcur, "Kcur", il);
 
-                // shared RoPE key
-                k_pe = lm_ggml_cont(ctx0, k_pe); // TODO: the CUDA backend used to not support non-cont. RoPE, investigate removing this
-                k_pe = lm_ggml_rope_ext(
-                        ctx0, k_pe, inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor_scaled, beta_fast, beta_slow
-                        );
-                cb(k_pe, "k_pe", il);
+                    // {kv_lora_rank, 1, n_tokens}
+                    lm_ggml_tensor * Vcur = kv_cmpr;
+                    cb(Vcur, "Vcur", il);
 
-                lm_ggml_tensor * q_states = lm_ggml_concat(ctx0, q_nope, q_pe, 0);
-                cb(q_states, "q_states", il);
+                    // note: MLA with the absorption optimzation converts into MQA (ie: GQA with 1 group)
+                    cur = build_attn(inp_attn, gf,
+                            model.layers[il].wo, NULL,
+                            Qcur, Kcur, Vcur, nullptr, model.layers[il].wv_b, kq_scale, il);
+                } else {
+                    lm_ggml_tensor * kv = lm_ggml_mul_mat(ctx0, model.layers[il].wkv_b, kv_cmpr);
+                    cb(kv, "kv", il);
+
+                    // split into {n_embd_head_qk_nope, n_head, n_tokens}
+                    lm_ggml_tensor * k_nope = lm_ggml_view_3d(ctx0, kv,
+                            n_embd_head_qk_nope, n_head, n_tokens,
+                            lm_ggml_row_size(kv->type, n_embd_head_qk_nope + n_embd_head_v),
+                            lm_ggml_row_size(kv->type, n_embd_head_qk_nope + n_embd_head_v) * n_head,
+                            0);
+                    cb(k_nope, "k_nope_view", il);
 
-                lm_ggml_tensor * k_states = lm_ggml_concat(ctx0, k_nope, lm_ggml_repeat(ctx0, k_pe, q_pe), 0);
-                cb(k_states, "k_states", il);
+                    // and {n_embd_head_v, n_head, n_tokens}
+                    lm_ggml_tensor * Vcur = lm_ggml_view_3d(ctx0, kv,
+                            n_embd_head_v, n_head, n_tokens,
+                            lm_ggml_row_size(kv->type, n_embd_head_qk_nope + n_embd_head_v),
+                            lm_ggml_row_size(kv->type, n_embd_head_qk_nope + n_embd_head_v) * n_head,
+                            lm_ggml_row_size(kv->type, n_embd_head_qk_nope));
+                    cb(Vcur, "Vcur_view", il);
 
-                cur = build_attn(inp_attn, gf,
-                        model.layers[il].wo, NULL,
-                        q_states, k_states, v_states, nullptr, kq_scale, il);
+                    Vcur = lm_ggml_cont(ctx0, Vcur);
+                    cb(Vcur, "Vcur_cont", il);
+
+                    // note: rope must go first for in-place context shifting in build_rope_shift()
+                    lm_ggml_tensor * Qcur = lm_ggml_concat(ctx0, q_pe, q_nope, 0);
+                    cb(Qcur, "Qcur", il);
+
+                    lm_ggml_tensor * Kcur = lm_ggml_concat(ctx0, lm_ggml_repeat(ctx0, k_pe, q_pe), k_nope, 0);
+                    cb(Kcur, "Kcur", il);
+
+                    // note: MLA without the absorption optimization converts into MHA (ie: GQA with full n_head groups)
+                    cur = build_attn(inp_attn, gf,
+                            model.layers[il].wo, NULL,
+                            Qcur, Kcur, Vcur, nullptr, nullptr, kq_scale, il);
+                }
             }
 
             if (il == n_layer - 1) {
@@ -10275,7 +10440,7 @@ struct llm_build_bitnet : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         NULL, NULL,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
 
                 cur = build_norm(cur,
                         model.layers[il].attn_sub_norm, NULL,
@@ -10398,7 +10563,7 @@ struct llm_build_t5_enc : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo_enc, nullptr,
-                        Qcur, Kcur, Vcur, kq_b, 1.0f, il);
+                        Qcur, Kcur, Vcur, kq_b, nullptr, 1.0f, il);
                 cb(cur, "kqv_out", il);
             }
 
@@ -10504,7 +10669,7 @@ struct llm_build_t5_dec : public llm_graph_context {
 
                 cur = build_attn(inp_attn_self, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, kq_b, 1.0f, il);
+                        Qcur, Kcur, Vcur, kq_b, nullptr, 1.0f, il);
                 cb(cur, "kqv_out", il);
             }
 
