ggml update to b6840 (#12791)

Makefile.sync

 UPSTREAM=https://github.com/ggml-org/llama.cpp.git
 WORKDIR=llama/vendor
-FETCH_HEAD=7049736b2dd9011bf819e298b844ebbc4b5afdc9
+FETCH_HEAD=3cfa9c3f125763305b4226bc032f1954f08990dc
 
 .PHONY: help
 help:

llama/build-info.cpp

 int LLAMA_BUILD_NUMBER = 0;
-char const *LLAMA_COMMIT = "7049736b2dd9011bf819e298b844ebbc4b5afdc9";
+char const *LLAMA_COMMIT = "3cfa9c3f125763305b4226bc032f1954f08990dc";
 char const *LLAMA_COMPILER = "";
 char const *LLAMA_BUILD_TARGET = "";

llama/llama.cpp/common/json-schema-to-grammar.cpp

@@ -41,9 +41,9 @@ static std::string build_repetition(const std::string & item_rule, int min_items
     return result;
 }
 
-static void _build_min_max_int(int min_value, int max_value, std::stringstream & out, int decimals_left = 16, bool top_level = true) {
-    auto has_min = min_value != std::numeric_limits<int>::min();
-    auto has_max = max_value != std::numeric_limits<int>::max();
+static void _build_min_max_int(int64_t min_value, int64_t max_value, std::stringstream & out, int decimals_left = 16, bool top_level = true) {
+    auto has_min = min_value != std::numeric_limits<int64_t>::min();
+    auto has_max = max_value != std::numeric_limits<int64_t>::max();
 
     auto digit_range = [&](char from, char to) {
         out << "[";
@@ -159,7 +159,7 @@ static void _build_min_max_int(int min_value, int max_value, std::stringstream &
     if (has_min) {
         if (min_value < 0) {
             out << "\"-\" (";
-            _build_min_max_int(std::numeric_limits<int>::min(), -min_value, out, decimals_left, /* top_level= */ false);
+            _build_min_max_int(std::numeric_limits<int64_t>::min(), -min_value, out, decimals_left, /* top_level= */ false);
             out << ") | [0] | [1-9] ";
             more_digits(0, decimals_left - 1);
         } else if (min_value == 0) {
@@ -194,7 +194,7 @@ static void _build_min_max_int(int min_value, int max_value, std::stringstream &
             }
             digit_range(c, c);
             out << " (";
-            _build_min_max_int(std::stoi(min_s.substr(1)), std::numeric_limits<int>::max(), out, less_decimals, /* top_level= */ false);
+            _build_min_max_int(std::stoll(min_s.substr(1)), std::numeric_limits<int64_t>::max(), out, less_decimals, /* top_level= */ false);
             out << ")";
             if (c < '9') {
                 out << " | ";
@@ -216,7 +216,7 @@ static void _build_min_max_int(int min_value, int max_value, std::stringstream &
             _build_min_max_int(0, max_value, out, decimals_left, /* top_level= */ true);
         } else {
             out << "\"-\" (";
-            _build_min_max_int(-max_value, std::numeric_limits<int>::max(), out, decimals_left, /* top_level= */ false);
+            _build_min_max_int(-max_value, std::numeric_limits<int64_t>::max(), out, decimals_left, /* top_level= */ false);
             out << ")";
         }
         return;
@@ -925,17 +925,17 @@ public:
             int max_len = schema.contains("maxLength") ? schema["maxLength"].get<int>() : std::numeric_limits<int>::max();
             return _add_rule(rule_name, "\"\\\"\" " + build_repetition(char_rule, min_len, max_len) + " \"\\\"\" space");
         } else if (schema_type == "integer" && (schema.contains("minimum") || schema.contains("exclusiveMinimum") || schema.contains("maximum") || schema.contains("exclusiveMaximum"))) {
-            int min_value = std::numeric_limits<int>::min();
-            int max_value = std::numeric_limits<int>::max();
+            int64_t min_value = std::numeric_limits<int64_t>::min();
+            int64_t max_value = std::numeric_limits<int64_t>::max();
             if (schema.contains("minimum")) {
-                min_value = schema["minimum"].get<int>();
+                min_value = schema["minimum"].get<int64_t>();
             } else if (schema.contains("exclusiveMinimum")) {
-                min_value = schema["exclusiveMinimum"].get<int>() + 1;
+                min_value = schema["exclusiveMinimum"].get<int64_t>() + 1;
             }
             if (schema.contains("maximum")) {
-                max_value = schema["maximum"].get<int>();
+                max_value = schema["maximum"].get<int64_t>();
             } else if (schema.contains("exclusiveMaximum")) {
-                max_value = schema["exclusiveMaximum"].get<int>() - 1;
+                max_value = schema["exclusiveMaximum"].get<int64_t>() - 1;
             }
             std::stringstream out;
             out << "(";

llama/llama.cpp/src/llama-arch.cpp

@@ -5,6 +5,7 @@
 #include <map>
 
 static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
+    { LLM_ARCH_CLIP,             "clip"             }, // dummy, only used by llama-quantize
     { LLM_ARCH_LLAMA,            "llama"            },
     { LLM_ARCH_LLAMA4,           "llama4"           },
     { LLM_ARCH_DECI,             "deci"             },
@@ -85,6 +86,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_WAVTOKENIZER_DEC, "wavtokenizer-dec" },
     { LLM_ARCH_PLM,              "plm"              },
     { LLM_ARCH_BAILINGMOE,       "bailingmoe"       },
+    { LLM_ARCH_BAILINGMOE2,      "bailingmoe2"      },
     { LLM_ARCH_DOTS1,            "dots1"            },
     { LLM_ARCH_ARCEE,            "arcee"            },
     { LLM_ARCH_ERNIE4_5,         "ernie4_5"         },
@@ -135,6 +137,8 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_EXPERT_COUNT,                      "%s.expert_count"                      },
     { LLM_KV_EXPERT_USED_COUNT,                 "%s.expert_used_count"                 },
     { LLM_KV_EXPERT_SHARED_COUNT,               "%s.expert_shared_count"               },
+    { LLM_KV_EXPERT_GROUP_COUNT,                "%s.expert_group_count"                },
+    { LLM_KV_EXPERT_GROUP_USED_COUNT,           "%s.expert_group_used_count"           },
     { LLM_KV_EXPERT_WEIGHTS_SCALE,              "%s.expert_weights_scale"              },
     { LLM_KV_EXPERT_WEIGHTS_NORM,               "%s.expert_weights_norm"               },
     { LLM_KV_EXPERT_GATING_FUNC,                "%s.expert_gating_func"                },
@@ -277,6 +281,10 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
 };
 
 static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_NAMES = {
+    {
+        LLM_ARCH_CLIP,
+        {},
+    },
     {
         LLM_ARCH_LLAMA,
         {
@@ -1961,6 +1969,38 @@ 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" },
         },
     },
+    {
+        LLM_ARCH_BAILINGMOE2,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,         "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,        "output_norm" },
+            { LLM_TENSOR_OUTPUT,             "output" },
+            { LLM_TENSOR_ATTN_NORM,          "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q_NORM,        "blk.%d.attn_q_norm" },
+            { LLM_TENSOR_ATTN_K_NORM,        "blk.%d.attn_k_norm" },
+            { LLM_TENSOR_ATTN_QKV,           "blk.%d.attn_qkv" },
+            { LLM_TENSOR_ATTN_OUT,           "blk.%d.attn_output" },
+            { LLM_TENSOR_FFN_GATE_INP,       "blk.%d.ffn_gate_inp" },
+            { LLM_TENSOR_FFN_EXP_PROBS_B,    "blk.%d.exp_probs_b" },
+            { LLM_TENSOR_FFN_NORM,           "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE,           "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_DOWN,           "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,             "blk.%d.ffn_up" },
+            { LLM_TENSOR_FFN_GATE_EXPS,      "blk.%d.ffn_gate_exps" },
+            { LLM_TENSOR_FFN_DOWN_EXPS,      "blk.%d.ffn_down_exps" },
+            { LLM_TENSOR_FFN_UP_EXPS,        "blk.%d.ffn_up_exps" },
+            { LLM_TENSOR_FFN_GATE_SHEXP,     "blk.%d.ffn_gate_shexp" },
+            { LLM_TENSOR_FFN_DOWN_SHEXP,     "blk.%d.ffn_down_shexp" },
+            { LLM_TENSOR_FFN_UP_SHEXP,       "blk.%d.ffn_up_shexp" },
+            { LLM_TENSOR_NEXTN_EH_PROJ,      "blk.%d.nextn.eh_proj" },
+            { LLM_TENSOR_NEXTN_EMBED_TOKENS, "blk.%d.nextn.embed_tokens" },
+            { LLM_TENSOR_NEXTN_ENORM,        "blk.%d.nextn.enorm" },
+            { LLM_TENSOR_NEXTN_HNORM,        "blk.%d.nextn.hnorm" },
+            { LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD, "blk.%d.nextn.shared_head_head" },
+            { LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, "blk.%d.nextn.shared_head_norm" },
+            { LLM_TENSOR_LAYER_OUT_NORM,     "blk.%d.layer_output_norm" },
+        },
+    },
     {
         LLM_ARCH_DOTS1,
         {

