llama-cpp-bin 9479.0.0__tar.gz → 9484.0.0__tar.gz

{llama_cpp_bin-9479.0.0 → llama_cpp_bin-9484.0.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: llama-cpp-bin
-Version: 9479.0.0
+Version: 9484.0.0
 Summary: llama.cpp server binary built from source
 Author-email: Vladyslav Tkachenko <vladlearns@gmail.com>
 License: AGPL-3.0-or-later

{llama_cpp_bin-9479.0.0 → llama_cpp_bin-9484.0.0}/llama.cpp/ggml/src/ggml-hexagon/htp-opnode.h

@@ -56,6 +56,20 @@ struct htp_opnode {
     }
 
     std::vector<const ggml_tensor *> get_inputs() const {
+        if (fused.empty()) {
+            int last_non_null = -1;
+            for (int i = 0; i < GGML_MAX_SRC; i++) {
+                if (node->src[i]) {
+                    last_non_null = i;
+                }
+            }
+            std::vector<const ggml_tensor *> inputs(last_non_null + 1, nullptr);
+            for (int i = 0; i <= last_non_null; i++) {
+                inputs[i] = node->src[i];
+            }
+            return inputs;
+        }
+
         std::vector<const ggml_tensor *> inputs(GGML_MAX_SRC, nullptr);
         std::vector<const ggml_tensor *> outputs;
         outputs.push_back(node);
@@ -82,12 +96,8 @@ struct htp_opnode {
         };
 
         for (int i = 0; i < GGML_MAX_SRC; i++) {
-            if (fused.empty()) {
-                inputs[i] = node->src[i];
-            } else {
-                if (node->src[i]) {
-                    add_input(node->src[i]);
-                }
+            if (node->src[i]) {
+                add_input(node->src[i]);
             }
         }
         for (const auto * f : fused) {
@@ -98,10 +108,7 @@ struct htp_opnode {
             }
         }
 
-        if (!fused.empty()) {
-            inputs.resize(count);
-        }
-
+        inputs.resize(count);
         return inputs;
     }
 

{llama_cpp_bin-9479.0.0 → llama_cpp_bin-9484.0.0}/llama.cpp/ggml/src/ggml-opencl/ggml-opencl.cpp

@@ -4950,6 +4950,21 @@ inline bool enable_adreno_trans_weight(const ggml_backend_opencl_context *backen
     return ((elem_num < 128 * 1024 * 1024) && adreno_kernel);  // max element num: 2**27
 }
 
+static inline bool use_flat_gemv_for_large_m_q4_K(const ggml_tensor *tensor) {
+    // gemv_noshuffle variant perf drops for large M, use flat variant for large M.
+    // threshold is well above typical hidden/FFN dims, but below typical vocab sizes.
+    // note that this forces large M weights to use LM GEMM.
+    return tensor->ne[1] >= 32768 && tensor->ne[2] == 1 && tensor->ne[3] == 1;
+}
+
+static inline bool use_flat_gemv_for_large_m_q6_K(const ggml_tensor *tensor) {
+    // gemv_noshuffle variant perf drops for large M, use flat variant for large M.
+    // threshold is well above typical hidden/FFN dims, but below typical vocab sizes.
+    // q6_K flat gemv is worse for smaller K; 2048 seems to be a reasonable threshold.
+    // note that this forces large M weights to use LM GEMM.
+    return tensor->ne[1] >= 32768 && tensor->ne[0] >= 2048 && tensor->ne[2] == 1 && tensor->ne[3] == 1;
+}
+
 static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_tensor * op) {
     ggml_backend_opencl_device_context * dev_ctx     = (ggml_backend_opencl_device_context *)dev->context;
     ggml_backend_opencl_context *        backend_ctx = dev_ctx->backend_ctx;
@@ -6595,7 +6610,7 @@ static void ggml_backend_opencl_buffer_set_tensor(ggml_backend_buffer_t buffer,
 