@@ -10536,7 +10701,7 @@ struct llm_build_t5_dec : public llm_graph_context {
 
                 cur = build_attn(inp_attn_cross, gf,
                         model.layers[il].wo_cross, nullptr,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f, il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f, il);
                 cb(cur, "kqv_out", il);
 
                 //lm_ggml_tensor * q =                 lm_ggml_permute(ctx0, Qcur, 0, 2, 1, 3);
@@ -10669,7 +10834,7 @@ struct llm_build_jais : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/float(n_embd_head), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/float(n_embd_head), il);
             }
 
             if (il == n_layer - 1) {
@@ -10801,7 +10966,7 @@ struct llm_build_chatglm : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -10854,6 +11019,157 @@ struct llm_build_chatglm : public llm_graph_context {
     }
 };
 
+struct llm_build_glm4 : public llm_graph_context {
+    llm_build_glm4(const llama_model & model, const llm_graph_params & params, lm_ggml_cgraph * gf) : llm_graph_context(params) {
+        const int64_t n_embd_head = hparams.n_embd_head_v;
+        const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
+
+        LM_GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+        lm_ggml_tensor * cur;
+        lm_ggml_tensor * inpL;
+
+        inpL = build_inp_embd(model.tok_embd);
+
+        // inp_pos - contains the positions
+        lm_ggml_tensor * inp_pos = build_inp_pos();
+
+        auto * inp_attn = build_attn_inp_kv_unified();
+
+        for (int il = 0; il < n_layer; ++il) {
+            lm_ggml_tensor * inpSA = inpL;
+
+            // Pre-attention norm
+            cur = build_norm(inpL,
+                    model.layers[il].attn_norm,
+                    NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "attn_norm", il);
+
+            // self-attention
+            {
+                lm_ggml_tensor * Qcur = nullptr;
+                lm_ggml_tensor * Kcur = nullptr;
+                lm_ggml_tensor * Vcur = nullptr;
+
+                if (model.layers[il].wqkv == nullptr) {
+                    Qcur = build_lora_mm(model.layers[il].wq, cur);
+                    if (model.layers[il].bq) {
+                        Qcur = lm_ggml_add(ctx0, Qcur, model.layers[il].bq);
+                    }
+                    Kcur = build_lora_mm(model.layers[il].wk, cur);
+                    if (model.layers[il].bk) {
+                        Kcur = lm_ggml_add(ctx0, Kcur, model.layers[il].bk);
+                    }
+                    Vcur = build_lora_mm(model.layers[il].wv, cur);
+                    if (model.layers[il].bv) {
+                        Vcur = lm_ggml_add(ctx0, Vcur, model.layers[il].bv);
+                    }
+                } else {
+                    cur = build_lora_mm(model.layers[il].wqkv, cur);
+                    cb(cur, "wqkv", il);
+                    if (model.layers[il].bqkv) {
+                        cur = lm_ggml_add(ctx0, cur, model.layers[il].bqkv);
+                        cb(cur, "bqkv", il);
+                    }
+                    Qcur = lm_ggml_cont(ctx0, lm_ggml_view_2d(ctx0, cur, n_embd,     n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd)));
+                    Kcur = lm_ggml_cont(ctx0, lm_ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd)));
+                    Vcur = lm_ggml_cont(ctx0, lm_ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)));
+                }
+
+                Qcur = lm_ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+                Kcur = lm_ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+                Vcur = lm_ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+                Qcur = lm_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 = lm_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, gf,
+                        model.layers[il].wo, NULL,
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+            }
+
+            if (il == n_layer - 1) {
+                // skip computing output for unused tokens
+                lm_ggml_tensor * inp_out_ids = build_inp_out_ids();
+                cur   = lm_ggml_get_rows(ctx0,   cur, inp_out_ids);
+                inpSA = lm_ggml_get_rows(ctx0, inpSA, inp_out_ids);
+            }
+
+            // Post-attention norm (new!)
+            cur = build_norm(cur,
+                    model.layers[il].attn_post_norm,
+                    NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "post_attn_norm", il);
+
+            // Add the input (residual connection after post-attention norm)
+            lm_ggml_tensor * ffn_inp = lm_ggml_add(ctx0, cur, inpSA);
+            cb(ffn_inp, "ffn_inp", il);
+
+            // FF
+            {
+                // Pre-MLP norm
+                cur = build_norm(ffn_inp,
+                        model.layers[il].ffn_norm,
+                        NULL,
+                        LLM_NORM_RMS, il);
+                cb(cur, "ffn_norm", il);
+
+                // MLP
+                cur = build_ffn(cur,
+                        model.layers[il].ffn_up,   NULL, NULL,
+                        NULL,                      NULL, NULL,
+                        model.layers[il].ffn_down, NULL, NULL,
+                        NULL,
+                        LLM_FFN_SWIGLU, LLM_FFN_SEQ, il);
+                cb(cur, "ffn_out", il);
+
+                // Post-MLP norm
+                cur = build_norm(cur,
+                        model.layers[il].ffn_post_norm,
+                        NULL,
+                        LLM_NORM_RMS, il);
+                cb(cur, "post_mlp_norm", il);
+            }
+
+            // Add residual connection after post-MLP norm
+            inpL = lm_ggml_add(ctx0, cur, ffn_inp);
+            cb(inpL, "l_out", il);
+        }
+
+        // Final norm
+        cur = build_norm(inpL,
+                model.output_norm,
+                NULL,
+                LLM_NORM_RMS, -1);
+
+        cb(cur, "result_norm", -1);
+        res->t_embd = cur;
+
+        // Output projection
+        cur = build_lora_mm(model.output, cur);
+
+        cb(cur, "result_output", -1);
+        res->t_logits = cur;
+
+        lm_ggml_build_forward_expand(gf, cur);
+    }
+};
+
 struct llm_build_nemotron : public llm_graph_context {
     llm_build_nemotron(const llama_model & model, const llm_graph_params & params, lm_ggml_cgraph * gf) : llm_graph_context(params) {
         const int64_t n_embd_head = hparams.n_embd_head_v;
@@ -10927,7 +11243,7 @@ struct llm_build_nemotron : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -11012,7 +11328,7 @@ struct llm_build_exaone : public llm_graph_context {
             // self-attention
             {
                 // rope freq factors for llama3; may return nullptr for llama2 and other models
-                lm_ggml_tensor * rope_factors = static_cast<const llama_kv_cache_unified *>(memory)->cbs.get_rope_factors(n_ctx_per_seq, il);
+                lm_ggml_tensor * rope_factors = model.get_rope_factors(n_ctx_per_seq, il);
 