llama/llama.cpp/src/llama-arch.h

@@ -9,6 +9,7 @@
 //
 
 enum llm_arch {
+    LLM_ARCH_CLIP,
     LLM_ARCH_LLAMA,
     LLM_ARCH_LLAMA4,
     LLM_ARCH_DECI,
@@ -89,6 +90,7 @@ enum llm_arch {
     LLM_ARCH_WAVTOKENIZER_DEC,
     LLM_ARCH_PLM,
     LLM_ARCH_BAILINGMOE,
+    LLM_ARCH_BAILINGMOE2,
     LLM_ARCH_DOTS1,
     LLM_ARCH_ARCEE,
     LLM_ARCH_ERNIE4_5,
@@ -139,6 +141,8 @@ enum llm_kv {
     LLM_KV_EXPERT_COUNT,
     LLM_KV_EXPERT_USED_COUNT,
     LLM_KV_EXPERT_SHARED_COUNT,
+    LLM_KV_EXPERT_GROUP_COUNT,
+    LLM_KV_EXPERT_GROUP_USED_COUNT,
     LLM_KV_EXPERT_WEIGHTS_SCALE,
     LLM_KV_EXPERT_WEIGHTS_NORM,
     LLM_KV_EXPERT_GATING_FUNC,

llama/llama.cpp/src/llama-batch.h

@@ -123,7 +123,7 @@ private:
     uint32_t n_seq_max;
     uint32_t n_outputs;
 
-    std::array<llama_seq_id, 1> seq_id_0 = { 0 }; // default sequence id
+    std::array<llama_seq_id, 1> seq_id_0 = {{ 0 }}; // default sequence id
 
     std::vector<llama_pos>      pos;
     std::vector<int32_t>        n_seq_id;

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

@@ -63,6 +63,8 @@ static const std::map<std::string, llm_chat_template> LLM_CHAT_TEMPLATES = {
     { "megrez",            LLM_CHAT_TEMPLATE_MEGREZ            },
     { "yandex",            LLM_CHAT_TEMPLATE_YANDEX            },
     { "bailing",           LLM_CHAT_TEMPLATE_BAILING           },
+    { "bailing-think",     LLM_CHAT_TEMPLATE_BAILING_THINK     },
+    { "bailing2",          LLM_CHAT_TEMPLATE_BAILING2          },
     { "llama4",            LLM_CHAT_TEMPLATE_LLAMA4            },
     { "smolvlm",           LLM_CHAT_TEMPLATE_SMOLVLM           },
     { "hunyuan-moe",       LLM_CHAT_TEMPLATE_HUNYUAN_MOE       },
@@ -191,6 +193,10 @@ llm_chat_template llm_chat_detect_template(const std::string & tmpl) {
         return LLM_CHAT_TEMPLATE_YANDEX;
     } else if (tmpl_contains("<role>ASSISTANT</role>") && tmpl_contains("'HUMAN'")) {
         return LLM_CHAT_TEMPLATE_BAILING;
+    } else if (tmpl_contains("<role>ASSISTANT</role>") && tmpl_contains("\"HUMAN\"") && tmpl_contains("<think>")) {
+        return LLM_CHAT_TEMPLATE_BAILING_THINK;
+    } else if (tmpl_contains("<role>ASSISTANT</role>") && tmpl_contains("<role>HUMAN</role>") && tmpl_contains("<|role_end|>")) {
+        return LLM_CHAT_TEMPLATE_BAILING2;
     } else if (tmpl_contains("<|header_start|>") && tmpl_contains("<|header_end|>")) {
         return LLM_CHAT_TEMPLATE_LLAMA4;
     } else if (tmpl_contains("<|endofuserprompt|>")) {
@@ -644,8 +650,8 @@ int32_t llm_chat_apply_template(
         if (add_ass) {
             ss << " Ассистент:[SEP]";
         }
-    }  else if (tmpl == LLM_CHAT_TEMPLATE_BAILING) {
-        // Bailing (Ling) template
+    } else if (tmpl == LLM_CHAT_TEMPLATE_BAILING || tmpl == LLM_CHAT_TEMPLATE_BAILING_THINK) {
+        // Bailing (Ling/Ring) template
         for (auto message : chat) {
             std::string role(message->role);
 
@@ -658,6 +664,33 @@ int32_t llm_chat_apply_template(
             ss << "<role>" << role << "</role>" << message->content;
         }
 
+        if (add_ass) {
+            ss << "<role>ASSISTANT</role>";
+
+            if (tmpl == LLM_CHAT_TEMPLATE_BAILING_THINK) {
+                ss << "<think>";
+            }
+        }
+    } else if (tmpl == LLM_CHAT_TEMPLATE_BAILING2) {
+        // Bailing2 (Ling 2.0) template
+        bool has_system = !chat.empty() && std::string(chat[0]->role) == "system";
+
+        if (!has_system) {
+            ss << "<role>SYSTEM</role>detailed thinking off<|role_end|>";
+        }
+
+        for (auto message : chat) {
+            std::string role(message->role);
+
+            if (role == "user") {
+                role = "HUMAN";
+            } else {
+                std::transform(role.begin(), role.end(), role.begin(), ::toupper);
+            }
+
+            ss << "<role>" << role << "</role>" << message->content << "<|role_end|>";
+        }
+
         if (add_ass) {
             ss << "<role>ASSISTANT</role>";
         }

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

@@ -42,6 +42,8 @@ enum llm_chat_template {
     LLM_CHAT_TEMPLATE_MEGREZ,
     LLM_CHAT_TEMPLATE_YANDEX,
     LLM_CHAT_TEMPLATE_BAILING,
+    LLM_CHAT_TEMPLATE_BAILING_THINK,
+    LLM_CHAT_TEMPLATE_BAILING2,
     LLM_CHAT_TEMPLATE_LLAMA4,
     LLM_CHAT_TEMPLATE_SMOLVLM,
     LLM_CHAT_TEMPLATE_DOTS1,

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

@@ -2345,7 +2345,8 @@ llama_context * llama_init_from_model(
         return nullptr;
     }
 
-    if (params.pooling_type != model->hparams.pooling_type) {
+    if (params.pooling_type != LLAMA_POOLING_TYPE_UNSPECIFIED &&
+        params.pooling_type != model->hparams.pooling_type) {
         //user-specified pooling-type is different from the model default
         LLAMA_LOG_WARN("%s: model default pooling_type is [%d], but [%d] was specified\n", __func__,
                        model->hparams.pooling_type, params.pooling_type);

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

@@ -261,12 +261,17 @@ void llm_graph_input_cross_embd::set_input(const llama_ubatch * ubatch) {
     }
 }
 
-static void print_mask(float * data, int64_t n_tokens, int64_t n_kv, int64_t n_swa, llama_swa_type swa_type) {
+static void print_mask(const float * data, int64_t n_tokens, int64_t n_kv, int64_t n_swa, llama_swa_type swa_type) {
     LLAMA_LOG_DEBUG("%s: === Attention mask ===\n", __func__);
-    const char * swa_type_str = (swa_type == LLAMA_SWA_TYPE_NONE) ? "LLAMA_SWA_TYPE_NONE" :
-                          (swa_type == LLAMA_SWA_TYPE_STANDARD) ? "LLAMA_SWA_TYPE_STANDARD" :
-                          (swa_type == LLAMA_SWA_TYPE_CHUNKED) ? "LLAMA_SWA_TYPE_CHUNKED" :
-                          (swa_type == LLAMA_SWA_TYPE_SYMMETRIC) ? "LLAMA_SWA_TYPE_SYMMETRIC" : "unknown";
+    const char * swa_type_str = "unknown";
+
+    switch (swa_type) {
+        case LLAMA_SWA_TYPE_NONE:      swa_type_str = "LLAMA_SWA_TYPE_NONE"; break;
+        case LLAMA_SWA_TYPE_STANDARD:  swa_type_str = "LLAMA_SWA_TYPE_STANDARD"; break;
+        case LLAMA_SWA_TYPE_CHUNKED:   swa_type_str = "LLAMA_SWA_TYPE_CHUNKED"; break;
+        case LLAMA_SWA_TYPE_SYMMETRIC: swa_type_str = "LLAMA_SWA_TYPE_SYMMETRIC"; break;
+    };
+
     LLAMA_LOG_DEBUG("%s: n_swa : %d, n_kv: %d, swq_type: %s\n", __func__, (int)n_swa, (int)n_kv, swa_type_str);
     LLAMA_LOG_DEBUG("%s: '0' = can attend, '∞' = masked\n", __func__);
     LLAMA_LOG_DEBUG("%s: Rows = query tokens, Columns = key/value tokens\n\n", __func__);
@@ -295,51 +300,68 @@ void llm_graph_input_attn_no_cache::set_input(const llama_ubatch * ubatch) {
     const int64_t n_kv     = ubatch->n_tokens;
     const int64_t n_tokens = ubatch->n_tokens;
 
-    GGML_ASSERT(kq_mask);
-    GGML_ASSERT(ggml_backend_buffer_is_host(kq_mask->buffer));
-
-    float * data = (float *) kq_mask->data;
-
-    // [TAG_NO_CACHE_ISWA]
-    GGML_ASSERT(hparams.swa_type == LLAMA_SWA_TYPE_NONE && "TODO: implement");
-
+    const auto fill_mask = [&](float * data, int n_swa, llama_swa_type swa_type) {
         for (int h = 0; h < 1; ++h) {
             for (int i1 = 0; i1 < n_tokens; ++i1) {
                 const llama_seq_id s1 = ubatch->seq_id[i1][0];
+                const llama_pos    p1 = ubatch->pos[i1];
 
-            for (int i0 = 0; i0 < n_tokens; ++i0) {
-                float f = -INFINITY;
+                const uint64_t idst = h*(n_kv*n_tokens) + i1*n_kv;
 
-                for (int s = 0; s < ubatch->n_seq_id[i0]; ++s) {
+                for (int i0 = 0; i0 < n_tokens; ++i0) {
                     const llama_seq_id s0 = ubatch->seq_id[i0][0];
+                    const llama_pos p0    = ubatch->pos[i0];
 
+                    // mask different sequences
                     if (s0 != s1) {
-                        continue; // skip different sequences
+                        continue;
                     }
 
-                    if (cparams.causal_attn && ubatch->pos[i0] > ubatch->pos[i1]) {
-                        continue; // skip future tokens for causal attention
+                    // mask future tokens
+                    if (cparams.causal_attn && p0 > p1) {
+                        continue;
                     }
 
-                    // TODO: this does not take into account that some layers are SWA and others are note (i.e. iSWA) [TAG_NO_CACHE_ISWA]
-                    //if (hparams.is_masked_swa(ubatch->pos[i0], ubatch->pos[i1])) {
-                    //    continue; // skip masked tokens for SWA
-                    //}
+                    // apply SWA if any
+                    if (llama_hparams::is_masked_swa(n_swa, swa_type, p0, p1)) {
+                        continue;
+                    }
 
-                    // TODO: reimplement this like in llama_kv_cache_unified
-                    if (hparams.use_alibi) {
-                        f = -std::abs(ubatch->pos[i0] - ubatch->pos[i1]);
-                    } else {
-                        f = 0.0f;
+                    data[idst + i0] = hparams.use_alibi ? -std::abs(p0 - p1) : 0.0f;
                 }
             }
-                data[h*(n_kv*n_tokens) + i1*n_kv + i0] = f;
         }
+    };
+
+    {
+        GGML_ASSERT(self_kq_mask);
+        GGML_ASSERT(ggml_backend_buffer_is_host(self_kq_mask->buffer));
+
+        float * data = (float *) self_kq_mask->data;
+
+        std::fill(data, data + ggml_nelements(self_kq_mask), -INFINITY);
+
+        fill_mask(data, 0, LLAMA_SWA_TYPE_NONE);
+
+        if (debug) {
+            print_mask(data, n_tokens, n_kv, 0, LLAMA_SWA_TYPE_NONE);
         }
     }
+
+    if (hparams.swa_type != LLAMA_SWA_TYPE_NONE) {
+        GGML_ASSERT(self_kq_mask_swa);
+        GGML_ASSERT(ggml_backend_buffer_is_host(self_kq_mask_swa->buffer));
+
+        float * data = (float *) self_kq_mask_swa->data;
+
+        std::fill(data, data + ggml_nelements(self_kq_mask_swa), -INFINITY);
+
+        fill_mask(data, hparams.n_swa, hparams.swa_type);
+
         if (debug) {
             print_mask(data, n_tokens, n_kv, hparams.n_swa, hparams.swa_type);
         }
+    }
 }
 
 void llm_graph_input_attn_kv::set_input(const llama_ubatch * ubatch) {
@@ -928,6 +950,31 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
         cb(selection_probs, "ffn_moe_probs_biased", il);
     }
 
+    // select top n_group_used expert groups
+    // https://huggingface.co/deepseek-ai/DeepSeek-V3/blob/e815299b0bcbac849fa540c768ef21845365c9eb/modeling_deepseek.py#L440-L457
+    if (hparams.n_expert_groups > 1 && n_tokens > 0) {
+        const int64_t n_exp_per_group = n_expert / hparams.n_expert_groups;
+
+        // organize experts into n_expert_groups
+        ggml_tensor * selection_groups = ggml_reshape_3d(ctx0, selection_probs, n_exp_per_group, hparams.n_expert_groups, n_tokens); // [n_exp_per_group, n_expert_groups, n_tokens]
+
+        ggml_tensor * group_scores = ggml_top_k(ctx0, selection_groups, 2); // [2, n_expert_groups, n_tokens]
+        group_scores = ggml_get_rows(ctx0, ggml_reshape_4d(ctx0, selection_groups, 1, selection_groups->ne[0], selection_groups->ne[1], selection_groups->ne[2]), group_scores); // [1, 2, n_expert_groups, n_tokens]
+
+        // get top n_group_used expert groups
+        group_scores = ggml_sum_rows(ctx0, ggml_reshape_3d(ctx0, group_scores, group_scores->ne[1], group_scores->ne[2], group_scores->ne[3])); // [1, n_expert_groups, n_tokens]
+        group_scores = ggml_reshape_2d(ctx0, group_scores, group_scores->ne[1], group_scores->ne[2]); // [n_expert_groups, n_tokens]
+
+        ggml_tensor * expert_groups = ggml_top_k(ctx0, group_scores, hparams.n_group_used); // [n_group_used, n_tokens]
+        cb(expert_groups, "ffn_moe_group_topk", il);
+
+        // mask out the other groups
+        selection_probs = ggml_get_rows(ctx0, selection_groups, expert_groups); // [n_exp_per_group, n_group_used, n_tokens]
+        selection_probs = ggml_set_rows(ctx0, ggml_scale_bias(ctx0, selection_groups, 0.0f, -INFINITY), selection_probs, expert_groups); // [n_exp_per_group, n_expert_groups, n_tokens]
+        selection_probs = ggml_reshape_2d(ctx0, selection_probs, n_expert, n_tokens); // [n_expert, n_tokens]
+        cb(selection_probs, "ffn_moe_probs_masked", il);
+    }
+
     // select experts
     ggml_tensor * selected_experts = ggml_top_k(ctx0, selection_probs, n_expert_used); // [n_expert_used, n_tokens]
     cb(selected_experts->src[0], "ffn_moe_argsort", il);
@@ -959,6 +1006,11 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
         ggml_tensor * weights_sum = ggml_sum_rows(ctx0, weights); // [1, n_tokens]
         cb(weights_sum, "ffn_moe_weights_sum", il);
 
+        if (arch == LLM_ARCH_BAILINGMOE2) {
+            weights_sum = ggml_scale_bias(ctx0, weights_sum, 1.0, 1e-20);
+            cb(weights_sum, "ffn_moe_weights_sum_biased", il);
+        }
+
         weights = ggml_div(ctx0, weights, weights_sum); // [n_expert_used, n_tokens]
         cb(weights, "ffn_moe_weights_norm", il);
 
@@ -1299,12 +1351,9 @@ ggml_tensor * llm_graph_context::build_attn_mha(
     k = ggml_permute(ctx0, k, 0, 2, 1, 3);
     v = ggml_permute(ctx0, v, 0, 2, 1, 3);
 
-    const auto n_kv = k->ne[1];
-
     ggml_tensor * cur;
 
-    // TODO: replace hardcoded padding with ggml-provided padding
-    if (cparams.flash_attn && (n_kv % 256 == 0) && kq_b == nullptr) {
+    if (cparams.flash_attn && kq_b == nullptr) {
         GGML_ASSERT(kq_b == nullptr && "Flash attention does not support KQ bias yet");
 
         if (v_trans) {
@@ -1419,10 +1468,20 @@ llm_graph_input_attn_no_cache * llm_graph_context::build_attn_inp_no_cache() con
     auto inp = std::make_unique<llm_graph_input_attn_no_cache>(hparams, cparams);
 
     // note: there is no KV cache, so the number of KV values is equal to the number of tokens in the batch
-    inp->kq_mask = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_tokens, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD), 1, 1);
-    ggml_set_input(inp->kq_mask);
+    inp->self_kq_mask = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_tokens, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD), 1, 1);
+    ggml_set_input(inp->self_kq_mask);
 