 #ifdef GGML_OPENCL_USE_ADRENO_KERNELS
         cl_kernel kernel = backend_ctx->kernel_convert_block_q4_K;
-        if (use_adreno_kernels(backend_ctx, tensor)) {
+        if (use_adreno_kernels(backend_ctx, tensor) && !use_flat_gemv_for_large_m_q4_K(tensor)) {
             kernel = backend_ctx->kernel_convert_block_q4_K_noshuffle;
         }
 #else
@@ -6623,7 +6638,7 @@ static void ggml_backend_opencl_buffer_set_tensor(ggml_backend_buffer_t buffer,
 
         tensor->extra  = extra;
 #ifdef GGML_OPENCL_USE_ADRENO_KERNELS
-        if (use_adreno_kernels(backend_ctx, tensor)) {
+        if (use_adreno_kernels(backend_ctx, tensor) && !use_flat_gemv_for_large_m_q4_K(tensor)) {
 
             int M = tensor->ne[1];
             int K = tensor->ne[0];
@@ -6923,7 +6938,7 @@ static void ggml_backend_opencl_buffer_set_tensor(ggml_backend_buffer_t buffer,
         cl_kernel kernel;
 #ifdef GGML_OPENCL_USE_ADRENO_KERNELS
         kernel = backend_ctx->kernel_convert_block_q6_K;
-        if (use_adreno_kernels(backend_ctx, tensor)) {
+        if (use_adreno_kernels(backend_ctx, tensor) && !use_flat_gemv_for_large_m_q6_K(tensor)) {
             kernel = backend_ctx->kernel_convert_block_q6_K_noshuffle;
         }
 #else
@@ -6956,7 +6971,7 @@ static void ggml_backend_opencl_buffer_set_tensor(ggml_backend_buffer_t buffer,
         tensor->extra  = extra;
 
 #ifdef GGML_OPENCL_USE_ADRENO_KERNELS
-        if (use_adreno_kernels(backend_ctx, tensor)) {
+        if (use_adreno_kernels(backend_ctx, tensor) && !use_flat_gemv_for_large_m_q6_K(tensor)) {
             cl_int M = tensor->ne[1];   // ne01
             cl_int K = tensor->ne[0];   // ne00
 
@@ -7599,7 +7614,7 @@ static void ggml_backend_opencl_buffer_get_tensor(ggml_backend_buffer_t buffer,
             CL_CHECK(clReleaseMemObject(data_device));
             return;
         }
-        if (use_adreno_kernels(backend_ctx, tensor)) {
+        if (use_adreno_kernels(backend_ctx, tensor) && !use_flat_gemv_for_large_m_q4_K(tensor)) {
             int M = tensor->ne[1];
             int K = tensor->ne[0];
 
@@ -7820,7 +7835,7 @@ static void ggml_backend_opencl_buffer_get_tensor(ggml_backend_buffer_t buffer,
             CL_CHECK(clReleaseMemObject(data_device));
             return;
         }
-        if (use_adreno_kernels(backend_ctx, tensor)) {
+        if (use_adreno_kernels(backend_ctx, tensor) && !use_flat_gemv_for_large_m_q6_K(tensor)) {
             static ggml_cl_buffer buf_trans_ql;
             static ggml_cl_buffer buf_trans_qh;
             static ggml_cl_buffer buf_trans_s;
@@ -13213,13 +13228,13 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
         }
 
         // q4_k x fp32
-        if (src0t == GGML_TYPE_Q4_K && src1t == GGML_TYPE_F32) {
+        if (src0t == GGML_TYPE_Q4_K && src1t == GGML_TYPE_F32 && !use_flat_gemv_for_large_m_q4_K(src0)) {
             ggml_cl_mul_mat_q4_k_f32_adreno(backend, src0, src1, dst);
             return;
         }
 
         // q6_K x fp32
-        if (src0t == GGML_TYPE_Q6_K && src1t == GGML_TYPE_F32) {
+        if (src0t == GGML_TYPE_Q6_K && src1t == GGML_TYPE_F32 && !use_flat_gemv_for_large_m_q6_K(src0)) {
             ggml_cl_mul_mat_q6_K_f32_adreno(backend, src0, src1, dst);
             return;
         }

{llama_cpp_bin-9479.0.0 → llama_cpp_bin-9484.0.0}/llama.cpp/requirements/requirements-convert_legacy_llama.txt

@@ -1,7 +1,7 @@
 numpy~=1.26.4
 sentencepiece>=0.1.98,<0.3.0
 
-transformers==5.5.1
+transformers==4.57.6
 
 gguf>=0.1.0
 protobuf>=4.21.0,<5.0.0

{llama_cpp_bin-9479.0.0 → llama_cpp_bin-9484.0.0}/llama.cpp/requirements/requirements-tool_bench.txt

@@ -1,6 +1,5 @@
 aiohttp~=3.9.3
 pytest~=8.3.3
-huggingface_hub>=1.5.0,<2.0
 matplotlib~=3.10.0
 numpy~=1.26.4
 openai~=2.14.0

{llama_cpp_bin-9479.0.0 → llama_cpp_bin-9484.0.0}/llama.cpp/src/llama-arch.cpp

@@ -135,6 +135,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_MAINCODER,        "maincoder"        },
     { LLM_ARCH_KIMI_LINEAR,      "kimi-linear"      },
     { LLM_ARCH_TALKIE,           "talkie"           },
+    { LLM_ARCH_MELLUM,           "mellum"           },
     { LLM_ARCH_UNKNOWN,          "(unknown)"        },
 };
 
@@ -195,6 +196,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_MOE_LATENT_SIZE,                   "%s.moe_latent_size"                   },
     { LLM_KV_NEXTN_PREDICT_LAYERS,              "%s.nextn_predict_layers"              },
     { LLM_KV_NUM_DEEPSTACK_LAYERS,              "%s.n_deepstack_layers"                },
+    { LLM_KV_HIDDEN_ACT,                        "%s.hidden_activation"                 },
     { LLM_KV_POOLING_TYPE,                      "%s.pooling_type"                      },
     { LLM_KV_LOGIT_SCALE,                       "%s.logit_scale"                       },
     { LLM_KV_DECODER_START_TOKEN_ID,            "%s.decoder_start_token_id"            },

{llama_cpp_bin-9479.0.0 → llama_cpp_bin-9484.0.0}/llama.cpp/src/llama-arch.h

@@ -139,6 +139,7 @@ enum llm_arch {
     LLM_ARCH_MAINCODER,
     LLM_ARCH_KIMI_LINEAR,
     LLM_ARCH_TALKIE,
+    LLM_ARCH_MELLUM,
     LLM_ARCH_UNKNOWN,
 };
 
@@ -199,6 +200,7 @@ enum llm_kv {
     LLM_KV_MOE_LATENT_SIZE,
     LLM_KV_NEXTN_PREDICT_LAYERS,
     LLM_KV_NUM_DEEPSTACK_LAYERS,
+    LLM_KV_HIDDEN_ACT,
     LLM_KV_POOLING_TYPE,
     LLM_KV_LOGIT_SCALE,
     LLM_KV_DECODER_START_TOKEN_ID,

{llama_cpp_bin-9479.0.0 → llama_cpp_bin-9484.0.0}/llama.cpp/src/llama-graph.h

@@ -36,7 +36,8 @@ enum llm_graph_type {
     LLM_GRAPH_TYPE_DECODER_MTP,
 };
 
-enum llm_ffn_op_type {
+enum llm_ffn_op_type : int {
+    LLM_FFN_NONE = 0,           // sentinel: unset; archs must assign before use
     LLM_FFN_SILU,
     LLM_FFN_GELU,
     LLM_FFN_RELU,

{llama_cpp_bin-9479.0.0 → llama_cpp_bin-9484.0.0}/llama.cpp/src/llama-hparams.h

@@ -23,6 +23,9 @@ enum llama_swa_type {
     LLAMA_SWA_TYPE_SYMMETRIC = 3,
 };
 
+// forward declaration; full definition in llama-graph.h
+enum llm_ffn_op_type : int;
+
 struct llama_hparams_posnet {
     uint32_t n_embd;
     uint32_t n_layer;
@@ -227,6 +230,14 @@ struct llama_hparams {
     enum llama_rope_scaling_type rope_scaling_type_train = LLAMA_ROPE_SCALING_TYPE_NONE;
 
 
+    // Resolved FFN gated activation flavor for archs that read
+    // `<arch>.hidden_activation` from the GGUF (e.g. ModernBert derivatives).
+    // Defaults to LLM_FFN_NONE (sentinel = 0); the mapping from the GGUF
+    // string to a real op is done at hparam-load time via
+    // llm_ffn_op_type_from_string() in llama-model.cpp, mirroring how
+    // rope_scaling_type_train is handled.
+    enum llm_ffn_op_type llm_ffn_op;
+
     // Step35: optional per-layer clamps for (Swi)GLU
     std::array<float, LLAMA_MAX_LAYERS> swiglu_clamp_exp; // clamping for expert FFN
     std::array<float, LLAMA_MAX_LAYERS> swiglu_clamp_shexp; // shared expert

{llama_cpp_bin-9479.0.0 → llama_cpp_bin-9484.0.0}/llama.cpp/src/llama-model-saver.cpp

@@ -29,6 +29,7 @@ bool llama_model_saver_supports_arch(llm_arch arch) {
         case LLM_ARCH_APERTUS:
         case LLM_ARCH_MIMO2:
         case LLM_ARCH_STEP35:
+        case LLM_ARCH_MELLUM:
             return false;
         default:
             return true;

{llama_cpp_bin-9479.0.0 → llama_cpp_bin-9484.0.0}/llama.cpp/src/llama-model.cpp

@@ -81,6 +81,8 @@ static llama_model * llama_model_mapping(llm_arch arch, const llama_model_params
             return new llama_model_mpt(params);
         case LLM_ARCH_STABLELM:
             return new llama_model_stablelm(params);
+        case LLM_ARCH_MELLUM:
+            return new llama_model_mellum(params);
         case LLM_ARCH_QWEN:
             return new llama_model_qwen(params);
         case LLM_ARCH_QWEN2:
@@ -764,6 +766,7 @@ const char * llm_type_name(llm_type type) {
         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_12B_A2_5B:     return "12B.A2.5B";
         case LLM_TYPE_16B_A1B:       return "16B.A1B";
         case LLM_TYPE_21B_A3B:       return "21B.A3B";
         case LLM_TYPE_24B_A2B:       return "24B.A2B";
@@ -822,6 +825,28 @@ static llama_rope_scaling_type llama_rope_scaling_type_from_string(const std::st
     return LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED;
 }
 
+// Maps the GGUF `<arch>.hidden_activation` string to the FFN op type used by the
+// graph builders. Only gated activations that map cleanly to llm_ffn_op_type are
+// listed; unrecognized values fall back to GeGLU, which matches the historical
+// default for ModernBert-style architectures.
+static const std::map<std::string, llm_ffn_op_type> LLM_FFN_OP_TYPES_FROM_STRING = {
+    { "gelu",   LLM_FFN_GEGLU  },
+    { "geglu",  LLM_FFN_GEGLU  },
+    { "silu",   LLM_FFN_SWIGLU },
+    { "swish",  LLM_FFN_SWIGLU },
+    { "swiglu", LLM_FFN_SWIGLU },
+    { "relu",   LLM_FFN_RELU   },
+    { "reglu",  LLM_FFN_REGLU  },
+};
+
+llm_ffn_op_type llm_ffn_op_type_from_string(const std::string & name, llm_ffn_op_type fallback) {
+    const auto it = LLM_FFN_OP_TYPES_FROM_STRING.find(name);
+    if (it != LLM_FFN_OP_TYPES_FROM_STRING.end()) {
+        return it->second;
+    }
+    return fallback;
+}
+
 // CPU: ACCEL -> GPU host -> CPU extra -> CPU
 static buft_list_t make_cpu_buft_list(const std::vector<llama_device> & devices, bool use_extra_bufts, bool no_host) {
     buft_list_t buft_list;
@@ -1794,7 +1819,11 @@ void llama_model::print_info() const {
             LLAMA_LOG_INFO("%s: n_ff_shexp            = %d\n",     __func__, hparams.n_ff_shexp);
         }
 
-        if (arch == LLM_ARCH_QWEN3MOE || arch == LLM_ARCH_OPENAI_MOE || arch == LLM_ARCH_QWEN3VLMOE || arch == LLM_ARCH_RND1) {
+        if (arch == LLM_ARCH_MELLUM ||
+                arch == LLM_ARCH_QWEN3MOE ||
+                arch == LLM_ARCH_OPENAI_MOE ||
+                arch == LLM_ARCH_QWEN3VLMOE ||
+                arch == LLM_ARCH_RND1) {
             LLAMA_LOG_INFO("%s: n_ff_exp              = %d\n",     __func__, hparams.n_ff_exp);
         }
 
@@ -2382,6 +2411,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
         case LLM_ARCH_MIMO2:
         case LLM_ARCH_STEP35:
         case LLM_ARCH_TALKIE:
+        case LLM_ARCH_MELLUM:
             return LLAMA_ROPE_TYPE_NEOX;
 
         case LLM_ARCH_QWEN2VL:

{llama_cpp_bin-9479.0.0 → llama_cpp_bin-9484.0.0}/llama.cpp/src/llama-model.h

@@ -116,6 +116,7 @@ enum llm_type {
     LLM_TYPE_A13B,
     LLM_TYPE_7B_A1B,
     LLM_TYPE_8B_A1B, // lfm2moe
+    LLM_TYPE_12B_A2_5B,
     LLM_TYPE_16B_A1B,
     LLM_TYPE_21B_A3B, // Ernie MoE small
     LLM_TYPE_24B_A2B, // lfm2moe
@@ -145,6 +146,10 @@ enum llm_type {
 
 std::string llama_rope_scaling_type_name(llama_rope_scaling_type rope_scaling_type);
 
+// Map a GGUF activation-name string to llm_ffn_op_type. Returns `fallback` if
+// the string is empty or not recognized.
+llm_ffn_op_type llm_ffn_op_type_from_string(const std::string & name, llm_ffn_op_type fallback);
+
 struct llama_layer_posnet {
     // resnet
     struct ggml_tensor * norm1   = nullptr;

{llama_cpp_bin-9479.0.0 → llama_cpp_bin-9484.0.0}/llama.cpp/src/llama-vocab.cpp

@@ -353,6 +353,7 @@ struct llm_tokenizer_bpe : llm_tokenizer {
             case LLAMA_VOCAB_PRE_TYPE_CODESHELL:
             case LLAMA_VOCAB_PRE_TYPE_EXAONE:
             case LLAMA_VOCAB_PRE_TYPE_MINERVA:
+            case LLAMA_VOCAB_PRE_TYPE_MELLUM2:
                 regex_exprs = {
                     "\\p{N}",
                     "'s|'t|'re|'ve|'m|'ll|'d| ?\\p{L}+| ?\\p{N}+| ?[^\\s\\p{L}\\p{N}]+|\\s+(?!\\S)",
@@ -432,6 +433,15 @@ struct llm_tokenizer_bpe : llm_tokenizer {
                     "[^\\r\\n\\p{L}\\p{N}]?((?=[\\p{L}])([^a-z]))*((?=[\\p{L}])([^A-Z]))+(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])?|[^\\r\\n\\p{L}\\p{N}]?((?=[\\p{L}])([^a-z]))+((?=[\\p{L}])([^A-Z]))*(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])?|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n/]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
                 };
                 break;
+            case LLAMA_VOCAB_PRE_TYPE_GRANITE_EMB_MULTI:
+                // Same lookaheads as GPT4O but with \p{M} added so combining marks
+                // (diacritics) attach to their base letters. Avoids excessive
+                // backtracking on scripts that use them heavily (Bengali, Hindi,
+                // Telugu, Thai, ...). See PR #22716 for benchmarks.
+                regex_exprs = {
+                    "[^\\r\\n\\p{L}\\p{N}]?((?=[\\p{L}\\p{M}])([^a-z]))*((?=[\\p{L}\\p{M}])([^A-Z]))+(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])?|[^\\r\\n\\p{L}\\p{N}]?((?=[\\p{L}\\p{M}])([^a-z]))+((?=[\\p{L}\\p{M}])([^A-Z]))*(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])?|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n/]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
+                };
+                break;
             case LLAMA_VOCAB_PRE_TYPE_TINY_AYA:
                 regex_exprs = {
                     // original regex from tokenizer.json: "\\d{1,3}(?=(?:\\d{3})*\\b)"
@@ -2142,7 +2152,8 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                     tokenizer_pre == "jais-2") {
                 pre_type = LLAMA_VOCAB_PRE_TYPE_JAIS2;
             } else if (
-                    tokenizer_pre == "gemma4") {
+                    tokenizer_pre == "gemma4" ||
+                    tokenizer_pre == "granite-embed-multi-311m") {
                 pre_type = LLAMA_VOCAB_PRE_TYPE_GEMMA4;
                 escape_whitespaces = true;
             } else if (
@@ -2252,6 +2263,11 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                 tokenizer_pre == "talkie") {
                 pre_type = LLAMA_VOCAB_PRE_TYPE_GPT4O;
                 clean_spaces = false;
+            } else if (
+                tokenizer_pre == "granite-embed-multi-97m") {
+                pre_type = LLAMA_VOCAB_PRE_TYPE_GRANITE_EMB_MULTI;
+                clean_spaces = false;
+                ignore_merges = true;
             } else if (
                 tokenizer_pre == "tiny_aya") {
                 pre_type = LLAMA_VOCAB_PRE_TYPE_TINY_AYA;
@@ -2310,6 +2326,9 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                 tokenizer_pre == "solar-open") {
                 pre_type = LLAMA_VOCAB_PRE_TYPE_SOLAR_OPEN;
                 clean_spaces = false;
+            } else if (
+                tokenizer_pre == "mellum2") {
+                pre_type = LLAMA_VOCAB_PRE_TYPE_MELLUM2;
             } else {
                 throw std::runtime_error(format("unknown pre-tokenizer type: '%s'", tokenizer_pre.c_str()));
             }