                 // compute Q and K and RoPE them
                 lm_ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
@@ -11058,7 +11374,7 @@ struct llm_build_exaone : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -11157,7 +11473,7 @@ struct llm_build_rwkv6_base : public llm_graph_context {
             lm_ggml_tensor * state_mask,
             const llama_ubatch & ubatch,
             int   il) const {
-        const llama_kv_cache_unified * kv_self = static_cast<const llama_kv_cache_unified *>(memory);
+        const llama_kv_cache_recurrent * kv_self = static_cast<const llama_kv_cache_recurrent *>(memory);
 
         const auto n_tokens = ubatch.n_tokens;
         const auto n_seqs = ubatch.n_seqs;
@@ -11553,7 +11869,7 @@ struct llm_build_rwkv7_base : public llm_graph_context {
             lm_ggml_tensor *& first_layer_value,
             const llama_ubatch & ubatch,
             int   il) const {
-        const llama_kv_cache_unified * kv_self = static_cast<const llama_kv_cache_unified *>(memory);
+        const llama_kv_cache_recurrent * kv_self = static_cast<const llama_kv_cache_recurrent *>(memory);
 
         const auto n_tokens = ubatch.n_tokens;
         const auto n_seqs = ubatch.n_seqs;
@@ -11960,7 +12276,7 @@ struct llm_build_chameleon : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, nullptr,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
 
                 if (hparams.swin_norm) {
                     cur = build_norm(cur,
@@ -12316,7 +12632,7 @@ struct llm_build_plm : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, NULL,
-                        q_states, k_states, v_states, nullptr, kq_scale, il);
+                        q_states, k_states, v_states, nullptr, nullptr, kq_scale, il);
             }
 
             if (il == n_layer - 1) {
@@ -12393,7 +12709,7 @@ struct llm_build_bailingmoe : public llm_graph_context {
             // self-attention
             {
                 // rope freq factors for llama3; may return nullptr for llama2 and other models
-                lm_ggml_tensor * rope_factors = static_cast<const llama_kv_cache_unified *>(memory)->cbs.get_rope_factors(n_ctx_per_seq, il);
+                lm_ggml_tensor * rope_factors = model.get_rope_factors(n_ctx_per_seq, il);
 
                 // compute Q and K and RoPE them
                 lm_ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
@@ -12439,7 +12755,7 @@ struct llm_build_bailingmoe : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_rot)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_rot)), il);
             }
 
             if (il == n_layer - 1) {
@@ -12513,7 +12829,7 @@ struct llm_build_bailingmoe : public llm_graph_context {
     }
 };
 