-    inp->kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->kq_mask, GGML_TYPE_F16) : inp->kq_mask;
+    inp->self_kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask, GGML_TYPE_F16) : inp->self_kq_mask;
+
+    if (hparams.swa_type != LLAMA_SWA_TYPE_NONE) {
+        inp->self_kq_mask_swa = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_tokens, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD), 1, 1);
+        ggml_set_input(inp->self_kq_mask_swa);
+
+        inp->self_kq_mask_swa_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask_swa, GGML_TYPE_F16) : inp->self_kq_mask_swa;
+    } else {
+        inp->self_kq_mask_swa     = nullptr;
+        inp->self_kq_mask_swa_cnv = nullptr;
+    }
 
     return (llm_graph_input_attn_no_cache *) res->add_input(std::move(inp));
 }
@@ -1447,7 +1506,9 @@ ggml_tensor * llm_graph_context::build_attn(
     ggml_build_forward_expand(gf, k_cur);
     ggml_build_forward_expand(gf, v_cur);
 
-    const auto & kq_mask = inp->get_kq_mask();
+    const bool is_swa = hparams.is_swa(il);
+
+    const auto & kq_mask = is_swa ? inp->get_kq_mask_swa() : inp->get_kq_mask();
 
     // [TAG_NO_CACHE_PAD]
     // TODO: if ubatch.equal_seqs() == true, we can split the three tensors below into ubatch.n_seqs_unq streams

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

@@ -257,10 +257,14 @@ public:
 
     void set_input(const llama_ubatch * ubatch) override;
 
-    ggml_tensor * get_kq_mask() const { return kq_mask_cnv; }
+    ggml_tensor * get_kq_mask()     const { return self_kq_mask_cnv; }
+    ggml_tensor * get_kq_mask_swa() const { return self_kq_mask_swa_cnv; }
 
-    ggml_tensor * kq_mask     = nullptr; // F32 [n_tokens, n_batch, 1, 1]
-    ggml_tensor * kq_mask_cnv = nullptr; //     [n_tokens, n_batch, 1, 1]
+    // n_tokens == n_batch
+    ggml_tensor * self_kq_mask         = nullptr; // F32 [n_tokens, n_batch/n_stream, 1, n_stream]
+    ggml_tensor * self_kq_mask_cnv     = nullptr; //     [n_tokens, n_batch/n_stream, 1, n_stream]
+    ggml_tensor * self_kq_mask_swa     = nullptr; // F32 [n_tokens, n_batch/n_stream, 1, n_stream]
+    ggml_tensor * self_kq_mask_swa_cnv = nullptr; //     [n_tokens, n_batch/n_stream, 1, n_stream]
 
     const llama_hparams hparams;
     const llama_cparams cparams;

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

@@ -74,6 +74,8 @@ struct llama_hparams {
     uint32_t n_ff_chexp         = 0;
     uint32_t n_expert_shared    = 0;
     uint32_t n_norm_groups      = 0;
+    uint32_t n_expert_groups    = 0;
+    uint32_t n_group_used       = 0;
     uint32_t n_group_experts    = 0;
 
     float    expert_group_scale   = 0.05f;

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

@@ -114,9 +114,12 @@ const char * llm_type_name(llm_type type) {
         case LLM_TYPE_17B_16E:       return "17Bx16E (Scout)";
         case LLM_TYPE_17B_128E:      return "17Bx128E (Maverick)";
         case LLM_TYPE_A13B:          return "A13B";
+        case LLM_TYPE_7B_A1B:        return "7B.A1B";
         case LLM_TYPE_8B_A1B:        return "8B.A1B";
+        case LLM_TYPE_16B_A1B:       return "16B.A1B";
         case LLM_TYPE_21B_A3B:       return "21B.A3B";
         case LLM_TYPE_30B_A3B:       return "30B.A3B";
+        case LLM_TYPE_100B_A6B:      return "100B.A6B";
         case LLM_TYPE_106B_A12B:     return "106B.A12B";
         case LLM_TYPE_235B_A22B:     return "235B.A22B";
         case LLM_TYPE_300B_A47B:     return "300B.A47B";
@@ -401,6 +404,19 @@ static buft_list_t make_gpu_buft_list(ggml_backend_dev_t dev, llama_split_mode s
     // add the device default buffer type
     buft_list.emplace_back(dev, ggml_backend_dev_buffer_type(dev));
+    // add the device extra buffer type (if any)
+    ggml_backend_reg_t reg = ggml_backend_dev_backend_reg(dev);
+    auto ggml_backend_dev_get_extra_bufts_fn = (ggml_backend_dev_get_extra_bufts_t)
+        ggml_backend_reg_get_proc_address(reg, "ggml_backend_dev_get_extra_bufts");
+
+    if (ggml_backend_dev_get_extra_bufts_fn) {
+        ggml_backend_buffer_type_t * extra_bufts = ggml_backend_dev_get_extra_bufts_fn(dev);
+        while (extra_bufts && *extra_bufts) {
+            buft_list.emplace_back(dev, *extra_bufts);
+            ++extra_bufts;
+        }
+    }
+
     return buft_list;
 }
@@ -421,11 +437,8 @@ struct llama_model::impl {
     llama_mlocks mlock_bufs;
     llama_mlocks mlock_mmaps;
-    // contexts where the model tensors metadata is stored
-    std::vector<ggml_context_ptr> ctxs;
-
-    // the model memory buffers for the tensor data
-    std::vector<ggml_backend_buffer_ptr> bufs;
+    // contexts where the model tensors metadata is stored as well ass the corresponding buffers:
+    std::vector<std::pair<ggml_context_ptr, ggml_backend_buffer_ptr>> ctxs_bufs;
     buft_list_t cpu_buft_list;
     std::map<ggml_backend_dev_t, buft_list_t> gpu_buft_list;
@@ -478,7 +491,8 @@ void llama_model::load_hparams(llama_model_loader & ml) {
     ml.get_key(LLM_KV_GENERAL_NAME, name, false);
     // everything past this point is not vocab-related
-    if (hparams.vocab_only) {
+    // for CLIP models, we only need to load tensors, no hparams
+    if (hparams.vocab_only || ml.get_arch() == LLM_ARCH_CLIP) {
         return;
     }
@@ -487,6 +501,8 @@ 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);
+    ml.get_key(LLM_KV_EXPERT_GROUP_COUNT,      hparams.n_expert_groups, false);
+    ml.get_key(LLM_KV_EXPERT_GROUP_USED_COUNT, hparams.n_group_used,    false);
     if (arch == LLM_ARCH_WAVTOKENIZER_DEC) {
         ml.get_key(LLM_KV_FEATURES_LENGTH, hparams.n_embd_features);
@@ -502,8 +518,15 @@ void llama_model::load_hparams(llama_model_loader & ml) {
     GGML_ASSERT(hparams.n_expert_used <= hparams.n_expert);
     if (hparams.n_expert > 0) {
         GGML_ASSERT(hparams.n_expert_used > 0);
+        GGML_ASSERT(hparams.n_expert_groups < hparams.n_expert);
+        if (hparams.n_expert_groups > 1) {
+            GGML_ASSERT(hparams.n_expert % hparams.n_expert_groups == 0);
+            GGML_ASSERT(hparams.n_group_used > 0);
+            GGML_ASSERT(hparams.n_group_used < hparams.n_expert_groups);
+        }
     } else {
         GGML_ASSERT(hparams.n_expert_used == 0);
+        GGML_ASSERT(hparams.n_expert_groups == 0);
     }
     std::fill(hparams.n_head_arr.begin(),    hparams.n_head_arr.end(),    0);
@@ -1845,8 +1868,10 @@ 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) {
-                    // TODO: Add llm type label (not sure this is useful)
+                switch (hparams.n_embd) {
+                    case 1536: type = LLM_TYPE_7B_A1B; break;
+                    case 2048: case 2560: type = LLM_TYPE_3B; break;
+                    case 4096: type = LLM_TYPE_32B; break;
                     default: type = LLM_TYPE_UNKNOWN;
                 }
@@ -1902,6 +1927,29 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                     default: type = LLM_TYPE_UNKNOWN;
                 }
             } break;
+        case LLM_ARCH_BAILINGMOE2:
+            {
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS,       hparams.f_norm_rms_eps);
+                ml.get_key(LLM_KV_LEADING_DENSE_BLOCK_COUNT,         hparams.n_layer_dense_lead);
+                ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,        hparams.n_ff_exp);
+                ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp);
+                ml.get_key(LLM_KV_EXPERT_SHARED_COUNT,               hparams.n_expert_shared);
+                ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE,              hparams.expert_weights_scale);
+                ml.get_key(LLM_KV_EXPERT_WEIGHTS_NORM,               hparams.expert_weights_norm, false);
+                ml.get_key(LLM_KV_EXPERT_GATING_FUNC,                hparams.expert_gating_func);
+                ml.get_key(LLM_KV_NEXTN_PREDICT_LAYERS,              hparams.nextn_predict_layers, false);
+
+                // TODO: when MTP is implemented, this should probably be updated if needed
+                hparams.n_layer_kv_from_start = hparams.n_layer - hparams.nextn_predict_layers;
+
+                switch (hparams.n_layer) {
+                    case 20: type = LLM_TYPE_16B_A1B; break;
+                    case 21: type = LLM_TYPE_16B_A1B; break;
+                    case 32: type = LLM_TYPE_100B_A6B; break;
+                    case 33: type = LLM_TYPE_100B_A6B; break;
+                    default: type = LLM_TYPE_UNKNOWN;
+                }
+            } break;
         case LLM_ARCH_DOTS1:
             {
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
@@ -2196,7 +2244,14 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
     max_n_tensors += n_layer*2; // duplicated rope freq tensors
     const size_t ctx_size = ggml_tensor_overhead()*max_n_tensors;
-    std::map<ggml_backend_buffer_type_t, ggml_context *> ctx_map;
+    // define a comparator for the buft -> ctx map to ensure that the order is well-defined:
+    struct ggml_backend_buft_comparator {
+        bool operator()(const ggml_backend_buffer_type_t & lhs, const ggml_backend_buffer_type_t & rhs) const {
+            return ggml_backend_buft_name(lhs) < ggml_backend_buft_name(rhs);
+        }
+    };
+    std::map<ggml_backend_buffer_type_t, ggml_context_ptr, ggml_backend_buft_comparator> ctx_map;
+
     auto ctx_for_buft = [&](ggml_backend_buffer_type_t buft) -> ggml_context * {
         auto it = ctx_map.find(buft);
         if (it == ctx_map.end()) {
@@ -2211,12 +2266,11 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                 throw std::runtime_error(format("failed to create ggml context"));
             }
-            ctx_map[buft] = ctx;
-            pimpl->ctxs.emplace_back(ctx);
+            ctx_map.emplace(buft, ctx);
             return ctx;
         }
-        return it->second;
+        return it->second.get();
     };
     const auto TENSOR_DUPLICATED   = llama_model_loader::TENSOR_DUPLICATED;
@@ -5534,6 +5588,70 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         layer.ffn_up_shexp   = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {n_embd, n_ff_exp * n_expert_shared}, 0);
                     }
                 } break;
+            case LLM_ARCH_BAILINGMOE2:
+                {
+                    const int64_t n_ff_exp        = hparams.n_ff_exp;
+                    const int64_t n_expert_shared = hparams.n_expert_shared;
+
+                    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}, 0);
+
+                    GGML_ASSERT(n_expert > 0 && "n_expert must be > 0 for bailingmoe2");
+                    GGML_ASSERT(n_expert_used > 0 && "n_expert_used must be > 0 for bailingmoe2");
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        int flags = 0;
+                        if (hparams.nextn_predict_layers > 0 && static_cast<uint32_t>(i) >= n_layer - hparams.nextn_predict_layers) {
+                            // skip all tensors in the NextN layers
+                            flags |= TENSOR_SKIP;
+                        }
+
+                        auto & layer = layers[i];
+
+                        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, flags);
+
+                        layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, flags);
+                        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, flags);
+
+                        layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k}, flags);
+                        layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k}, flags);
+
+                        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, flags);
+
+                        if (static_cast<uint32_t>(i) >= hparams.n_layer_dense_lead) { // MoE layers
+                            const int64_t n_ff_shexp = (hparams.n_ff_shexp ? hparams.n_ff_shexp : n_ff_exp) * n_expert_shared;
+
+                            layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, flags);
+                            layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), {n_expert}, TENSOR_NOT_REQUIRED | flags);
+
+                            layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {  n_embd, n_ff_exp, n_expert}, flags);
+                            layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp,   n_embd, n_expert}, flags);
+                            layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {  n_embd, n_ff_exp, n_expert}, flags);
+
+                            layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, n_ff_shexp}, flags);
+                            layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {n_ff_shexp, n_embd}, flags);
+                            layer.ffn_up_shexp   = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {n_embd, n_ff_shexp}, flags);
+                        } else { // Dense layers
+                            layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, flags);
+                            layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, flags);
+                            layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, flags);
+                        }
+
+                        // NextN/MTP tensors (preserved but unused) - conditionally load for last nextn_predict_layers
+                        if (hparams.nextn_predict_layers > 0 && static_cast<uint32_t>(i) >= n_layer - hparams.nextn_predict_layers) {
+                            layer.nextn.eh_proj          = create_tensor(tn(LLM_TENSOR_NEXTN_EH_PROJ, "weight", i), { 2 * n_embd, n_embd }, flags);
+                            layer.nextn.embed_tokens     = create_tensor(tn(LLM_TENSOR_NEXTN_EMBED_TOKENS, "weight", i), { n_embd, n_vocab }, TENSOR_NOT_REQUIRED | flags);
+                            layer.nextn.enorm            = create_tensor(tn(LLM_TENSOR_NEXTN_ENORM, "weight", i), { n_embd }, flags);
+                            layer.nextn.hnorm            = create_tensor(tn(LLM_TENSOR_NEXTN_HNORM, "weight", i), { n_embd }, flags);
+                            layer.nextn.shared_head_head = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD, "weight", i), { n_embd, n_vocab }, TENSOR_NOT_REQUIRED | flags);
+                            layer.nextn.shared_head_norm = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, "weight", i), { n_embd }, TENSOR_NOT_REQUIRED | flags);
+                            layer.layer_out_norm         = create_tensor(tn(LLM_TENSOR_LAYER_OUT_NORM, "weight", i), {n_embd}, flags);
+                        }
+                    }
+                } break;
             case LLM_ARCH_DOTS1:
                 {
                     const int64_t n_ff_exp        = hparams.n_ff_exp;
@@ -6079,16 +6197,15 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
     pimpl->mappings.reserve(ml.mappings.size());
     // create the backend buffers
-    std::vector<std::pair<ggml_context *, llama_buf_map>> ctx_bufs;
-    ctx_bufs.reserve(ctx_map.size());
+    std::vector<std::pair<ggml_context *, llama_buf_map>> ctx_buf_maps;
+    ctx_buf_maps.reserve(ctx_map.size());
     // Ensure we have enough capacity for the maximum backend buffer we will potentially create
     const size_t n_max_backend_buffer = ctx_map.size() * ml.files.size();
-    pimpl->bufs.reserve(n_max_backend_buffer);
-    for (auto & it : ctx_map) {
-        ggml_backend_buffer_type_t buft = it.first;
-        ggml_context * ctx              = it.second;
+    pimpl->ctxs_bufs.reserve(n_max_backend_buffer);
+    for (auto & [buft, ctx_ptr] : ctx_map) {
+        ggml_context * ctx = ctx_ptr.get();
         // skip contexts without tensors
         if (ggml_get_first_tensor(ctx) == nullptr) {
@@ -6112,6 +6229,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
         bool buffer_from_host_ptr_supported = props.caps.buffer_from_host_ptr;
         bool is_default_buft = buft == ggml_backend_dev_buffer_type(dev);
+        ggml_backend_buffer_t buf = nullptr;
         if (ml.use_mmap && use_mmap_buffer && buffer_from_host_ptr_supported && is_default_buft) {
             for (uint32_t idx = 0; idx < ml.files.size(); idx++) {
                 // only the mmap region containing the tensors in the model is mapped to the backend buffer
@@ -6124,20 +6242,18 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                     continue;
                 }
                 const size_t max_size = ggml_get_max_tensor_size(ctx);
-                ggml_backend_buffer_t buf = ggml_backend_dev_buffer_from_host_ptr(dev, (char *) addr + first, last - first, max_size);
+                buf = ggml_backend_dev_buffer_from_host_ptr(dev, (char *) addr + first, last - first, max_size);
                 if (buf == nullptr) {
                     throw std::runtime_error(format("unable to allocate %s buffer", ggml_backend_buft_name(buft)));
                 }
-                pimpl->bufs.emplace_back(buf);
                 buf_map.emplace(idx, buf);
             }
         }
         else {
-            ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx, buft);
+            buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx, buft);
             if (buf == nullptr) {
                 throw std::runtime_error(format("unable to allocate %s buffer", ggml_backend_buft_name(buft)));
             }
-            pimpl->bufs.emplace_back(buf);
             if (use_mlock && ggml_backend_buffer_is_host(buf)) {
                 pimpl->mlock_bufs.emplace_back(new llama_mlock);
                 auto & mlock_buf = pimpl->mlock_bufs.back();
@@ -6148,10 +6264,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                 buf_map.emplace(idx, buf);
             }
         }
-
-        if (pimpl->bufs.empty()) {
-            throw std::runtime_error("failed to allocate buffer");
-        }
+        pimpl->ctxs_bufs.emplace_back(std::move(ctx_ptr), buf);
         for (auto & buf : buf_map) {
             // indicate that this buffer contains weights
@@ -6159,7 +6272,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
             ggml_backend_buffer_set_usage(buf.second, GGML_BACKEND_BUFFER_USAGE_WEIGHTS);
         }
-        ctx_bufs.emplace_back(ctx, buf_map);
+        ctx_buf_maps.emplace_back(ctx, buf_map);
     }
     if (llama_supports_gpu_offload()) {
@@ -6177,22 +6290,20 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
     }
     // print memory requirements per buffer type
-    for (auto & buf : pimpl->bufs) {
+    for (auto & [_, buf] : pimpl->ctxs_bufs) {
         LLAMA_LOG_INFO("%s: %12s model buffer size = %8.2f MiB\n", __func__, ggml_backend_buffer_name(buf.get()), ggml_backend_buffer_get_size(buf.get()) / 1024.0 / 1024.0);
     }
     // populate tensors_by_name
-    for (auto & ctx : pimpl->ctxs) {
+    for (auto & [ctx, _] : pimpl->ctxs_bufs) {
         for (auto * cur = ggml_get_first_tensor(ctx.get()); cur != NULL; cur = ggml_get_next_tensor(ctx.get(), cur)) {
             tensors_by_name.emplace_back(ggml_get_name(cur), cur);
         }
     }
     // load tensor data
-    for (auto & it : ctx_bufs) {
-        ggml_context * ctx = it.first;
-        auto & bufs = it.second;
-        if (!ml.load_all_data(ctx, bufs, use_mlock ? &pimpl->mlock_mmaps : NULL, params.progress_callback, params.progress_callback_user_data)) {
+    for (auto & [ctx, buf_map] : ctx_buf_maps) {
+        if (!ml.load_all_data(ctx, buf_map, use_mlock ? &pimpl->mlock_mmaps : NULL, params.progress_callback, params.progress_callback_user_data)) {
             return false;
         }
     }
@@ -6232,8 +6343,8 @@ size_t llama_model::n_devices() const {
 std::map<ggml_backend_buffer_type_t, size_t> llama_model::memory_breakdown() const {
     std::map<ggml_backend_buffer_type_t, size_t> ret;
-    for (const ggml_backend_buffer_ptr & buf_ptr : pimpl->bufs) {
-        ret[ggml_backend_buffer_get_type(buf_ptr.get())] += ggml_backend_buffer_get_size(buf_ptr.get());
+    for (const auto & [_, buf] : pimpl->ctxs_bufs) {
+        ret[ggml_backend_buffer_get_type(buf.get())] += ggml_backend_buffer_get_size(buf.get());
     }
     return ret;
 }
@@ -6396,6 +6507,19 @@ void llama_model::print_info() const {
         LLAMA_LOG_INFO("%s: expert_weights_norm  = %d\n",     __func__, hparams.expert_weights_norm);
     }
+    if (arch == LLM_ARCH_BAILINGMOE2) {
+        LLAMA_LOG_INFO("%s: n_layer_dense_lead   = %d\n",     __func__, hparams.n_layer_dense_lead);
+        LLAMA_LOG_INFO("%s: n_ff_exp             = %d\n",     __func__, hparams.n_ff_exp);
+        LLAMA_LOG_INFO("%s: n_ff_shexp           = %d\n",     __func__, hparams.n_ff_shexp);
+        LLAMA_LOG_INFO("%s: n_expert_shared      = %d\n",     __func__, hparams.n_expert_shared);
+        LLAMA_LOG_INFO("%s: n_expert_groups      = %d\n",     __func__, hparams.n_expert_groups);
+        LLAMA_LOG_INFO("%s: n_group_used         = %d\n",     __func__, hparams.n_group_used);
+        LLAMA_LOG_INFO("%s: expert_weights_scale = %.1f\n",   __func__, hparams.expert_weights_scale);
+        LLAMA_LOG_INFO("%s: expert_weights_norm  = %d\n",     __func__, hparams.expert_weights_norm);
+        LLAMA_LOG_INFO("%s: expert_gating_func   = %s\n",     __func__, llama_expert_gating_func_name((llama_expert_gating_func_type) hparams.expert_gating_func));
+        LLAMA_LOG_INFO("%s: nextn_predict_layers = %d\n",     __func__, hparams.nextn_predict_layers);
+    }
+
     if (arch == LLM_ARCH_SMALLTHINKER || arch == LLM_ARCH_LFM2MOE) {
         LLAMA_LOG_INFO("%s: n_ff_exp             = %d\n",     __func__, hparams.n_ff_exp);
         LLAMA_LOG_INFO("%s: expert_gating_func   = %s\n",     __func__, llama_expert_gating_func_name((llama_expert_gating_func_type) hparams.expert_gating_func));
@@ -11401,8 +11525,8 @@ struct llm_build_gemma3n_iswa : public llm_graph_context {
     }
 };
-struct llm_build_gemma_embedding_iswa : public llm_graph_context {
-    llm_build_gemma_embedding_iswa(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+struct llm_build_gemma_embedding : public llm_graph_context {
+    llm_build_gemma_embedding(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
         const int64_t n_embd_head = hparams.n_embd_head_k;
         ggml_tensor * cur;
@@ -11419,8 +11543,7 @@ struct llm_build_gemma_embedding_iswa : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
-        // TODO: support cacheless iSWA embeddings [TAG_NO_CACHE_ISWA]
-        auto * inp_attn = build_attn_inp_kv_iswa();
+        auto * inp_attn = build_attn_inp_no_cache();
         ggml_tensor * inp_out_ids = build_inp_out_ids();
@@ -17245,6 +17368,150 @@ struct llm_build_bailingmoe : public llm_graph_context {
     }
 };
+struct llm_build_bailingmoe2 : public llm_graph_context {
+    llm_build_bailingmoe2(const llama_model & model, const llm_graph_params & params) : 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();
+
+        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+        ggml_tensor * cur;
+        ggml_tensor * inpL;
+
+        inpL = build_inp_embd(model.tok_embd);
+
+        // inp_pos - contains the positions
+        ggml_tensor * inp_pos = build_inp_pos();
+
+        auto * inp_attn = build_attn_inp_kv();
+
+        ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+        const int n_transformer_layers = n_layer - hparams.nextn_predict_layers;
+        for (int il = 0; il < n_transformer_layers; ++il) {
+            ggml_tensor * inpSA = inpL;
+
+            // norm
+            cur = build_norm(inpL,
+                    model.layers[il].attn_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "attn_norm", il);
+
+            // self_attention
+            {
+                cur = build_lora_mm(model.layers[il].wqkv, cur);
+                cb(cur, "wqkv", il);
+
+                ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head,    n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd));
+                ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd));
+                ggml_tensor * Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa));
+
+                Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
+                cb(Qcur, "Qcur_normed", il);
+
+                Qcur = ggml_rope_ext(
+                        ctx0, Qcur, inp_pos, nullptr,
+                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                        ext_factor, attn_factor, beta_fast, beta_slow
+                        );
+
+                Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
+                cb(Kcur, "Kcur_normed", il);
+
+                Kcur = ggml_rope_ext(
+                        ctx0, Kcur, inp_pos, nullptr,
+                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                        ext_factor, attn_factor, beta_fast, beta_slow
+                        );
+
+                cb(Qcur, "Qcur", il);
+                cb(Kcur, "Kcur", il);
+                cb(Vcur, "Vcur", il);
+
+                cur = build_attn(inp_attn,
+                        model.layers[il].wo, model.layers[il].bo,
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+            }
+
+            if (il == n_transformer_layers - 1 && inp_out_ids) {
+                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+            }
+
+            ggml_tensor * sa_out = ggml_add(ctx0, cur, inpSA);
+            cb(sa_out, "sa_out", il);
+
+            // MoE branch
+            cur = build_norm(sa_out,
+                    model.layers[il].ffn_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "ffn_norm", il);
+
+            if (static_cast<uint32_t>(il) < hparams.n_layer_dense_lead) {
+                cur = build_ffn(cur,
+                        model.layers[il].ffn_up,   NULL, NULL,
+                        model.layers[il].ffn_gate, NULL, NULL,
+                        model.layers[il].ffn_down, NULL, NULL,
+                        NULL,
+                        LLM_FFN_SILU, LLM_FFN_PAR, il);
+                cb(cur, "ffn_out", il);
+            } else {
+                ggml_tensor * moe_out =
+                    build_moe_ffn(cur,
+                            model.layers[il].ffn_gate_inp,
+                            model.layers[il].ffn_up_exps,
+                            model.layers[il].ffn_gate_exps,
+                            model.layers[il].ffn_down_exps,
+                            model.layers[il].ffn_exp_probs_b,
+                            n_expert, n_expert_used,
+                            LLM_FFN_SILU, hparams.expert_weights_norm,
+                            true, hparams.expert_weights_scale,
+                            (llama_expert_gating_func_type) hparams.expert_gating_func,
+                            il);
+                cb(moe_out, "ffn_moe_out", il);
+
+                {
+                    ggml_tensor * ffn_shexp = build_ffn(cur,
+                            model.layers[il].ffn_up_shexp,   NULL, NULL,
+                            model.layers[il].ffn_gate_shexp, NULL, NULL,
+                            model.layers[il].ffn_down_shexp, NULL, NULL,
+                            NULL,
+                            LLM_FFN_SILU, LLM_FFN_PAR, il);
+                    cb(ffn_shexp, "ffn_shexp", il);
+
+                    cur = ggml_add(ctx0, moe_out, ffn_shexp);
+                    cb(cur, "ffn_out", il);
+                }
+            }
+
+            cur = ggml_add(ctx0, cur, sa_out);
+
+            cur = build_cvec(cur, il);
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+
+        cur = inpL;
+
+        cur = build_norm(cur,
+                model.output_norm, NULL,
+                LLM_NORM_RMS, -1);
+
+        cb(cur, "result_norm", -1);
+        res->t_embd = cur;
+
+        // lm_head
+        cur = build_lora_mm(model.output, cur);
+
+        cb(cur, "result_output", -1);
+        res->t_logits = cur;
+
+        ggml_build_forward_expand(gf, cur);
+    }
+};
+
 struct llm_build_dots1 : public llm_graph_context {
     llm_build_dots1(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
         const int64_t n_embd_head = hparams.n_embd_head_v;
@@ -17900,6 +18167,8 @@ struct llm_build_plamo2 : public llm_graph_context_mamba {
         cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
         cb(cur, "result_norm", -1);
+        res->t_embd = cur;
+
         // lm_head
         cur = build_lora_mm(model.output, cur);
         cb(cur, "result_output", -1);
@@ -19580,7 +19849,7 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
         case LLM_ARCH_NOMIC_BERT_MOE:
         case LLM_ARCH_NEO_BERT:
         case LLM_ARCH_WAVTOKENIZER_DEC:
-        //case LLM_ARCH_GEMMA_EMBEDDING: // TODO: disabled until the cacheless SWA logic is fixed [TAG_NO_CACHE_ISWA]
+        case LLM_ARCH_GEMMA_EMBEDDING:
         case LLM_ARCH_DREAM:
         case LLM_ARCH_LLADA:
         case LLM_ARCH_LLADA_MOE:
@@ -19873,7 +20142,7 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
             } break;
         case LLM_ARCH_GEMMA_EMBEDDING:
             {
-                llm = std::make_unique<llm_build_gemma_embedding_iswa>(*this, params);
+                llm = std::make_unique<llm_build_gemma_embedding>(*this, params);
             } break;
         case LLM_ARCH_STARCODER2:
             {
@@ -20045,6 +20314,10 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
             {
                 llm = std::make_unique<llm_build_bailingmoe>(*this, params);
             } break;
+        case LLM_ARCH_BAILINGMOE2:
+            {
+                llm = std::make_unique<llm_build_bailingmoe2>(*this, params);
+            } break;
         case LLM_ARCH_SEED_OSS:
             {
                 llm = std::make_unique<llm_build_seed_oss>(*this, params);
@@ -20220,6 +20493,7 @@ int32_t llama_n_head(const llama_model * model) {
 llama_rope_type llama_model_rope_type(const llama_model * model) {
     switch (model->arch) {
         // these models do not use RoPE
+        case LLM_ARCH_CLIP:
         case LLM_ARCH_GPT2:
         case LLM_ARCH_GPTJ:
         case LLM_ARCH_MPT:
@@ -20311,6 +20585,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
         case LLM_ARCH_EXAONE:
         case LLM_ARCH_EXAONE4:
         case LLM_ARCH_MINICPM3:
+        case LLM_ARCH_BAILINGMOE2:
         case LLM_ARCH_DOTS1:
         case LLM_ARCH_HUNYUAN_MOE:
         case LLM_ARCH_OPENAI_MOE:

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

@@ -108,9 +108,12 @@ enum llm_type {
     LLM_TYPE_17B_16E, // llama4 Scout
     LLM_TYPE_17B_128E, // llama4 Maverick
     LLM_TYPE_A13B,
+    LLM_TYPE_7B_A1B,
     LLM_TYPE_8B_A1B, // lfm2moe
+    LLM_TYPE_16B_A1B,
     LLM_TYPE_21B_A3B, // Ernie MoE small
     LLM_TYPE_30B_A3B,
+    LLM_TYPE_100B_A6B,
     LLM_TYPE_106B_A12B, // GLM-4.5-Air
     LLM_TYPE_235B_A22B,
     LLM_TYPE_300B_A47B, // Ernie MoE big

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

@@ -701,6 +701,7 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
         });
     }
 
+    bool is_clip_model = false;
     for (const auto * it : tensors) {
         const struct ggml_tensor * tensor = it->tensor;
 
@@ -714,12 +715,14 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
         } else if (name == LLM_TN(model.arch)(LLM_TENSOR_OUTPUT, "weight")) {
             qs.has_output = true;
         }
+
+        is_clip_model |= name.rfind("mm.", 0) == 0; // check the "mm." prefix
     }
 
     qs.n_ffn_down = qs.n_ffn_gate = qs.n_ffn_up = (int)model.hparams.n_layer;
 
     // sanity checks for models that have attention layers
-    if (qs.n_attention_wv != 0)
+    if (qs.n_attention_wv != 0 && !is_clip_model)
     {
         const auto & n_head_kv_iter = model.hparams.n_head_kv_arr.begin();
         // attention layers have a non-zero number of kv heads
@@ -881,6 +884,9 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
         // do not quantize relative position bias (T5)
         quantize &= name.find("attn_rel_b.weight") == std::string::npos;
 
+        // do not quantize specific multimodal tensors
+        quantize &= name.find(".position_embd.") == std::string::npos;
+
         ggml_type new_type;
         void * new_data;
         size_t new_size;

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

@@ -1957,6 +1957,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                 clean_spaces = false;
             } else if (
                 tokenizer_pre == "bailingmoe" ||
+                tokenizer_pre == "bailingmoe2" ||
                 tokenizer_pre == "llada-moe") {
                 pre_type = LLAMA_VOCAB_PRE_TYPE_BAILINGMOE;
                 clean_spaces = false;

llama/llama.cpp/src/llama.cpp

@@ -124,6 +124,9 @@ static int llama_model_load(const std::string & fname, std::vector<std::string>
         } catch(const std::exception & e) {
             throw std::runtime_error("error loading model hyperparameters: " + std::string(e.what()));
         }
+        if (model.arch == LLM_ARCH_CLIP) {
+            throw std::runtime_error("CLIP cannot be used as main model, use it with --mmproj instead");
+        }
         try {
             model.load_vocab(ml);
         } catch(const std::exception & e) {
@@ -314,6 +317,7 @@ struct llama_model * llama_model_load_from_splits(
         LLAMA_LOG_ERROR("%s: list of splits is empty\n", __func__);
         return nullptr;
     }
+    splits.reserve(n_paths);
     for (size_t i = 0; i < n_paths; ++i) {
         splits.push_back(paths[i]);
     }

llama/llama.cpp/tools/mtmd/clip-impl.h

@@ -30,6 +30,7 @@
 #define KEY_LAYER_NORM_EPS      "clip.%s.attention.layer_norm_epsilon"
 
 // vision-specific
+#define KEY_VISION_PROJ_TYPE    "clip.vision.projector_type" // for models with mixed modalities
 #define KEY_IMAGE_SIZE          "clip.vision.image_size"
 #define KEY_PREPROC_IMAGE_SIZE  "clip.vision.preproc_image_size"
 #define KEY_PATCH_SIZE          "clip.vision.patch_size"
@@ -48,6 +49,7 @@
 #define KEY_MINICPMV_QUERY_NUM    "clip.minicpmv_query_num"
 
 // audio-specific
+#define KEY_AUDIO_PROJ_TYPE     "clip.audio.projector_type" // for models with mixed modalities
 #define KEY_A_NUM_MEL_BINS      "clip.audio.num_mel_bins"
 #define KEY_A_PROJ_STACK_FACTOR "clip.audio.projector.stack_factor"
 

llama/llama.cpp/tools/mtmd/clip.cpp

@@ -2234,15 +2234,27 @@ struct clip_model_loader {
         // projector type
         std::string proj_type;
         {
+            // default key
             get_string(KEY_PROJ_TYPE, proj_type, false);
-            if (!proj_type.empty()) {
-                model.proj_type = clip_projector_type_from_string(proj_type);
+
+            // for models with mixed modalities
+            if (proj_type.empty()) {
+                if (modality == CLIP_MODALITY_VISION) {
+                    get_string(KEY_VISION_PROJ_TYPE, proj_type, false);
+                } else if (modality == CLIP_MODALITY_AUDIO) {
+                    get_string(KEY_AUDIO_PROJ_TYPE, proj_type, false);
+                } else {
+                    GGML_ABORT("unknown modality");
                 }
+            }
+
+            model.proj_type = clip_projector_type_from_string(proj_type);
+
             if (model.proj_type == PROJECTOR_TYPE_UNKNOWN) {
                 throw std::runtime_error(string_format("%s: unknown projector type: %s\n", __func__, proj_type.c_str()));
             }
 