{llama_cpp_bin-9479.0.0 → llama_cpp_bin-9484.0.0}/llama.cpp/src/llama-vocab.h

@@ -8,60 +8,62 @@
 
 // pre-tokenization types
 enum llama_vocab_pre_type {
-    LLAMA_VOCAB_PRE_TYPE_DEFAULT         = 0,
-    LLAMA_VOCAB_PRE_TYPE_LLAMA3          = 1,
-    LLAMA_VOCAB_PRE_TYPE_DEEPSEEK_LLM    = 2,
-    LLAMA_VOCAB_PRE_TYPE_DEEPSEEK_CODER  = 3,
-    LLAMA_VOCAB_PRE_TYPE_FALCON          = 4,
-    LLAMA_VOCAB_PRE_TYPE_MPT             = 5,
-    LLAMA_VOCAB_PRE_TYPE_STARCODER       = 6,
-    LLAMA_VOCAB_PRE_TYPE_GPT2            = 7,
-    LLAMA_VOCAB_PRE_TYPE_REFACT          = 8,
-    LLAMA_VOCAB_PRE_TYPE_COMMAND_R       = 9,
-    LLAMA_VOCAB_PRE_TYPE_STABLELM2       = 10,
-    LLAMA_VOCAB_PRE_TYPE_QWEN2           = 11,
-    LLAMA_VOCAB_PRE_TYPE_OLMO            = 12,
-    LLAMA_VOCAB_PRE_TYPE_DBRX            = 13,
-    LLAMA_VOCAB_PRE_TYPE_SMAUG           = 14,
-    LLAMA_VOCAB_PRE_TYPE_PORO            = 15,
-    LLAMA_VOCAB_PRE_TYPE_CHATGLM3        = 16,
-    LLAMA_VOCAB_PRE_TYPE_CHATGLM4        = 17,
-    LLAMA_VOCAB_PRE_TYPE_VIKING          = 18,
-    LLAMA_VOCAB_PRE_TYPE_JAIS            = 19,
-    LLAMA_VOCAB_PRE_TYPE_TEKKEN          = 20,
-    LLAMA_VOCAB_PRE_TYPE_SMOLLM          = 21,
-    LLAMA_VOCAB_PRE_TYPE_CODESHELL       = 22,
-    LLAMA_VOCAB_PRE_TYPE_BLOOM           = 23,
-    LLAMA_VOCAB_PRE_TYPE_GPT3_FINNISH    = 24,
-    LLAMA_VOCAB_PRE_TYPE_EXAONE          = 25,
-    LLAMA_VOCAB_PRE_TYPE_CHAMELEON       = 26,
-    LLAMA_VOCAB_PRE_TYPE_MINERVA         = 27,
-    LLAMA_VOCAB_PRE_TYPE_DEEPSEEK3_LLM   = 28,
-    LLAMA_VOCAB_PRE_TYPE_GPT4O           = 29,
-    LLAMA_VOCAB_PRE_TYPE_SUPERBPE        = 30,
-    LLAMA_VOCAB_PRE_TYPE_TRILLION        = 31,
-    LLAMA_VOCAB_PRE_TYPE_BAILINGMOE      = 32,
-    LLAMA_VOCAB_PRE_TYPE_LLAMA4          = 33,
-    LLAMA_VOCAB_PRE_TYPE_PIXTRAL         = 34,
-    LLAMA_VOCAB_PRE_TYPE_SEED_CODER      = 35,
-    LLAMA_VOCAB_PRE_TYPE_HUNYUAN         = 36,
-    LLAMA_VOCAB_PRE_TYPE_KIMI_K2         = 37,
-    LLAMA_VOCAB_PRE_TYPE_HUNYUAN_DENSE   = 38,
-    LLAMA_VOCAB_PRE_TYPE_GROK_2          = 39,
-    LLAMA_VOCAB_PRE_TYPE_GRANITE_DOCLING = 40,
-    LLAMA_VOCAB_PRE_TYPE_MINIMAX_M2      = 41,
-    LLAMA_VOCAB_PRE_TYPE_AFMOE           = 42,
-    LLAMA_VOCAB_PRE_TYPE_SOLAR_OPEN      = 43,
-    LLAMA_VOCAB_PRE_TYPE_YOUTU           = 44,
-    LLAMA_VOCAB_PRE_TYPE_EXAONE_MOE      = 45,
-    LLAMA_VOCAB_PRE_TYPE_QWEN35          = 46,
-    LLAMA_VOCAB_PRE_TYPE_TINY_AYA        = 47,