-llama_memory_i * llama_model::create_memory() const {
+llama_memory_i * llama_model::create_memory(const llama_memory_params & params, llama_cparams & cparams) const {
     llama_memory_i * res;
 
     switch (arch) {
@@ -12523,26 +12839,29 @@ llama_memory_i * llama_model::create_memory() const {
         case LLM_ARCH_RWKV7:
         case LLM_ARCH_ARWKV7:
             {
-                res = new llama_kv_cache_unified(hparams, {
-                    /*.get_rope_factors =*/ nullptr
-                });
+                res = new llama_kv_cache_recurrent(
+                        *this,
+                        LM_GGML_TYPE_F32,
+                        LM_GGML_TYPE_F32,
+                        cparams.offload_kqv,
+                        std::max((uint32_t) 1, cparams.n_seq_max));
             } break;
         default:
             {
-                res = new llama_kv_cache_unified(hparams, {
-                    /*.get_rope_factors =*/ [this](uint32_t n_ctx_per_seq, int il) {
-                        // choose long/short freq factors based on the context size
-                        if (layers[il].rope_freqs != nullptr) {
-                            return layers[il].rope_freqs;
-                        }
+                const auto padding = llama_kv_cache_unified::get_padding(cparams);
 
-                        if (n_ctx_per_seq > hparams.n_ctx_orig_yarn) {
-                            return layers[il].rope_long;
-                        }
+                cparams.n_ctx = LM_GGML_PAD(cparams.n_ctx, padding);
 
-                        return layers[il].rope_short;
-                    }
-                });
+                LLAMA_LOG_DEBUG("%s: n_ctx = %u (padded)\n", __func__, cparams.n_ctx);
+
+                res = new llama_kv_cache_unified(
+                        *this,
+                        params.type_k,
+                        params.type_v,
+                        !cparams.flash_attn,
+                        cparams.offload_kqv,
+                        cparams.n_ctx,
+                        padding);
             }
     }
 
@@ -12591,6 +12910,7 @@ llm_graph_result_ptr llama_model::build_graph(
         case LLM_ARCH_BERT:
         case LLM_ARCH_JINA_BERT_V2:
         case LLM_ARCH_NOMIC_BERT:
+        case LLM_ARCH_NOMIC_BERT_MOE:
             {
                 llm = std::make_unique<llm_build_bert>(*this, params, gf);
             } break;
@@ -12735,6 +13055,10 @@ llm_graph_result_ptr llama_model::build_graph(
             {
                 llm = std::make_unique<llm_build_chatglm>(*this, params, gf);
             } break;
+        case LLM_ARCH_GLM4:
+            {
+                llm = std::make_unique<llm_build_glm4>(*this, params, gf);
+            } break;
         case LLM_ARCH_BITNET:
             {
                 llm = std::make_unique<llm_build_bitnet>(*this, params, gf);
@@ -12919,8 +13243,6 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
         case LLM_ARCH_DECI:
         case LLM_ARCH_BAICHUAN:
         case LLM_ARCH_STARCODER:
-        case LLM_ARCH_PLAMO:
-        case LLM_ARCH_ORION:
         case LLM_ARCH_INTERNLM2:
         case LLM_ARCH_MINICPM:
         case LLM_ARCH_XVERSE:
@@ -12932,6 +13254,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
         case LLM_ARCH_DEEPSEEK2:
         case LLM_ARCH_PLM:
         case LLM_ARCH_CHATGLM:
+        case LLM_ARCH_GLM4:
         case LLM_ARCH_GRANITE:
         case LLM_ARCH_GRANITE_MOE:
         case LLM_ARCH_CHAMELEON:
@@ -12944,6 +13267,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
         case LLM_ARCH_DBRX:
         case LLM_ARCH_BERT:
         case LLM_ARCH_NOMIC_BERT:
+        case LLM_ARCH_NOMIC_BERT_MOE:
         case LLM_ARCH_STABLELM:
         case LLM_ARCH_BITNET:
         case LLM_ARCH_QWEN:
@@ -12956,6 +13280,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
         case LLM_ARCH_PHI2:
         case LLM_ARCH_PHI3:
         case LLM_ARCH_PHIMOE:
+        case LLM_ARCH_PLAMO:
         case LLM_ARCH_GEMMA:
         case LLM_ARCH_GEMMA2:
         case LLM_ARCH_GEMMA3:
@@ -12963,6 +13288,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
         case LLM_ARCH_OPENELM:
         case LLM_ARCH_GPTNEOX:
         case LLM_ARCH_CODESHELL:
+        case LLM_ARCH_ORION:
         case LLM_ARCH_NEMOTRON:
         case LLM_ARCH_EXAONE:
         case LLM_ARCH_MINICPM3: