Merge branch 'main' into drifkin/array-head-count-simple

llama/llama.cpp/src/llama-chat.h

@@ -29,8 +29,8 @@ enum llm_chat_template {
     LLM_CHAT_TEMPLATE_DEEPSEEK_3,
     LLM_CHAT_TEMPLATE_COMMAND_R,
     LLM_CHAT_TEMPLATE_LLAMA_3,
-    LLM_CHAT_TEMPLATE_CHATGML_3,
-    LLM_CHAT_TEMPLATE_CHATGML_4,
+    LLM_CHAT_TEMPLATE_CHATGLM_3,
+    LLM_CHAT_TEMPLATE_CHATGLM_4,
     LLM_CHAT_TEMPLATE_GLMEDGE,
     LLM_CHAT_TEMPLATE_MINICPM,
     LLM_CHAT_TEMPLATE_EXAONE_3,
@@ -41,6 +41,7 @@ enum llm_chat_template {
     LLM_CHAT_TEMPLATE_YANDEX,
     LLM_CHAT_TEMPLATE_BAILING,
     LLM_CHAT_TEMPLATE_LLAMA4,
+    LLM_CHAT_TEMPLATE_SMOLVLM,
     LLM_CHAT_TEMPLATE_UNKNOWN,
 };
 

llama/llama.cpp/src/llama-context.cpp

@@ -114,7 +114,7 @@ llama_context::llama_context(
     }
 
     if (n_ctx_per_seq > hparams.n_ctx_train) {
-        LLAMA_LOG_WARN("%s: n_ctx_pre_seq (%u) > n_ctx_train (%u) -- possible training context overflow\n",
+        LLAMA_LOG_WARN("%s: n_ctx_per_seq (%u) > n_ctx_train (%u) -- possible training context overflow\n",
                 __func__, n_ctx_per_seq, hparams.n_ctx_train);
     }
 
@@ -469,8 +469,7 @@ ggml_tensor * llama_context::build_rope_shift(
         ggml_tensor * shift,
         ggml_tensor * factors,
               float   freq_base,
-              float   freq_scale,
-        ggml_backend_buffer * bbuf) const {
+              float   freq_scale) const {
     const auto & n_ctx_orig = cparams.n_ctx_orig_yarn;
 
     const auto & yarn_ext_factor  = cparams.yarn_ext_factor;
@@ -492,17 +491,7 @@ ggml_tensor * llama_context::build_rope_shift(
         // dequantize to f32 -> RoPE -> quantize back
         tmp = ggml_cast(ctx0, cur, GGML_TYPE_F32);
 
-        if (bbuf) {
-            for (const auto & backend : backends) {
-                // Figure out which backend KV cache belongs to
-                if (ggml_backend_supports_buft(backend.get(), ggml_backend_buffer_get_type(bbuf))) {
-                    ggml_backend_sched_set_tensor_backend(sched.get(), tmp, backend.get());
-                    break;
-                }
-            }
-        }
-
-        tmp = ggml_rope_ext_inplace(ctx0, tmp,
+        tmp = ggml_rope_ext(ctx0, tmp,
                 shift, factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
                 yarn_ext_factor, yarn_attn_factor, yarn_beta_fast, yarn_beta_slow);
 
@@ -582,7 +571,7 @@ llm_graph_result_ptr llama_context::build_kv_self_shift(
                 ggml_row_size(kv_self->k_l[il]->type, n_embd_k_gqa),
                 0);
 
-        ggml_tensor * cur = build_rope_shift(ctx0, k, inp->k_shift, rope_factors, freq_base_l, freq_scale_l, kv_self->k_l[il]->buffer);
+        ggml_tensor * cur = build_rope_shift(ctx0, k, inp->k_shift, rope_factors, freq_base_l, freq_scale_l);
 
         ggml_build_forward_expand(gf, cur);
     }
@@ -1510,8 +1499,6 @@ int32_t llama_context::output_reserve(int32_t n_outputs) {
     // set all ids as invalid (negative)
     std::fill(output_ids.begin(), output_ids.end(), -1);
 
-    ggml_backend_buffer_clear(buf_output.get(), 0);
-
     this->n_outputs     = 0;
     this->n_outputs_max = n_outputs_max;
 

llama/llama.cpp/src/llama-context.h

@@ -172,8 +172,7 @@ private:
         ggml_tensor * shift,
         ggml_tensor * factors,
               float   freq_base,
-              float   freq_scale,
-        ggml_backend_buffer * bbuf) const;
+              float   freq_scale) const;
 
     llm_graph_result_ptr build_kv_self_shift(
             ggml_context * ctx0,

llama/llama.cpp/src/llama-graph.cpp

@@ -55,7 +55,21 @@ void llm_graph_input_pos::set_input(const llama_ubatch * ubatch) {
     if (ubatch->pos && pos) {
         const int64_t n_tokens = ubatch->n_tokens;
 
-        ggml_backend_tensor_set(pos, ubatch->pos, 0, n_tokens*n_pos_per_token*ggml_element_size(pos));
+        if (ubatch->token && n_pos_per_embd == 4) {
+            // in case we're using M-RoPE with text tokens, convert the 1D positions to 4D
+            // the 3 first dims are the same, and 4th dim is all 0
+            std::vector<llama_pos> pos_data(n_tokens*n_pos_per_embd);
+            // copy the first dimension
+            for (int i = 0; i < n_tokens; ++i) {
+                pos_data[               i] = ubatch->pos[i];
+                pos_data[    n_tokens + i] = ubatch->pos[i];
+                pos_data[2 * n_tokens + i] = ubatch->pos[i];
+                pos_data[3 * n_tokens + i] = 0; // 4th dim is 0
+            }
+            ggml_backend_tensor_set(pos, pos_data.data(), 0, pos_data.size()*ggml_element_size(pos));
+        } else {
+            ggml_backend_tensor_set(pos, ubatch->pos, 0, n_tokens*n_pos_per_embd*ggml_element_size(pos));
+        }
     }
 }
 
@@ -71,7 +85,7 @@ void llm_graph_input_attn_temp::set_input(const llama_ubatch * ubatch) {
             ) * f_attn_temp_scale + 1.0;
         }
 
-        ggml_backend_tensor_set(attn_scale, attn_scale_data.data(), 0, n_tokens*n_pos_per_token*ggml_element_size(attn_scale));
+        ggml_backend_tensor_set(attn_scale, attn_scale_data.data(), 0, n_tokens*ggml_element_size(attn_scale));
     }
 }
 
@@ -598,7 +612,7 @@ llm_graph_context::llm_graph_context(const llm_graph_params & params) :
     res              (std::make_unique<llm_graph_result>()) {
     }
 
-int64_t llm_graph_context::n_pos_per_token() const {
+int64_t llm_graph_context::n_pos_per_embd() const {
     return arch == LLM_ARCH_QWEN2VL ? 4 : 1;
 }
 
@@ -809,6 +823,10 @@ ggml_tensor * llm_graph_context::build_ffn(
 
     if (down) {
         cur = build_lora_mm(down, cur);
+        if (arch == LLM_ARCH_GLM4) {
+            // GLM4 seems to have numerical issues with half-precision accumulators
+            ggml_mul_mat_set_prec(cur, GGML_PREC_F32);
+        }
     }
 
     if (down_b) {
@@ -916,28 +934,35 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
     ggml_tensor * up = build_lora_mm_id(up_exps, cur, selected_experts); // [n_ff, n_expert_used, n_tokens]
     cb(up, "ffn_moe_up", il);
 
-    ggml_tensor * gate = build_lora_mm_id(gate_exps, cur, selected_experts); // [n_ff, n_expert_used, n_tokens]
-    cb(gate, "ffn_moe_gate", il);
+    ggml_tensor * experts = nullptr;
+    if (gate_exps) {
+        cur = build_lora_mm_id(gate_exps, cur, selected_experts); // [n_ff, n_expert_used, n_tokens]
+        cb(cur, "ffn_moe_gate", il);
+    } else {
+        cur = up;
+    }
 
     switch (type_op) {
         case LLM_FFN_SILU:
             {
-                gate = ggml_silu(ctx0, gate);
-                cb(gate, "ffn_moe_silu", il);
+                cur = ggml_silu(ctx0, cur);
+                cb(cur, "ffn_moe_silu", il);
             } break;
         case LLM_FFN_GELU:
             {
-                gate = ggml_gelu(ctx0, gate);
-                cb(gate, "ffn_moe_gelu", il);
+                cur = ggml_gelu(ctx0, cur);
+                cb(cur, "ffn_moe_gelu", il);
             } break;
         default:
             GGML_ABORT("fatal error");
     }
 
-    ggml_tensor * par = ggml_mul(ctx0, up, gate); // [n_ff, n_expert_used, n_tokens]
-    cb(par, "ffn_moe_gate_par", il);
+    if (gate_exps) {
+        cur = ggml_mul(ctx0, cur, up); // [n_ff, n_expert_used, n_tokens]
+        cb(cur, "ffn_moe_gate_par", il);
+    }
 
-    ggml_tensor * experts = build_lora_mm_id(down_exps, par, selected_experts); // [n_embd, n_expert_used, n_tokens]
+    experts = build_lora_mm_id(down_exps, cur, selected_experts); // [n_embd, n_expert_used, n_tokens]
     cb(experts, "ffn_moe_down", il);
 
     if (!weight_before_ffn) {
@@ -1020,11 +1045,11 @@ ggml_tensor * llm_graph_context::build_inp_embd(ggml_tensor * tok_embd) const {
 }
 
 ggml_tensor * llm_graph_context::build_inp_pos() const {
-    auto inp = std::make_unique<llm_graph_input_pos>(n_pos_per_token());
+    auto inp = std::make_unique<llm_graph_input_pos>(n_pos_per_embd());
 
     auto & cur = inp->pos;
 
-    cur = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens*n_pos_per_token());
+    cur = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens*n_pos_per_embd());
     ggml_set_input(cur);
 
     res->add_input(std::move(inp));
@@ -1033,11 +1058,12 @@ ggml_tensor * llm_graph_context::build_inp_pos() const {
 }
 
 ggml_tensor * llm_graph_context::build_inp_attn_scale() const {
-    auto inp = std::make_unique<llm_graph_input_attn_temp>(n_pos_per_token(), hparams.n_attn_temp_floor_scale, hparams.f_attn_temp_scale);
+    auto inp = std::make_unique<llm_graph_input_attn_temp>(hparams.n_attn_temp_floor_scale, hparams.f_attn_temp_scale);
 
     auto & cur = inp->attn_scale;
 
-    cur = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, 1, 1, n_tokens*n_pos_per_token());
+    // this need to be 1x1xN for broadcasting
+    cur = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, 1, 1, n_tokens);
     ggml_set_input(cur);
 
     res->add_input(std::move(inp));

llama/llama.cpp/src/llama-graph.h

@@ -91,29 +91,27 @@ public:
 
 class llm_graph_input_pos : public llm_graph_input_i {
 public:
-    llm_graph_input_pos(int64_t n_pos_per_token) : n_pos_per_token(n_pos_per_token) {}
+    llm_graph_input_pos(int64_t n_pos_per_embd) : n_pos_per_embd(n_pos_per_embd) {}
     virtual ~llm_graph_input_pos() = default;
 
     void set_input(const llama_ubatch * ubatch) override;
 
     ggml_tensor * pos = nullptr; // I32 [n_batch]
 
-    const int64_t n_pos_per_token = 1;
+    const int64_t n_pos_per_embd = 1;
 };
 
 // temperature tuning, used by llama4
 class llm_graph_input_attn_temp : public llm_graph_input_i {
 public:
-    llm_graph_input_attn_temp(int64_t n_pos_per_token, uint32_t n_attn_temp_floor_scale, float f_attn_temp_scale)
-        : n_pos_per_token(n_pos_per_token), n_attn_temp_floor_scale(n_attn_temp_floor_scale), f_attn_temp_scale(f_attn_temp_scale) {}
+    llm_graph_input_attn_temp(uint32_t n_attn_temp_floor_scale, float f_attn_temp_scale)
+        : n_attn_temp_floor_scale(n_attn_temp_floor_scale), f_attn_temp_scale(f_attn_temp_scale) {}
     virtual ~llm_graph_input_attn_temp() = default;
 
     void set_input(const llama_ubatch * ubatch) override;
 
     ggml_tensor * attn_scale = nullptr; // F32 [n_batch]
 
-    const int64_t n_pos_per_token = 1;
-
     const uint32_t n_attn_temp_floor_scale;
     const float    f_attn_temp_scale;
 };
@@ -430,7 +428,7 @@ struct llm_graph_context {
 
     llm_graph_context(const llm_graph_params & params);
 
-    int64_t n_pos_per_token() const;
+    int64_t n_pos_per_embd() const;
 
     void cb(ggml_tensor * cur, const char * name, int il) const;
 

llama/llama.cpp/src/llama-hparams.h

@@ -72,6 +72,7 @@ struct llama_hparams {
     float    expert_weights_scale = 0.0;
     bool     expert_weights_norm  = false;
     uint32_t expert_gating_func   = LLAMA_EXPERT_GATING_FUNC_TYPE_NONE;
+    uint32_t moe_every_n_layers   = 0;
 
     float f_norm_eps;
     float f_norm_rms_eps;

llama/llama.cpp/src/llama-model.cpp

@@ -43,11 +43,13 @@ const char * llm_type_name(llm_type type) {
         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 +68,7 @@ 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 +77,7 @@ 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 +92,10 @@ 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";
     }
 }
@@ -709,10 +713,12 @@ 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;
@@ -805,6 +811,10 @@ 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;
@@ -814,6 +824,8 @@ 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;
@@ -1425,7 +1437,6 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                     default: type = LLM_TYPE_UNKNOWN;
                 }
             } break;
-        case LLM_ARCH_MISTRAL3: break;
         default: throw std::runtime_error("unsupported model architecture");
     }
@@ -2133,6 +2144,7 @@ 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);
@@ -2166,20 +2178,31 @@ 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);
@@ -6074,6 +6097,11 @@ 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 = ggml_add(ctx0, cur, model.layers[il].bqkv);
+                    cb(cur, "bqkv", il);
+                }
+
                 Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd,     n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd)));
                 Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd)));
                 Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)));
@@ -6126,13 +6154,29 @@ 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,
@@ -6140,6 +6184,7 @@ 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,
@@ -6147,8 +6192,8 @@ 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 = ggml_add(ctx0, cur, ffn_inp);
@@ -13349,6 +13394,7 @@ 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;
@@ -13705,7 +13751,6 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
         case LLM_ARCH_CHAMELEON:
         case LLM_ARCH_SOLAR:
         case LLM_ARCH_BAILINGMOE:
-        case LLM_ARCH_MISTRAL3:
             return LLAMA_ROPE_TYPE_NORM;
         // the pairs of head values are offset by n_rot/2
@@ -13714,6 +13759,7 @@ 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:

llama/llama.cpp/src/llama-model.h

@@ -40,11 +40,13 @@ enum llm_type {
     LLM_TYPE_770M,
     LLM_TYPE_780M,
     LLM_TYPE_0_5B,
+    LLM_TYPE_0_6B,
     LLM_TYPE_1B,
     LLM_TYPE_1_3B,
     LLM_TYPE_1_4B,
     LLM_TYPE_1_5B,
     LLM_TYPE_1_6B,
+    LLM_TYPE_1_7B,
     LLM_TYPE_1_8B,
     LLM_TYPE_2B,
     LLM_TYPE_2_8B,
@@ -64,6 +66,7 @@ enum llm_type {
     LLM_TYPE_16B,
     LLM_TYPE_20B,
     LLM_TYPE_22B,
+    LLM_TYPE_27B,
     LLM_TYPE_30B,
     LLM_TYPE_32B,
     LLM_TYPE_34B,
@@ -73,6 +76,7 @@ enum llm_type {
     LLM_TYPE_70B,
     LLM_TYPE_90B,
     LLM_TYPE_236B,
+    LLM_TYPE_290B,
     LLM_TYPE_314B,
     LLM_TYPE_671B,
     LLM_TYPE_SMALL,
@@ -87,10 +91,10 @@ enum llm_type {
     LLM_TYPE_16x3_8B,
     LLM_TYPE_10B_128x3_66B,
     LLM_TYPE_57B_A14B,
-    LLM_TYPE_27B,
-    LLM_TYPE_290B,
     LLM_TYPE_17B_16E, // llama4 Scout
     LLM_TYPE_17B_128E, // llama4 Maverick
+    LLM_TYPE_30B_A3B,
+    LLM_TYPE_235B_A22B,
 };
 
 struct llama_layer_posnet {

llama/llama.cpp/src/llama-quant.cpp

@@ -744,10 +744,6 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
         // This used to be a regex, but <regex> has an extreme cost to compile times.
         bool quantize = name.rfind("weight") == name.size() - 6; // ends with 'weight'?
 
-        // don't quantize vision stuff
-        quantize &= name.find("v.") == std::string::npos;
-        quantize &= name.find("mm.") == std::string::npos;
-
         // quantize only 2D and 3D tensors (experts)
         quantize &= (ggml_n_dims(tensor) >= 2);
 

llama/llama.cpp/src/llama-sampling.cpp

@@ -232,7 +232,7 @@ static void llama_sampler_top_k_impl(llama_token_data_array * cur_p, int32_t k)
     // }
 
     if (k <= 0) {
-        k = cur_p->size;
+        return;
     }
 
     k = std::min(k, (int) cur_p->size);
@@ -298,6 +298,7 @@ static void llama_sampler_top_k_impl(llama_token_data_array * cur_p, int32_t k)
         }
         cur_p->sorted = true;
     }
+
     cur_p->size = k;
 }
 

llama/llama.cpp/src/llama-vocab.cpp

@@ -1497,7 +1497,8 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                     tokenizer_pre == "llama3"   ||
                     tokenizer_pre == "llama-v3" ||
                     tokenizer_pre == "llama-bpe"||
-                    tokenizer_pre == "falcon3") {
+                    tokenizer_pre == "falcon3"  ||
+                    tokenizer_pre == "pixtral") {
                 pre_type = LLAMA_VOCAB_PRE_TYPE_LLAMA3;
                 ignore_merges = true;
                 add_bos = true;

llama/llama.go

@@ -2,6 +2,7 @@ package llama
 
 /*
 #cgo CFLAGS: -std=c11
+#cgo windows CFLAGS: -Wno-dll-attribute-on-redeclaration
 #cgo CXXFLAGS: -std=c++17
 #cgo CPPFLAGS: -I${SRCDIR}/llama.cpp/include
 #cgo CPPFLAGS: -I${SRCDIR}/llama.cpp/common
@@ -198,7 +199,6 @@ type ModelParams struct {
 	NumGpuLayers int
 	MainGpu      int
 	UseMmap      bool
-	UseMlock     bool
 	TensorSplit  []float32
 	Progress     func(float32)
 	VocabOnly    bool
@@ -217,7 +217,6 @@ func LoadModelFromFile(modelPath string, params ModelParams) (*Model, error) {
 	cparams.n_gpu_layers = C.int(params.NumGpuLayers)
 	cparams.main_gpu = C.int32_t(params.MainGpu)
 	cparams.use_mmap = C.bool(params.UseMmap)
-	cparams.use_mlock = C.bool(params.UseMlock)
 	cparams.vocab_only = C.bool(params.VocabOnly)
 
 	if len(params.TensorSplit) > 0 {
@@ -461,24 +460,6 @@ func (m *Model) NEmbd() int {
 	return int(C.llama_model_n_embd(m.c))
 }
 
-func Quantize(infile, outfile string, ftype uint32) error {
-	cinfile := C.CString(infile)
-	defer C.free(unsafe.Pointer(cinfile))
-
-	coutfile := C.CString(outfile)
-	defer C.free(unsafe.Pointer(coutfile))
-
-	params := C.llama_model_quantize_default_params()
-	params.nthread = -1
-	params.ftype = ftype
-
-	if rc := C.llama_model_quantize(cinfile, coutfile, &params); rc != 0 {
-		return fmt.Errorf("llama_model_quantize: %d", rc)
-	}
-
-	return nil
-}
-
 // vision processing
 type ClipContext struct {
 	c *C.struct_clip_ctx
@@ -606,9 +587,6 @@ type SamplingParams struct {
 	PenaltyRepeat  float32
 	PenaltyFreq    float32
 	PenaltyPresent float32
-	Mirostat       int
-	MirostatTau    float32
-	MirostatEta    float32
 	PenalizeNl     bool
 	Seed           uint32
 	Grammar        string
@@ -625,9 +603,6 @@ func NewSamplingContext(model *Model, params SamplingParams) (*SamplingContext,
 	cparams.penalty_repeat = C.float(params.PenaltyRepeat)
 	cparams.penalty_freq = C.float(params.PenaltyFreq)
 	cparams.penalty_present = C.float(params.PenaltyFreq)
-	cparams.mirostat = C.int32_t(params.Mirostat)
-	cparams.mirostat_tau = C.float(params.MirostatTau)
-	cparams.mirostat_eta = C.float(params.MirostatEta)
 	cparams.seed = C.uint32_t(params.Seed)
 
 	grammar := C.CString(params.Grammar)

llama/patches/0001-ggml-backend-malloc-and-free-using-the-same-compiler.patch

@@ -85,7 +85,7 @@ index e2617b06..242e50a7 100644
  
  /**
 diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu
-index a7febef7..31750b6f 100644
+index 9fb2134f..04ce764e 100644
 --- a/ggml/src/ggml-cuda/ggml-cuda.cu
 +++ b/ggml/src/ggml-cuda/ggml-cuda.cu
 @@ -534,6 +534,7 @@ struct ggml_backend_cuda_buffer_context {
@@ -125,10 +125,10 @@ index 50579227..2799a0a5 100644
  
  static void * ggml_backend_kompute_buffer_get_base(ggml_backend_buffer_t buffer) {
 diff --git a/ggml/src/ggml-metal/ggml-metal.m b/ggml/src/ggml-metal/ggml-metal.m
-index 266d8af4..12886cd3 100644
+index d92392ed..425524d0 100644
 --- a/ggml/src/ggml-metal/ggml-metal.m
 +++ b/ggml/src/ggml-metal/ggml-metal.m
-@@ -4759,6 +4759,7 @@ static void ggml_backend_metal_buffer_free_buffer(ggml_backend_buffer_t buffer)
+@@ -5077,6 +5077,7 @@ static void ggml_backend_metal_buffer_free_buffer(ggml_backend_buffer_t buffer)
      }
  
      free(ctx);
@@ -149,10 +149,10 @@ index 05a2f4e6..392cc18d 100644
  
  static void * ggml_backend_opencl_buffer_get_base(ggml_backend_buffer_t buffer) {
 diff --git a/ggml/src/ggml-rpc/ggml-rpc.cpp b/ggml/src/ggml-rpc/ggml-rpc.cpp
-index a0667b7d..bd83adc5 100644
+index 140a775f..e33c4ba0 100644
 --- a/ggml/src/ggml-rpc/ggml-rpc.cpp
 +++ b/ggml/src/ggml-rpc/ggml-rpc.cpp
-@@ -468,6 +468,7 @@ static void ggml_backend_rpc_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+@@ -477,6 +477,7 @@ static void ggml_backend_rpc_buffer_free_buffer(ggml_backend_buffer_t buffer) {
      bool status = send_rpc_cmd(ctx->sock, RPC_CMD_FREE_BUFFER, &request, sizeof(request), nullptr, 0);
      GGML_ASSERT(status);
      delete ctx;
@@ -161,10 +161,10 @@ index a0667b7d..bd83adc5 100644
  
  static void * ggml_backend_rpc_buffer_get_base(ggml_backend_buffer_t buffer) {
 diff --git a/ggml/src/ggml-sycl/ggml-sycl.cpp b/ggml/src/ggml-sycl/ggml-sycl.cpp
-index 1de34c96..4600f61e 100644
+index 66b6f2cc..e3e6deae 100644
 --- a/ggml/src/ggml-sycl/ggml-sycl.cpp
 +++ b/ggml/src/ggml-sycl/ggml-sycl.cpp
-@@ -316,6 +316,7 @@ ggml_backend_sycl_buffer_free_buffer(ggml_backend_buffer_t buffer) try {
+@@ -317,6 +317,7 @@ ggml_backend_sycl_buffer_free_buffer(ggml_backend_buffer_t buffer) try {
      ggml_sycl_set_device(ctx->device);
  
      delete ctx;
@@ -172,7 +172,7 @@ index 1de34c96..4600f61e 100644
  }
  catch (sycl::exception const &exc) {
    std::cerr << exc.what() << "Exception caught at file:" << __FILE__
-@@ -761,6 +762,7 @@ struct ggml_backend_sycl_split_buffer_context {
+@@ -762,6 +763,7 @@ struct ggml_backend_sycl_split_buffer_context {
  static void ggml_backend_sycl_split_buffer_free_buffer(ggml_backend_buffer_t buffer) {
      ggml_backend_sycl_split_buffer_context * ctx = (ggml_backend_sycl_split_buffer_context *)buffer->context;
      delete ctx;
@@ -180,7 +180,7 @@ index 1de34c96..4600f61e 100644
  }
  
  static void * ggml_backend_sycl_split_buffer_get_base(ggml_backend_buffer_t buffer) {
-@@ -1095,6 +1097,7 @@ static const char * ggml_backend_sycl_host_buffer_type_name(ggml_backend_buffer_
+@@ -1096,6 +1098,7 @@ static const char * ggml_backend_sycl_host_buffer_type_name(ggml_backend_buffer_
  
  static void ggml_backend_sycl_host_buffer_free_buffer(ggml_backend_buffer_t buffer) {
      ggml_sycl_host_free(buffer->context);
@@ -189,10 +189,10 @@ index 1de34c96..4600f61e 100644
  
  static ggml_backend_buffer_t ggml_backend_sycl_host_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
 diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
-index 39f3cd34..c569a8a5 100644
+index c0bdb9e1..03d03064 100644
 --- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp
 +++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
-@@ -8653,6 +8653,7 @@ static void ggml_backend_vk_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+@@ -8660,6 +8660,7 @@ static void ggml_backend_vk_buffer_free_buffer(ggml_backend_buffer_t buffer) {
      ggml_backend_vk_buffer_context * ctx = (ggml_backend_vk_buffer_context *)buffer->context;
      ggml_vk_destroy_buffer(ctx->dev_buffer);
      delete ctx;
@@ -200,7 +200,7 @@ index 39f3cd34..c569a8a5 100644
  }
  
  static void * ggml_backend_vk_buffer_get_base(ggml_backend_buffer_t buffer) {
-@@ -8796,6 +8797,7 @@ static const char * ggml_backend_vk_host_buffer_name(ggml_backend_buffer_t buffe
+@@ -8803,6 +8804,7 @@ static const char * ggml_backend_vk_host_buffer_name(ggml_backend_buffer_t buffe
  static void ggml_backend_vk_host_buffer_free_buffer(ggml_backend_buffer_t buffer) {
      VK_LOG_MEMORY("ggml_backend_vk_host_buffer_free_buffer()");
      ggml_vk_host_free(vk_instance.devices[0], buffer->context);

llama/patches/0002-pretokenizer.patch

@@ -10,7 +10,7 @@ logs instead of throwing an error
  1 file changed, 3 insertions(+), 11 deletions(-)
 
 diff --git a/src/llama-vocab.cpp b/src/llama-vocab.cpp
-index 48060517..a35b498c 100644
+index 50ded286..a9ee9f03 100644
 --- a/src/llama-vocab.cpp
 +++ b/src/llama-vocab.cpp
 @@ -1491,16 +1491,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
@@ -31,7 +31,7 @@ index 48060517..a35b498c 100644
                  pre_type = LLAMA_VOCAB_PRE_TYPE_DEFAULT;
              } else if (
                      tokenizer_pre == "llama3"   ||
-@@ -1634,7 +1625,8 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
+@@ -1635,7 +1626,8 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                  pre_type = LLAMA_VOCAB_PRE_TYPE_BAILINGMOE;
                  clean_spaces = false;
              } else {

llama/patches/0003-embeddings.patch

@@ -11,10 +11,10 @@ instead of forcing one or the error
  1 file changed, 3 insertions(+), 3 deletions(-)
 
 diff --git a/src/llama-context.cpp b/src/llama-context.cpp
-index 983385f8..32f59819 100644
+index 5a2eef9b..9c1fe93f 100644
 --- a/src/llama-context.cpp
 +++ b/src/llama-context.cpp
-@@ -1236,7 +1236,7 @@ int llama_context::decode(llama_batch & inp_batch) {
+@@ -1225,7 +1225,7 @@ int llama_context::decode(llama_batch & inp_batch) {
      int64_t n_outputs_all = 0;
  
      // count outputs
@@ -23,7 +23,7 @@ index 983385f8..32f59819 100644
          for (uint32_t i = 0; i < n_tokens_all; ++i) {
              n_outputs_all += batch.logits[i] != 0;
          }
-@@ -1348,7 +1348,7 @@ int llama_context::decode(llama_batch & inp_batch) {
+@@ -1337,7 +1337,7 @@ int llama_context::decode(llama_batch & inp_batch) {
          //    ggml_graph_dump_dot(gf, NULL, "llama.dot");
          //}
  
@@ -32,7 +32,7 @@ index 983385f8..32f59819 100644
          auto * t_embd   = cparams.embeddings ? res->get_embd() : nullptr;
  
          if (t_embd && res->get_embd_pooled()) {
-@@ -1492,7 +1492,7 @@ int32_t llama_context::output_reserve(int32_t n_outputs) {
+@@ -1481,7 +1481,7 @@ int32_t llama_context::output_reserve(int32_t n_outputs) {
      const auto n_embd  = hparams.n_embd;
  
      // TODO: use a per-batch flag for logits presence instead

llama/patches/0004-clip-unicode.patch

@@ -10,12 +10,12 @@ filesystems for paths that include wide characters
  1 file changed, 39 insertions(+)
 
 diff --git a/examples/llava/clip.cpp b/examples/llava/clip.cpp
-index 75970615..d57b4bd6 100644
+index ad3e7df1..b3218c78 100644
 --- a/examples/llava/clip.cpp
 +++ b/examples/llava/clip.cpp
-@@ -29,6 +29,19 @@
- #include <limits>
+@@ -30,6 +30,19 @@
  #include <array>
+ #include <numeric>
  
 +#if defined(_WIN32)
 +#define WIN32_LEAN_AND_MEAN
@@ -33,7 +33,7 @@ index 75970615..d57b4bd6 100644
  struct clip_logger_state g_logger_state = {GGML_LOG_LEVEL_CONT, clip_log_callback_default, NULL};
  
  //#define CLIP_DEBUG_FUNCTIONS
-@@ -1430,7 +1443,29 @@ struct clip_model_loader {
+@@ -1971,7 +1984,29 @@ struct clip_model_loader {
          {
              std::vector<uint8_t> read_buf;
  
@@ -63,7 +63,7 @@ index 75970615..d57b4bd6 100644
              if (!fin) {
                  throw std::runtime_error(string_format("%s: failed to open %s\n", __func__, fname.c_str()));
              }
-@@ -1457,7 +1492,11 @@ struct clip_model_loader {
+@@ -1998,7 +2033,11 @@ struct clip_model_loader {
                      ggml_backend_tensor_set(cur, read_buf.data(), 0, num_bytes);
                  }
              }

llama/patches/0005-solar-pro.patch

@@ -15,10 +15,10 @@ adds support for the Solar Pro architecture
  7 files changed, 248 insertions(+)
 
 diff --git a/src/llama-arch.cpp b/src/llama-arch.cpp
-index 62e1480b..f754bc8f 100644
+index f2bc8ca7..5ab3f572 100644
 --- a/src/llama-arch.cpp
 +++ b/src/llama-arch.cpp
-@@ -68,6 +68,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
+@@ -69,6 +69,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
      { LLM_ARCH_GRANITE,          "granite"          },
      { LLM_ARCH_GRANITE_MOE,      "granitemoe"       },
      { LLM_ARCH_CHAMELEON,        "chameleon"        },
@@ -26,7 +26,7 @@ index 62e1480b..f754bc8f 100644
      { LLM_ARCH_WAVTOKENIZER_DEC, "wavtokenizer-dec" },
      { LLM_ARCH_PLM,              "plm"              },
      { LLM_ARCH_BAILINGMOE,       "bailingmoe"       },
-@@ -140,6 +141,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
+@@ -142,6 +143,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
      { LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT,       "%s.attention.relative_buckets_count"       },
      { LLM_KV_ATTENTION_SLIDING_WINDOW,               "%s.attention.sliding_window"               },
      { LLM_KV_ATTENTION_SCALE,                        "%s.attention.scale"                        },
@@ -34,7 +34,7 @@ index 62e1480b..f754bc8f 100644
      { LLM_KV_ATTENTION_KEY_LENGTH_MLA,               "%s.attention.key_length_mla"               },
      { LLM_KV_ATTENTION_VALUE_LENGTH_MLA,             "%s.attention.value_length_mla"             },
  
-@@ -1482,6 +1484,24 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
+@@ -1502,6 +1504,24 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
              { LLM_TENSOR_ATTN_K_NORM,     "blk.%d.attn_k_norm" },
          },
      },
@@ -59,7 +59,7 @@ index 62e1480b..f754bc8f 100644
      {
          LLM_ARCH_WAVTOKENIZER_DEC,
          {
-@@ -1660,6 +1680,7 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
+@@ -1680,6 +1700,7 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
      {LLM_TENSOR_FFN_EXP_PROBS_B,            {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}},
      // this tensor is loaded for T5, but never used
      {LLM_TENSOR_DEC_CROSS_ATTN_REL_B,       {LLM_TENSOR_LAYER_REPEATING, GGML_OP_NONE}},
@@ -68,10 +68,10 @@ index 62e1480b..f754bc8f 100644
      {LLM_TENSOR_POS_NET_NORM,               {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
      {LLM_TENSOR_POS_NET_NORM1,              {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
 diff --git a/src/llama-arch.h b/src/llama-arch.h
-index 98ca00a1..439aaeab 100644
+index 41a023da..525c1b7d 100644
 --- a/src/llama-arch.h
 +++ b/src/llama-arch.h
-@@ -72,6 +72,7 @@ enum llm_arch {
+@@ -73,6 +73,7 @@ enum llm_arch {
      LLM_ARCH_GRANITE,
      LLM_ARCH_GRANITE_MOE,
      LLM_ARCH_CHAMELEON,
@@ -79,7 +79,7 @@ index 98ca00a1..439aaeab 100644
      LLM_ARCH_WAVTOKENIZER_DEC,
      LLM_ARCH_PLM,
      LLM_ARCH_BAILINGMOE,
-@@ -144,6 +145,7 @@ enum llm_kv {
+@@ -146,6 +147,7 @@ enum llm_kv {
      LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT,
      LLM_KV_ATTENTION_SLIDING_WINDOW,
      LLM_KV_ATTENTION_SCALE,
@@ -87,7 +87,7 @@ index 98ca00a1..439aaeab 100644
      LLM_KV_ATTENTION_KEY_LENGTH_MLA,
      LLM_KV_ATTENTION_VALUE_LENGTH_MLA,
  
-@@ -344,6 +346,7 @@ enum llm_tensor {
+@@ -346,6 +348,7 @@ enum llm_tensor {
      LLM_TENSOR_ENC_OUTPUT_NORM,
      LLM_TENSOR_CLS,
      LLM_TENSOR_CLS_OUT,
@@ -115,7 +115,7 @@ index 90dfe7a7..8a667960 100644
      if (il < n_layer) {
          return n_swa > 0 && n_swa_pattern > 0 && il % n_swa_pattern < (n_swa_pattern - 1);
 diff --git a/src/llama-hparams.h b/src/llama-hparams.h
-index 80fcd65d..6e278945 100644
+index 7ee6a5b7..48dce407 100644
 --- a/src/llama-hparams.h
 +++ b/src/llama-hparams.h
 @@ -55,6 +55,8 @@ struct llama_hparams {
@@ -127,7 +127,7 @@ index 80fcd65d..6e278945 100644
      uint32_t n_layer_dense_lead = 0;
      uint32_t n_lora_q           = 0;
      uint32_t n_lora_kv          = 0;
-@@ -153,6 +155,9 @@ struct llama_hparams {
+@@ -154,6 +156,9 @@ struct llama_hparams {
      // dimension of the recurrent state embeddings
      uint32_t n_embd_v_s() const;
  
@@ -150,10 +150,10 @@ index ea73a8a7..a012aeae 100644
  llama_model_loader::llama_model_loader(
          const std::string & fname,
 diff --git a/src/llama-model.cpp b/src/llama-model.cpp
-index 6b7bfecf..aba42819 100644
+index 822e2bb2..572378c9 100644
 --- a/src/llama-model.cpp
 +++ b/src/llama-model.cpp
-@@ -1374,6 +1374,21 @@ void llama_model::load_hparams(llama_model_loader & ml) {
+@@ -1386,6 +1386,21 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                      default: type = LLM_TYPE_UNKNOWN;
                 }
              } break;
@@ -175,7 +175,7 @@ index 6b7bfecf..aba42819 100644
          case LLM_ARCH_WAVTOKENIZER_DEC:
              {
                  ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS,    hparams.f_norm_eps);
-@@ -3717,6 +3732,34 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
+@@ -3741,6 +3756,34 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
  
                          layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
  
@@ -210,7 +210,7 @@ index 6b7bfecf..aba42819 100644
                          layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 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}, 0);
-@@ -12296,6 +12339,165 @@ struct llm_build_chameleon : public llm_graph_context {
+@@ -12342,6 +12385,165 @@ struct llm_build_chameleon : public llm_graph_context {
      }
  };
  
@@ -376,7 +376,7 @@ index 6b7bfecf..aba42819 100644
  struct llm_build_wavtokenizer_dec : public llm_graph_context {
      llm_build_wavtokenizer_dec(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) {
          ggml_tensor * cur;
-@@ -13045,6 +13247,10 @@ llm_graph_result_ptr llama_model::build_graph(
+@@ -13092,6 +13294,10 @@ llm_graph_result_ptr llama_model::build_graph(
              {
                  llm = std::make_unique<llm_build_chameleon>(*this, params, gf);
              } break;
@@ -387,7 +387,7 @@ index 6b7bfecf..aba42819 100644
          case LLM_ARCH_WAVTOKENIZER_DEC:
              {
                  llm = std::make_unique<llm_build_wavtokenizer_dec>(*this, params, gf);
-@@ -13191,6 +13397,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
+@@ -13238,6 +13444,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
          case LLM_ARCH_GRANITE:
          case LLM_ARCH_GRANITE_MOE:
          case LLM_ARCH_CHAMELEON:
@@ -396,18 +396,18 @@ index 6b7bfecf..aba42819 100644
              return LLAMA_ROPE_TYPE_NORM;
  
 diff --git a/src/llama-model.h b/src/llama-model.h
-index fd82d106..5865d5e9 100644
+index 95eca002..856e6042 100644
 --- a/src/llama-model.h
 +++ b/src/llama-model.h
-@@ -62,6 +62,7 @@ enum llm_type {
+@@ -64,6 +64,7 @@ enum llm_type {
      LLM_TYPE_15B,
      LLM_TYPE_16B,
      LLM_TYPE_20B,
 +    LLM_TYPE_22B,
+     LLM_TYPE_27B,
      LLM_TYPE_30B,
      LLM_TYPE_32B,
-     LLM_TYPE_34B,
-@@ -307,6 +308,8 @@ struct llama_layer {
+@@ -311,6 +312,8 @@ struct llama_layer {
      struct ggml_tensor * ffn_up_scale   = nullptr;
      struct ggml_tensor * ffn_down_scale = nullptr;
  

llama/patches/0006-add-mllama-support.patch

@@ -5,7 +5,6 @@ Subject: [PATCH] add mllama support
 
 adds support for the llama 3.2 vision architecture
 ---
- examples/llava/gemma3-cli.cpp |   3 +-
  examples/llava/llava.cpp      |   5 +-
  examples/llava/mtmd.cpp       |   6 +-
  ggml/src/ggml-backend-reg.cpp |   6 +-
@@ -25,34 +24,13 @@ adds support for the llama 3.2 vision architecture
  src/llama-model.cpp           | 309 +++++++++++++++++++++++++++++++++-
  src/llama-model.h             |  12 ++
  src/llama-quant.cpp           |   4 +-
- 20 files changed, 475 insertions(+), 22 deletions(-)
+ 19 files changed, 473 insertions(+), 21 deletions(-)
 
-diff --git a/examples/llava/gemma3-cli.cpp b/examples/llava/gemma3-cli.cpp
-index 3d566475..654d1358 100644
---- a/examples/llava/gemma3-cli.cpp
-+++ b/examples/llava/gemma3-cli.cpp
-@@ -106,7 +106,7 @@ struct decode_embd_batch {
-     std::vector<llama_seq_id *> seq_ids;
-     std::vector<int8_t>         logits;
-     llama_batch batch;
--    decode_embd_batch(float * embd, int32_t n_tokens, llama_pos pos_0, llama_seq_id seq_id) {
-+    decode_embd_batch(float * embd, int32_t n_embd, int32_t n_tokens, llama_pos pos_0, llama_seq_id seq_id) {
-         pos     .resize(n_tokens);
-         n_seq_id.resize(n_tokens);
-         seq_ids .resize(n_tokens + 1);
-@@ -118,6 +118,7 @@ struct decode_embd_batch {
-             /*n_tokens       =*/ n_tokens,
-             /*tokens         =*/ nullptr,
-             /*embd           =*/ embd,
-+            /*n_embd         =*/ n_embd,
-             /*pos            =*/ pos.data(),
-             /*n_seq_id       =*/ n_seq_id.data(),
-             /*seq_id         =*/ seq_ids.data(),
 diff --git a/examples/llava/llava.cpp b/examples/llava/llava.cpp
-index 03a22cbb..5eb40bcd 100644
+index c00d16ae..bab027b5 100644
 --- a/examples/llava/llava.cpp
 +++ b/examples/llava/llava.cpp
-@@ -456,7 +456,7 @@ struct llava_embd_batch {
+@@ -457,7 +457,7 @@ struct llava_embd_batch {
      std::vector<llama_seq_id *> seq_ids;
      std::vector<int8_t>         logits;
      llama_batch batch;
@@ -61,7 +39,7 @@ index 03a22cbb..5eb40bcd 100644
          pos     .resize(n_tokens);
          n_seq_id.resize(n_tokens);
          seq_ids .resize(n_tokens + 1);
-@@ -468,6 +468,7 @@ struct llava_embd_batch {
+@@ -469,6 +469,7 @@ struct llava_embd_batch {
              /*n_tokens       =*/ n_tokens,
              /*tokens         =*/ nullptr,
              /*embd           =*/ embd,
@@ -69,7 +47,7 @@ index 03a22cbb..5eb40bcd 100644
              /*pos            =*/ pos.data(),
              /*n_seq_id       =*/ n_seq_id.data(),
              /*seq_id         =*/ seq_ids.data(),
-@@ -491,7 +492,7 @@ bool llava_eval_image_embed(llama_context * ctx_llama, const struct llava_image_
+@@ -492,7 +493,7 @@ bool llava_eval_image_embed(llama_context * ctx_llama, const struct llava_image_
              n_eval = n_batch;
          }
          float * embd = image_embed->embed+i*n_embd;
@@ -79,19 +57,19 @@ index 03a22cbb..5eb40bcd 100644
              LOG_ERR("%s : failed to eval\n", __func__);
              return false;
 diff --git a/examples/llava/mtmd.cpp b/examples/llava/mtmd.cpp
-index 3fd5bebc..f0cec596 100644
+index 7081fd73..c14ac501 100644
 --- a/examples/llava/mtmd.cpp
 +++ b/examples/llava/mtmd.cpp
-@@ -233,7 +233,7 @@ struct decode_embd_batch {
+@@ -476,7 +476,7 @@ struct decode_embd_batch {
      std::vector<llama_seq_id *> seq_ids;
      std::vector<int8_t>         logits;
      llama_batch batch;
--    decode_embd_batch(float * embd, int32_t n_tokens, llama_pos pos_0, llama_seq_id seq_id) {
-+    decode_embd_batch(float * embd, int32_t n_embd, int32_t n_tokens, llama_pos pos_0, llama_seq_id seq_id) {
-         pos     .resize(n_tokens);
+-    decode_embd_batch(float * embd, int32_t n_tokens, int n_pos_per_embd, int n_mmproj_embd) : n_pos_per_embd(n_pos_per_embd), n_mmproj_embd(n_mmproj_embd) {
++    decode_embd_batch(float * embd, int32_t n_embd, int32_t n_tokens, llama_pos pos_0, llama_seq_id seq_id) : n_pos_per_embd(n_pos_per_embd), n_mmproj_embd(n_mmproj_embd) {
+         pos     .resize(n_tokens * n_pos_per_embd);
          n_seq_id.resize(n_tokens);
          seq_ids .resize(n_tokens + 1);
-@@ -245,6 +245,7 @@ struct decode_embd_batch {
+@@ -487,6 +487,7 @@ struct decode_embd_batch {
              /*n_tokens       =*/ n_tokens,
              /*tokens         =*/ nullptr,
              /*embd           =*/ embd,
@@ -99,16 +77,16 @@ index 3fd5bebc..f0cec596 100644
              /*pos            =*/ pos.data(),
              /*n_seq_id       =*/ n_seq_id.data(),
              /*seq_id         =*/ seq_ids.data(),
-@@ -311,7 +312,8 @@ int32_t mtmd_helper_eval(mtmd_context * ctx,
- 
-             int32_t n_tokens = mtmd_image_tokens_get_n_tokens(chunk.tokens_image.get());
+@@ -610,7 +611,8 @@ int32_t mtmd_helper_eval(mtmd_context * ctx,
+             int32_t i_batch = 0;
+             int32_t n_img_batches = GGML_PAD(n_tokens, n_batch) / n_batch;
              float * embd = mtmd_get_output_embd(ctx);
--            decode_embd_batch batch_img(embd, n_tokens, n_past, 0);
+-            decode_embd_batch batch_embd(embd, n_tokens, n_pos_per_embd, n_mmproj_embd);
 +            int n_embd  = llama_model_n_embd(llama_get_model(lctx));
-+            decode_embd_batch batch_img(embd, n_embd, n_tokens, n_past, 0);
-             int64_t t1 = ggml_time_ms();
-             ret = llama_decode(lctx, batch_img.batch);
-             if (ret != 0) {
++            decode_embd_batch batch_embd(embd, n_embd, n_tokens, n_past, 0);
+ 
+             const int nx = mtmd_image_tokens_get_nx(chunk.tokens_image.get());
+             const int ny = mtmd_image_tokens_get_ny(chunk.tokens_image.get());
 diff --git a/ggml/src/ggml-backend-reg.cpp b/ggml/src/ggml-backend-reg.cpp
 index 405d8e31..82ae1b5b 100644
 --- a/ggml/src/ggml-backend-reg.cpp
@@ -127,10 +105,10 @@ index 405d8e31..82ae1b5b 100644
          register_backend(ggml_backend_rpc_reg());
  #endif
 diff --git a/include/llama.h b/include/llama.h
-index 5657fbf0..f91896e4 100644
+index 06c56395..f1628e88 100644
 --- a/include/llama.h
 +++ b/include/llama.h
-@@ -255,6 +255,7 @@ extern "C" {
+@@ -256,6 +256,7 @@ extern "C" {
  
          llama_token  *  token;
          float        *  embd;
@@ -138,7 +116,7 @@ index 5657fbf0..f91896e4 100644
          llama_pos    *  pos;
          int32_t      *  n_seq_id;
          llama_seq_id ** seq_id;
-@@ -357,6 +358,7 @@ extern "C" {
+@@ -358,6 +359,7 @@ extern "C" {
          bool offload_kqv; // whether to offload the KQV ops (including the KV cache) to GPU
          bool flash_attn;  // whether to use flash attention [EXPERIMENTAL]
          bool no_perf;     // whether to measure performance timings
@@ -146,7 +124,7 @@ index 5657fbf0..f91896e4 100644
  
          // Abort callback
          // if it returns true, execution of llama_decode() will be aborted
-@@ -458,6 +460,10 @@ extern "C" {
+@@ -459,6 +461,10 @@ extern "C" {
              struct llama_context_params   params),
              "use llama_init_from_model instead");
  
@@ -158,7 +136,7 @@ index 5657fbf0..f91896e4 100644
      LLAMA_API void llama_free(struct llama_context * ctx);
  
 diff --git a/src/llama-arch.cpp b/src/llama-arch.cpp
-index f754bc8f..0568565f 100644
+index 5ab3f572..eb7b5325 100644
 --- a/src/llama-arch.cpp
 +++ b/src/llama-arch.cpp
 @@ -6,6 +6,7 @@
@@ -169,7 +147,7 @@ index f754bc8f..0568565f 100644
      { LLM_ARCH_LLAMA4,           "llama4"           },
      { LLM_ARCH_DECI,             "deci"             },
      { LLM_ARCH_FALCON,           "falcon"           },
-@@ -142,6 +143,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
+@@ -144,6 +145,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
      { LLM_KV_ATTENTION_SLIDING_WINDOW,               "%s.attention.sliding_window"               },
      { LLM_KV_ATTENTION_SCALE,                        "%s.attention.scale"                        },
      { LLM_KV_ATTENTION_BLOCK_SKIP_CONNECTION,        "%s.attention.block_skip_connection"        },
@@ -177,7 +155,7 @@ index f754bc8f..0568565f 100644
      { LLM_KV_ATTENTION_KEY_LENGTH_MLA,               "%s.attention.key_length_mla"               },
      { LLM_KV_ATTENTION_VALUE_LENGTH_MLA,             "%s.attention.value_length_mla"             },
  
-@@ -271,6 +273,40 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
+@@ -273,6 +275,40 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
              { LLM_TENSOR_FFN_UP_SHEXP,    "blk.%d.ffn_up_shexp" },
          },
      },
@@ -218,7 +196,7 @@ index f754bc8f..0568565f 100644
      {
          LLM_ARCH_DECI,
          {
-@@ -1681,6 +1717,14 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
+@@ -1701,6 +1737,14 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
      // this tensor is loaded for T5, but never used
      {LLM_TENSOR_DEC_CROSS_ATTN_REL_B,       {LLM_TENSOR_LAYER_REPEATING, GGML_OP_NONE}},
      {LLM_TENSOR_BSKCN_TV,                   {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
@@ -234,7 +212,7 @@ index f754bc8f..0568565f 100644
      {LLM_TENSOR_POS_NET_NORM,               {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
      {LLM_TENSOR_POS_NET_NORM1,              {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
 diff --git a/src/llama-arch.h b/src/llama-arch.h
-index 439aaeab..6a989034 100644
+index 525c1b7d..bc8a4f0b 100644
 --- a/src/llama-arch.h
 +++ b/src/llama-arch.h
 @@ -11,6 +11,7 @@
@@ -245,7 +223,7 @@ index 439aaeab..6a989034 100644
      LLM_ARCH_DECI,
      LLM_ARCH_FALCON,
      LLM_ARCH_BAICHUAN,
-@@ -146,6 +147,7 @@ enum llm_kv {
+@@ -148,6 +149,7 @@ enum llm_kv {
      LLM_KV_ATTENTION_SLIDING_WINDOW,
      LLM_KV_ATTENTION_SCALE,
      LLM_KV_ATTENTION_BLOCK_SKIP_CONNECTION,
@@ -253,7 +231,7 @@ index 439aaeab..6a989034 100644
      LLM_KV_ATTENTION_KEY_LENGTH_MLA,
      LLM_KV_ATTENTION_VALUE_LENGTH_MLA,
  
-@@ -347,6 +349,14 @@ enum llm_tensor {
+@@ -349,6 +351,14 @@ enum llm_tensor {
      LLM_TENSOR_CLS,
      LLM_TENSOR_CLS_OUT,
      LLM_TENSOR_BSKCN_TV,
@@ -297,10 +275,10 @@ index 01d5ca57..8682b0e6 100644
          batch.token = (llama_token *) malloc(sizeof(llama_token) * n_tokens_alloc);
      }
 diff --git a/src/llama-context.cpp b/src/llama-context.cpp
-index 32f59819..0343ba8a 100644
+index 9c1fe93f..cd06ad91 100644
 --- a/src/llama-context.cpp
 +++ b/src/llama-context.cpp
-@@ -862,7 +862,7 @@ float * llama_context::get_logits_ith(int32_t i) {
+@@ -851,7 +851,7 @@ float * llama_context::get_logits_ith(int32_t i) {
              throw std::runtime_error(format("corrupt output buffer (j=%d, n_outputs=%d)", j, n_outputs));
          }
  
@@ -309,7 +287,7 @@ index 32f59819..0343ba8a 100644
      } catch (const std::exception & err) {
          LLAMA_LOG_ERROR("%s: invalid logits id %d, reason: %s\n", __func__, i, err.what());
  #ifndef NDEBUG
-@@ -983,6 +983,10 @@ void llama_context::set_warmup(bool value) {
+@@ -972,6 +972,10 @@ void llama_context::set_warmup(bool value) {
      cparams.warmup = value;
  }
  
@@ -320,7 +298,7 @@ index 32f59819..0343ba8a 100644
  void llama_context::set_adapter_lora(
              llama_adapter_lora * adapter,
              float scale) {
-@@ -1058,7 +1062,7 @@ int llama_context::encode(llama_batch & inp_batch) {
+@@ -1047,7 +1051,7 @@ int llama_context::encode(llama_batch & inp_batch) {
  
      const int64_t n_embd = hparams.n_embd;
  
@@ -329,7 +307,7 @@ index 32f59819..0343ba8a 100644
  
      const llama_ubatch ubatch = sbatch.split_simple(n_tokens);
  
-@@ -1198,10 +1202,9 @@ int llama_context::decode(llama_batch & inp_batch) {
+@@ -1187,10 +1191,9 @@ int llama_context::decode(llama_batch & inp_batch) {
  
      const llama_batch & batch = batch_allocr.batch;
  
@@ -341,7 +319,7 @@ index 32f59819..0343ba8a 100644
  
      const int64_t n_tokens_all = batch.n_tokens;
      const int64_t n_embd       = hparams.n_embd;
-@@ -1249,7 +1252,7 @@ int llama_context::decode(llama_batch & inp_batch) {
+@@ -1238,7 +1241,7 @@ int llama_context::decode(llama_batch & inp_batch) {
  
      const bool logits_all = n_outputs_all == n_tokens_all;
  
@@ -350,7 +328,7 @@ index 32f59819..0343ba8a 100644
              /* simple_split */ !kv_self->recurrent,
              /* logits_all   */ logits_all);
  
-@@ -1483,12 +1486,11 @@ int llama_context::decode(llama_batch & inp_batch) {
+@@ -1472,12 +1475,11 @@ int llama_context::decode(llama_batch & inp_batch) {
  
  int32_t llama_context::output_reserve(int32_t n_outputs) {
      const auto & hparams = model.hparams;
@@ -364,7 +342,7 @@ index 32f59819..0343ba8a 100644
      const auto n_embd  = hparams.n_embd;
  
      // TODO: use a per-batch flag for logits presence instead
-@@ -1558,7 +1560,7 @@ int32_t llama_context::output_reserve(int32_t n_outputs) {
+@@ -1545,7 +1547,7 @@ int32_t llama_context::output_reserve(int32_t n_outputs) {
  void llama_context::output_reorder() {
      auto & out_ids = sbatch.out_ids;
      if (!out_ids.empty()) {
@@ -373,7 +351,7 @@ index 32f59819..0343ba8a 100644
          const uint32_t n_embd  = model.hparams.n_embd;
  
          GGML_ASSERT((size_t) n_outputs == out_ids.size());
-@@ -2065,7 +2067,7 @@ size_t llama_context::state_write_data(llama_io_write_i & io) {
+@@ -2052,7 +2054,7 @@ size_t llama_context::state_write_data(llama_io_write_i & io) {
      {
          LLAMA_LOG_DEBUG("%s: - writing logits\n", __func__);
  
@@ -382,7 +360,7 @@ index 32f59819..0343ba8a 100644
  
          io.write(&logits_size, sizeof(logits_size));
  
-@@ -2248,6 +2250,7 @@ llama_context_params llama_context_default_params() {
+@@ -2235,6 +2237,7 @@ llama_context_params llama_context_default_params() {
          /*.offload_kqv                 =*/ true,
          /*.flash_attn                  =*/ false,
          /*.no_perf                     =*/ true,
@@ -390,7 +368,7 @@ index 32f59819..0343ba8a 100644
          /*.abort_callback              =*/ nullptr,
          /*.abort_callback_data         =*/ nullptr,
      };
-@@ -2375,6 +2378,10 @@ void llama_set_warmup(llama_context * ctx, bool warmup) {
+@@ -2362,6 +2365,10 @@ void llama_set_warmup(llama_context * ctx, bool warmup) {
      ctx->set_warmup(warmup);
  }
  
@@ -402,7 +380,7 @@ index 32f59819..0343ba8a 100644
      ctx->synchronize();
  }
 diff --git a/src/llama-context.h b/src/llama-context.h
-index 04facb54..baa03276 100644
+index 5457f077..a50c4afa 100644
 --- a/src/llama-context.h
 +++ b/src/llama-context.h
 @@ -65,6 +65,7 @@ struct llama_context {
@@ -426,10 +404,10 @@ index 30e550f0..85ad91b9 100644
  
      enum llama_pooling_type pooling_type;
 diff --git a/src/llama-graph.cpp b/src/llama-graph.cpp
-index a85e9728..d740c120 100644
+index fabb9ca2..b67216a4 100644
 --- a/src/llama-graph.cpp
 +++ b/src/llama-graph.cpp
-@@ -546,6 +546,12 @@ void llm_graph_input_attn_cross::set_input(const llama_ubatch * ubatch) {
+@@ -560,6 +560,12 @@ void llm_graph_input_attn_cross::set_input(const llama_ubatch * ubatch) {
      }
  }
  
@@ -442,7 +420,7 @@ index a85e9728..d740c120 100644
  //
  // llm_graph_context
  //
-@@ -1506,6 +1512,25 @@ llm_graph_input_attn_cross * llm_graph_context::build_attn_inp_cross() const {
+@@ -1532,6 +1538,25 @@ llm_graph_input_attn_cross * llm_graph_context::build_attn_inp_cross() const {
      return (llm_graph_input_attn_cross *) res->add_input(std::move(inp));
  }
  
@@ -469,7 +447,7 @@ index a85e9728..d740c120 100644
          llm_graph_input_attn_cross * inp,
          ggml_cgraph * gf,
 diff --git a/src/llama-graph.h b/src/llama-graph.h
-index d192dc14..260a2af2 100644
+index d0c8d321..0fe18150 100644
 --- a/src/llama-graph.h
 +++ b/src/llama-graph.h
 @@ -86,6 +86,7 @@ public:
@@ -480,7 +458,7 @@ index d192dc14..260a2af2 100644
  };
  
  class llm_graph_input_pos : public llm_graph_input_i {
-@@ -285,6 +286,16 @@ public:
+@@ -283,6 +284,16 @@ public:
      const llama_cross * cross = nullptr;
  };
  
@@ -497,7 +475,7 @@ index d192dc14..260a2af2 100644
  //
  // llm_graph_result
  //
-@@ -493,6 +504,7 @@ struct llm_graph_context {
+@@ -491,6 +502,7 @@ struct llm_graph_context {
      ggml_tensor * build_inp_cls() const;
      ggml_tensor * build_inp_s_copy() const;
      ggml_tensor * build_inp_s_mask() const;
@@ -518,7 +496,7 @@ index 8a667960..6a02de03 100644
 +    return std::find(cross_attn_layers.begin(), cross_attn_layers.end(), il) != cross_attn_layers.end();
 +}
 diff --git a/src/llama-hparams.h b/src/llama-hparams.h
-index 6e278945..c8a34d52 100644
+index 48dce407..b6fc7e6d 100644
 --- a/src/llama-hparams.h
 +++ b/src/llama-hparams.h
 @@ -2,6 +2,8 @@
@@ -546,7 +524,7 @@ index 6e278945..c8a34d52 100644
  
      uint32_t n_layer_dense_lead = 0;
      uint32_t n_lora_q           = 0;
-@@ -158,6 +162,9 @@ struct llama_hparams {
+@@ -159,6 +163,9 @@ struct llama_hparams {
      // Block skip connection
      bool n_bskcn(uint32_t n, uint32_t il) const;
  
@@ -593,10 +571,10 @@ index a012aeae..2e11507d 100644
      bool llama_model_loader::get_arr(const std::string & key, std::array<T, N_MAX> & result, bool required) {
          const int kid = gguf_find_key(meta.get(), key.c_str());
 diff --git a/src/llama-model.cpp b/src/llama-model.cpp
-index aba42819..d051696c 100644
+index 572378c9..9d099f11 100644
 --- a/src/llama-model.cpp
 +++ b/src/llama-model.cpp
-@@ -419,6 +419,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
+@@ -423,6 +423,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
  
      // get general kv
      ml.get_key(LLM_KV_GENERAL_NAME, name, false);
@@ -604,7 +582,7 @@ index aba42819..d051696c 100644
  
      // everything past this point is not vocab-related
      if (hparams.vocab_only) {
-@@ -430,6 +431,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
+@@ -434,6 +435,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
      ml.get_key(LLM_KV_BLOCK_COUNT,       hparams.n_layer);
      ml.get_key(LLM_KV_EXPERT_COUNT,      hparams.n_expert,      false);
      ml.get_key(LLM_KV_EXPERT_USED_COUNT, hparams.n_expert_used, false);
@@ -612,7 +590,7 @@ index aba42819..d051696c 100644
  
      if (arch == LLM_ARCH_WAVTOKENIZER_DEC) {
          ml.get_key(LLM_KV_FEATURES_LENGTH, hparams.n_embd_features);
-@@ -453,9 +455,11 @@ void llama_model::load_hparams(llama_model_loader & ml) {
+@@ -457,9 +459,11 @@ void llama_model::load_hparams(llama_model_loader & ml) {
      std::fill(hparams.n_head_arr.begin(),    hparams.n_head_arr.end(),    0);
      std::fill(hparams.n_head_kv_arr.begin(), hparams.n_head_kv_arr.end(), 0);
      std::fill(hparams.n_ff_arr.begin(),      hparams.n_ff_arr.end(),      0);
@@ -624,7 +602,7 @@ index aba42819..d051696c 100644
  
      // n_head_kv is optional, default to n_head
      hparams.n_head_kv_arr = hparams.n_head_arr;
-@@ -508,7 +512,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
+@@ -512,7 +516,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
  
          ml.get_key(LLM_KV_ROPE_DIMENSION_COUNT, hparams.n_rot, false);
  
@@ -633,7 +611,7 @@ index aba42819..d051696c 100644
              if (hparams.n_rot != hparams.n_embd_head_k) {
                  throw std::runtime_error(format("invalid n_rot: %u, expected %u", hparams.n_rot, hparams.n_embd_head_k));
              }
-@@ -571,6 +575,16 @@ void llama_model::load_hparams(llama_model_loader & ml) {
+@@ -575,6 +579,16 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                      hparams.use_kq_norm = false;
                  }
              } break;
@@ -650,7 +628,7 @@ index aba42819..d051696c 100644
          case LLM_ARCH_DECI:
              {
                  ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
-@@ -1550,7 +1564,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
+@@ -1562,7 +1576,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
          const int64_t n_embd_head_v = hparams.n_embd_head_v;
          const int64_t n_ff          = hparams.n_ff();
          const int64_t n_embd_gqa    = n_embd_v_gqa;
@@ -659,7 +637,7 @@ index aba42819..d051696c 100644
          const int64_t n_token_types = vocab.n_token_types();
          const int64_t n_rot         = hparams.n_rot;
          const int64_t n_expert      = hparams.n_expert;
-@@ -1803,6 +1817,52 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
+@@ -1815,6 +1829,52 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                          }
                      }
                  } break;
@@ -712,7 +690,7 @@ index aba42819..d051696c 100644
              case LLM_ARCH_DECI:
                  {
                      tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
-@@ -4683,6 +4743,246 @@ struct llm_build_llama : public llm_graph_context {
+@@ -4707,6 +4767,246 @@ struct llm_build_llama : public llm_graph_context {
      }
  };
  
@@ -959,7 +937,7 @@ index aba42819..d051696c 100644
  struct llm_build_deci : public llm_graph_context {
      llm_build_deci(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) {
          const int64_t n_embd_head = hparams.n_embd_head_v;
-@@ -13017,6 +13317,10 @@ llm_graph_result_ptr llama_model::build_graph(
+@@ -13063,6 +13363,10 @@ llm_graph_result_ptr llama_model::build_graph(
              {
                  llm = std::make_unique<llm_build_llama>(*this, params, gf);
              } break;
@@ -970,7 +948,7 @@ index aba42819..d051696c 100644
          case LLM_ARCH_DECI:
              {
                  llm = std::make_unique<llm_build_deci>(*this, params, gf);
-@@ -13377,6 +13681,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
+@@ -13424,6 +13728,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
          // use what we call a normal RoPE, operating on pairs of consecutive head values
          case LLM_ARCH_LLAMA:
          case LLM_ARCH_LLAMA4:
@@ -979,7 +957,7 @@ index aba42819..d051696c 100644
          case LLM_ARCH_BAICHUAN:
          case LLM_ARCH_STARCODER:
 diff --git a/src/llama-model.h b/src/llama-model.h
-index 5865d5e9..72bab5be 100644
+index 856e6042..6be91282 100644
 --- a/src/llama-model.h
 +++ b/src/llama-model.h
 @@ -11,6 +11,7 @@
@@ -990,15 +968,15 @@ index 5865d5e9..72bab5be 100644
  
  struct llama_cparams;
  struct llama_ubatch;
-@@ -70,6 +71,7 @@ enum llm_type {
+@@ -73,6 +74,7 @@ enum llm_type {
      LLM_TYPE_40B,
      LLM_TYPE_65B,
      LLM_TYPE_70B,
 +    LLM_TYPE_90B,
      LLM_TYPE_236B,
+     LLM_TYPE_290B,
      LLM_TYPE_314B,
-     LLM_TYPE_671B,
-@@ -310,6 +312,16 @@ struct llama_layer {
+@@ -314,6 +316,16 @@ struct llama_layer {
  
      struct ggml_tensor * bskcn_tv = nullptr;
  

llama/patches/0007-add-unpad-operator.patch

@@ -18,10 +18,10 @@ adds the unpad operator to GGML
  10 files changed, 223 insertions(+), 2 deletions(-)
 
 diff --git a/ggml/include/ggml.h b/ggml/include/ggml.h
-index 8fcc16df..d19fc167 100644
+index 1b8603e7..53ef31b2 100644
 --- a/ggml/include/ggml.h
 +++ b/ggml/include/ggml.h
-@@ -488,6 +488,7 @@ extern "C" {
+@@ -489,6 +489,7 @@ extern "C" {
          GGML_OP_UPSCALE, // nearest interpolate
          GGML_OP_PAD,
          GGML_OP_PAD_REFLECT_1D,
@@ -29,7 +29,7 @@ index 8fcc16df..d19fc167 100644
          GGML_OP_ARANGE,
          GGML_OP_TIMESTEP_EMBEDDING,
          GGML_OP_ARGSORT,
-@@ -1757,6 +1758,15 @@ extern "C" {
+@@ -1777,6 +1778,15 @@ extern "C" {
              int                   p0,
              int                   p1);
  