-            // correct arch for multimodal models
+            // correct arch for multimodal models (legacy method)
             if (model.proj_type == PROJECTOR_TYPE_QWEN25O) {
                 model.proj_type = modality == CLIP_MODALITY_VISION
                                     ? PROJECTOR_TYPE_QWEN25VL

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

@@ -23,7 +23,7 @@ problem.
  8 files changed, 21 insertions(+), 2 deletions(-)
 
 diff --git a/ggml/src/ggml-backend.cpp b/ggml/src/ggml-backend.cpp
-index ff9135fe..8ba86f82 100644
+index ff9135fe2..8ba86f824 100644
 --- a/ggml/src/ggml-backend.cpp
 +++ b/ggml/src/ggml-backend.cpp
 @@ -113,7 +113,6 @@ void ggml_backend_buffer_free(ggml_backend_buffer_t buffer) {
@@ -64,18 +64,18 @@ index ff9135fe..8ba86f82 100644
      /* .init_tensor     = */ NULL, // no initialization required
      /* .memset_tensor   = */ ggml_backend_cpu_buffer_memset_tensor,
 diff --git a/ggml/src/ggml-cann/ggml-cann.cpp b/ggml/src/ggml-cann/ggml-cann.cpp
-index ad1adba6..7d44f74f 100755
+index 8bd5449f1..01e2df61a 100644
 --- a/ggml/src/ggml-cann/ggml-cann.cpp
 +++ b/ggml/src/ggml-cann/ggml-cann.cpp
-@@ -843,6 +843,7 @@ static void ggml_backend_cann_buffer_free_buffer(
-     ggml_backend_cann_buffer_context* ctx =
-         (ggml_backend_cann_buffer_context*)buffer->context;
+@@ -820,6 +820,7 @@ static bool ggml_backend_buffer_is_cann(ggml_backend_buffer_t buffer) {
+ static void ggml_backend_cann_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+     ggml_backend_cann_buffer_context * ctx = (ggml_backend_cann_buffer_context *) buffer->context;
      delete ctx;
 +    delete buffer;
  }
  
  /**
-@@ -1630,6 +1631,7 @@ static const char * ggml_backend_cann_host_buffer_name(ggml_backend_buffer_t buf
+@@ -1560,6 +1561,7 @@ static const char * ggml_backend_cann_host_buffer_name(ggml_backend_buffer_t buf
   */
  static void ggml_backend_cann_host_buffer_free(ggml_backend_buffer_t buffer) {
      ACL_CHECK(aclrtFreeHost(buffer->context));
@@ -84,10 +84,10 @@ index ad1adba6..7d44f74f 100755
  
  /**
 diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu
-index 856e9de2..c0b1e4c1 100644
+index bc396b521..aefc6935e 100644
 --- a/ggml/src/ggml-cuda/ggml-cuda.cu
 +++ b/ggml/src/ggml-cuda/ggml-cuda.cu
-@@ -567,6 +567,7 @@ struct ggml_backend_cuda_buffer_context {
+@@ -576,6 +576,7 @@ struct ggml_backend_cuda_buffer_context {
  static void ggml_backend_cuda_buffer_free_buffer(ggml_backend_buffer_t buffer) {
      ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context;
      delete ctx;
@@ -95,7 +95,7 @@ index 856e9de2..c0b1e4c1 100644
  }
  
  static bool ggml_backend_buffer_is_cuda(ggml_backend_buffer_t buffer) {
-@@ -822,6 +823,7 @@ struct ggml_backend_cuda_split_buffer_context {
+@@ -831,6 +832,7 @@ struct ggml_backend_cuda_split_buffer_context {
  static void ggml_backend_cuda_split_buffer_free_buffer(ggml_backend_buffer_t buffer) {
      ggml_backend_cuda_split_buffer_context * ctx = (ggml_backend_cuda_split_buffer_context *)buffer->context;
      delete ctx;
@@ -103,7 +103,7 @@ index 856e9de2..c0b1e4c1 100644
  }
  
  static void * ggml_backend_cuda_split_buffer_get_base(ggml_backend_buffer_t buffer) {
-@@ -1103,6 +1105,7 @@ static bool ggml_backend_buft_is_cuda_host(ggml_backend_buffer_type_t buft) {
+@@ -1112,6 +1114,7 @@ static bool ggml_backend_buft_is_cuda_host(ggml_backend_buffer_type_t buft) {
  
  static void ggml_backend_cuda_host_buffer_free_buffer(ggml_backend_buffer_t buffer) {
      CUDA_CHECK(cudaFreeHost(buffer->context));
@@ -112,7 +112,7 @@ index 856e9de2..c0b1e4c1 100644
  
  static void * ggml_cuda_host_malloc(size_t size) {
 diff --git a/ggml/src/ggml-metal/ggml-metal.cpp b/ggml/src/ggml-metal/ggml-metal.cpp
-index 7afc881f..bf096227 100644
+index 7afc881fa..bf0962274 100644
 --- a/ggml/src/ggml-metal/ggml-metal.cpp
 +++ b/ggml/src/ggml-metal/ggml-metal.cpp
 @@ -25,6 +25,7 @@ static void ggml_backend_metal_buffer_shared_free_buffer(ggml_backend_buffer_t b
@@ -132,10 +132,10 @@ index 7afc881f..bf096227 100644
  
  static void * ggml_backend_metal_buffer_private_get_base(ggml_backend_buffer_t buffer) {
 diff --git a/ggml/src/ggml-opencl/ggml-opencl.cpp b/ggml/src/ggml-opencl/ggml-opencl.cpp
-index 79d21487..38c75018 100644
+index db33a4ab6..c42ee26e1 100644
 --- a/ggml/src/ggml-opencl/ggml-opencl.cpp
 +++ b/ggml/src/ggml-opencl/ggml-opencl.cpp
-@@ -3212,6 +3212,7 @@ struct ggml_backend_opencl_buffer_context {
+@@ -3266,6 +3266,7 @@ struct ggml_backend_opencl_buffer_context {
  static void ggml_backend_opencl_buffer_free_buffer(ggml_backend_buffer_t buffer) {
      ggml_backend_opencl_buffer_context * ctx = (ggml_backend_opencl_buffer_context *) buffer->context;
      delete ctx;
@@ -144,7 +144,7 @@ index 79d21487..38c75018 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 aad48d62..a46c0f52 100644
+index a38df5a97..fd07e4a21 100644
 --- a/ggml/src/ggml-rpc/ggml-rpc.cpp
 +++ b/ggml/src/ggml-rpc/ggml-rpc.cpp
 @@ -528,6 +528,7 @@ static void ggml_backend_rpc_buffer_free_buffer(ggml_backend_buffer_t buffer) {
@@ -156,10 +156,10 @@ index aad48d62..a46c0f52 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 45b8c216..4ec9a592 100644
+index b695ba051..37e853120 100644
 --- a/ggml/src/ggml-sycl/ggml-sycl.cpp
 +++ b/ggml/src/ggml-sycl/ggml-sycl.cpp
-@@ -334,6 +334,7 @@ ggml_backend_sycl_buffer_free_buffer(ggml_backend_buffer_t buffer) try {
+@@ -352,6 +352,7 @@ ggml_backend_sycl_buffer_free_buffer(ggml_backend_buffer_t buffer) try {
      ggml_sycl_set_device(ctx->device);
  
      delete ctx;
@@ -167,7 +167,7 @@ index 45b8c216..4ec9a592 100644
  }
  catch (sycl::exception const &exc) {
    std::cerr << exc.what() << "Exception caught at file:" << __FILE__
-@@ -795,6 +796,7 @@ struct ggml_backend_sycl_split_buffer_context {
+@@ -813,6 +814,7 @@ 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;
@@ -175,7 +175,7 @@ index 45b8c216..4ec9a592 100644
  }
  
  static void * ggml_backend_sycl_split_buffer_get_base(ggml_backend_buffer_t buffer) {
-@@ -1137,6 +1139,7 @@ static const char * ggml_backend_sycl_host_buffer_type_name(ggml_backend_buffer_
+@@ -1155,6 +1157,7 @@ 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);
@@ -184,10 +184,10 @@ index 45b8c216..4ec9a592 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 3cd89c71..ed83236f 100644
+index b783f7805..216dc167c 100644
 --- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp
 +++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
-@@ -11600,6 +11600,7 @@ static void ggml_backend_vk_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+@@ -11828,6 +11828,7 @@ 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;
@@ -195,7 +195,7 @@ index 3cd89c71..ed83236f 100644
  }
  
  static void * ggml_backend_vk_buffer_get_base(ggml_backend_buffer_t buffer) {
-@@ -11743,6 +11744,7 @@ static const char * ggml_backend_vk_host_buffer_name(ggml_backend_buffer_t buffe
+@@ -11971,6 +11972,7 @@ 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);