@@ -46,10 +46,10 @@ index 8fcc16df..d19fc167 100644
      // timesteps: [N,]
      // return: [N, dim]
 diff --git a/ggml/src/ggml-cpu/ggml-cpu.c b/ggml/src/ggml-cpu/ggml-cpu.c
-index 50400328..432942bf 100644
+index 64405449..34624cca 100644
 --- a/ggml/src/ggml-cpu/ggml-cpu.c
 +++ b/ggml/src/ggml-cpu/ggml-cpu.c
-@@ -1960,6 +1960,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
+@@ -1964,6 +1964,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
              {
                  ggml_compute_forward_pad_reflect_1d(params, tensor);
              } break;
@@ -60,7 +60,7 @@ index 50400328..432942bf 100644
          case GGML_OP_ARANGE:
              {
                  ggml_compute_forward_arange(params, tensor);
-@@ -2282,6 +2286,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
+@@ -2287,6 +2291,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
          case GGML_OP_UPSCALE:
          case GGML_OP_PAD:
          case GGML_OP_PAD_REFLECT_1D:
@@ -69,10 +69,10 @@ index 50400328..432942bf 100644
          case GGML_OP_TIMESTEP_EMBEDDING:
          case GGML_OP_ARGSORT:
 diff --git a/ggml/src/ggml-cpu/ops.cpp b/ggml/src/ggml-cpu/ops.cpp
-index 6050147b..66b8da68 100644
+index 7413192b..becdae07 100644
 --- a/ggml/src/ggml-cpu/ops.cpp
 +++ b/ggml/src/ggml-cpu/ops.cpp
-@@ -6531,6 +6531,61 @@ void ggml_compute_forward_pad_reflect_1d(
+@@ -6703,6 +6703,61 @@ void ggml_compute_forward_pad_reflect_1d(
      }
  }
  
@@ -135,10 +135,10 @@ index 6050147b..66b8da68 100644
  
  static void ggml_compute_forward_arange_f32(
 diff --git a/ggml/src/ggml-cpu/ops.h b/ggml/src/ggml-cpu/ops.h
-index 410a3720..3eca1cf8 100644
+index dc081b9e..a7125555 100644
 --- a/ggml/src/ggml-cpu/ops.h
 +++ b/ggml/src/ggml-cpu/ops.h
-@@ -71,6 +71,7 @@ void ggml_compute_forward_pool_2d_back(const struct ggml_compute_params * params
+@@ -72,6 +72,7 @@ void ggml_compute_forward_pool_2d_back(const struct ggml_compute_params * params
  void ggml_compute_forward_upscale(const struct ggml_compute_params * params, struct ggml_tensor * dst);
  void ggml_compute_forward_pad(const struct ggml_compute_params * params, struct ggml_tensor * dst);
  void ggml_compute_forward_pad_reflect_1d(const struct ggml_compute_params * params, struct ggml_tensor * dst);
@@ -147,10 +147,10 @@ index 410a3720..3eca1cf8 100644
  void ggml_compute_forward_timestep_embedding(const struct ggml_compute_params * params, struct ggml_tensor * dst);
  void ggml_compute_forward_argsort(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu
-index 31750b6f..0fef9522 100644
+index 04ce764e..491acccb 100644
 --- a/ggml/src/ggml-cuda/ggml-cuda.cu
 +++ b/ggml/src/ggml-cuda/ggml-cuda.cu
-@@ -2246,6 +2246,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
+@@ -2223,6 +2223,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
          case GGML_OP_PAD:
              ggml_cuda_op_pad(ctx, dst);
              break;
@@ -160,7 +160,7 @@ index 31750b6f..0fef9522 100644
          case GGML_OP_ARANGE:
              ggml_cuda_op_arange(ctx, dst);
              break;
-@@ -3222,6 +3225,7 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
+@@ -3197,6 +3200,7 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
          case GGML_OP_UPSCALE:
              return op->src[0]->type == GGML_TYPE_F32 && op->op_params[0] == GGML_SCALE_MODE_NEAREST;
          case GGML_OP_PAD:
@@ -233,7 +233,7 @@ index 8fd386b0..e2ededc3 100644
  void ggml_cuda_op_pad(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 +void ggml_cuda_op_unpad(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 diff --git a/ggml/src/ggml-metal/ggml-metal.m b/ggml/src/ggml-metal/ggml-metal.m
-index 12886cd3..b2e95a66 100644
+index 425524d0..112abef6 100644
 --- a/ggml/src/ggml-metal/ggml-metal.m
 +++ b/ggml/src/ggml-metal/ggml-metal.m
 @@ -341,6 +341,7 @@ static void ggml_backend_metal_device_rel(struct ggml_backend_metal_device_conte
@@ -244,7 +244,7 @@ index 12886cd3..b2e95a66 100644
      GGML_METAL_KERNEL_TYPE_ARANGE_F32,
      GGML_METAL_KERNEL_TYPE_TIMESTEP_EMBEDDING_F32,
      GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_ASC,
-@@ -1020,6 +1021,7 @@ @implementation GGMLMetalClass
+@@ -1277,6 +1278,7 @@ @implementation GGMLMetalClass
          GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_UPSCALE_F32,                     upscale_f32,                     true);
          GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_PAD_F32,                         pad_f32,                         true);
          GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_PAD_REFLECT_1D_F32,              pad_reflect_1d_f32,              true);
@@ -252,7 +252,7 @@ index 12886cd3..b2e95a66 100644
          GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_TIMESTEP_EMBEDDING_F32,          timestep_embedding_f32,          true);
          GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARANGE_F32,                      arange_f32,                      true);
          GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_ASC,             argsort_f32_i32_asc,             true);
-@@ -1384,6 +1386,7 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_device_contex
+@@ -1647,6 +1649,7 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_device_contex
          case GGML_OP_POOL_2D:
          case GGML_OP_PAD:
          case GGML_OP_PAD_REFLECT_1D:
@@ -260,7 +260,7 @@ index 12886cd3..b2e95a66 100644
          case GGML_OP_TIMESTEP_EMBEDDING:
          case GGML_OP_ARGSORT:
          case GGML_OP_LEAKY_RELU:
-@@ -3731,6 +3734,36 @@ static void ggml_metal_encode_node(
+@@ -4047,6 +4050,36 @@ static bool ggml_metal_encode_node(
  
                  const int nth = MIN(1024, ne0);
  
@@ -298,7 +298,7 @@ index 12886cd3..b2e95a66 100644
              } break;
          case GGML_OP_ARANGE:
 diff --git a/ggml/src/ggml-metal/ggml-metal.metal b/ggml/src/ggml-metal/ggml-metal.metal
-index 8d6e99e6..71f0f97f 100644
+index 9f4147e9..6ceb3cef 100644
 --- a/ggml/src/ggml-metal/ggml-metal.metal
 +++ b/ggml/src/ggml-metal/ggml-metal.metal
 @@ -2975,6 +2975,51 @@ kernel void kernel_pad_reflect_1d_f32(
@@ -354,10 +354,10 @@ index 8d6e99e6..71f0f97f 100644
      device        char * dst,
      constant   ggml_metal_kargs_arange & args,
 diff --git a/ggml/src/ggml.c b/ggml/src/ggml.c
-index 950772c7..2276b631 100644
+index 7654ae17..3c57aff8 100644
 --- a/ggml/src/ggml.c
 +++ b/ggml/src/ggml.c
-@@ -963,6 +963,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
+@@ -923,6 +923,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
      "UPSCALE",
      "PAD",
      "PAD_REFLECT_1D",
@@ -365,16 +365,16 @@ index 950772c7..2276b631 100644
      "ARANGE",
      "TIMESTEP_EMBEDDING",
      "ARGSORT",
-@@ -993,7 +994,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
+@@ -953,7 +954,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
      "OPT_STEP_ADAMW",
  };
  
--static_assert(GGML_OP_COUNT == 81, "GGML_OP_COUNT != 81");
-+static_assert(GGML_OP_COUNT == 82, "GGML_OP_COUNT != 82");
+-static_assert(GGML_OP_COUNT == 82, "GGML_OP_COUNT != 82");
++static_assert(GGML_OP_COUNT == 83, "GGML_OP_COUNT != 83");
  
  static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
      "none",
-@@ -1057,6 +1058,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
+@@ -1018,6 +1019,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
      "upscale(x)",
      "pad(x)",
      "pad_reflect_1d(x)",
@@ -382,16 +382,16 @@ index 950772c7..2276b631 100644
      "arange(start, stop, step)",
      "timestep_embedding(timesteps, dim, max_period)",
      "argsort(x)",
-@@ -1087,7 +1089,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
+@@ -1048,7 +1050,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
      "adamw(x)",
  };
  
--static_assert(GGML_OP_COUNT == 81, "GGML_OP_COUNT != 81");
-+static_assert(GGML_OP_COUNT == 82, "GGML_OP_COUNT != 82");
+-static_assert(GGML_OP_COUNT == 82, "GGML_OP_COUNT != 82");
++static_assert(GGML_OP_COUNT == 83, "GGML_OP_COUNT != 83");
  
  static_assert(GGML_OP_POOL_COUNT == 2, "GGML_OP_POOL_COUNT != 2");
  
-@@ -4262,6 +4264,25 @@ struct ggml_tensor * ggml_pad_reflect_1d(
+@@ -4270,6 +4272,25 @@ struct ggml_tensor * ggml_pad_reflect_1d(
      return result;
  }
  

llama/patches/0008-fix-deepseek-deseret-regex.patch

@@ -12,7 +12,7 @@ regex
  2 files changed, 22 insertions(+), 1 deletion(-)
 
 diff --git a/src/llama-vocab.cpp b/src/llama-vocab.cpp
-index a35b498c..032019c9 100644
+index a9ee9f03..1306864e 100644
 --- a/src/llama-vocab.cpp
 +++ b/src/llama-vocab.cpp
 @@ -296,7 +296,7 @@ struct llm_tokenizer_bpe : llm_tokenizer {