-    LLAMA_VOCAB_PRE_TYPE_JOYAI_LLM       = 48,
-    LLAMA_VOCAB_PRE_TYPE_JAIS2           = 49,
-    LLAMA_VOCAB_PRE_TYPE_GEMMA4          = 50,
-    LLAMA_VOCAB_PRE_TYPE_SARVAM_MOE      = 51,
-    LLAMA_VOCAB_PRE_TYPE_MINICPM5        = 52,
-    LLAMA_VOCAB_PRE_TYPE_WHITESPACE      = 53,
+    LLAMA_VOCAB_PRE_TYPE_DEFAULT           = 0,
+    LLAMA_VOCAB_PRE_TYPE_LLAMA3            = 1,
+    LLAMA_VOCAB_PRE_TYPE_DEEPSEEK_LLM      = 2,
+    LLAMA_VOCAB_PRE_TYPE_DEEPSEEK_CODER    = 3,
+    LLAMA_VOCAB_PRE_TYPE_FALCON            = 4,
+    LLAMA_VOCAB_PRE_TYPE_MPT               = 5,
+    LLAMA_VOCAB_PRE_TYPE_STARCODER         = 6,
+    LLAMA_VOCAB_PRE_TYPE_GPT2              = 7,
+    LLAMA_VOCAB_PRE_TYPE_REFACT            = 8,
+    LLAMA_VOCAB_PRE_TYPE_COMMAND_R         = 9,
+    LLAMA_VOCAB_PRE_TYPE_STABLELM2         = 10,
+    LLAMA_VOCAB_PRE_TYPE_QWEN2             = 11,
+    LLAMA_VOCAB_PRE_TYPE_OLMO              = 12,
+    LLAMA_VOCAB_PRE_TYPE_DBRX              = 13,
+    LLAMA_VOCAB_PRE_TYPE_SMAUG             = 14,
+    LLAMA_VOCAB_PRE_TYPE_PORO              = 15,
+    LLAMA_VOCAB_PRE_TYPE_CHATGLM3          = 16,
+    LLAMA_VOCAB_PRE_TYPE_CHATGLM4          = 17,
+    LLAMA_VOCAB_PRE_TYPE_VIKING            = 18,
+    LLAMA_VOCAB_PRE_TYPE_JAIS              = 19,
+    LLAMA_VOCAB_PRE_TYPE_TEKKEN            = 20,
+    LLAMA_VOCAB_PRE_TYPE_SMOLLM            = 21,
+    LLAMA_VOCAB_PRE_TYPE_CODESHELL         = 22,
+    LLAMA_VOCAB_PRE_TYPE_BLOOM             = 23,
+    LLAMA_VOCAB_PRE_TYPE_GPT3_FINNISH      = 24,
+    LLAMA_VOCAB_PRE_TYPE_EXAONE            = 25,
+    LLAMA_VOCAB_PRE_TYPE_CHAMELEON         = 26,
+    LLAMA_VOCAB_PRE_TYPE_MINERVA           = 27,
+    LLAMA_VOCAB_PRE_TYPE_DEEPSEEK3_LLM     = 28,
+    LLAMA_VOCAB_PRE_TYPE_GPT4O             = 29,
+    LLAMA_VOCAB_PRE_TYPE_SUPERBPE          = 30,
+    LLAMA_VOCAB_PRE_TYPE_TRILLION          = 31,
+    LLAMA_VOCAB_PRE_TYPE_BAILINGMOE        = 32,
+    LLAMA_VOCAB_PRE_TYPE_LLAMA4            = 33,
+    LLAMA_VOCAB_PRE_TYPE_PIXTRAL           = 34,
+    LLAMA_VOCAB_PRE_TYPE_SEED_CODER        = 35,
+    LLAMA_VOCAB_PRE_TYPE_HUNYUAN           = 36,
+    LLAMA_VOCAB_PRE_TYPE_KIMI_K2           = 37,
+    LLAMA_VOCAB_PRE_TYPE_HUNYUAN_DENSE     = 38,
+    LLAMA_VOCAB_PRE_TYPE_GROK_2            = 39,
+    LLAMA_VOCAB_PRE_TYPE_GRANITE_DOCLING   = 40,
+    LLAMA_VOCAB_PRE_TYPE_MINIMAX_M2        = 41,
+    LLAMA_VOCAB_PRE_TYPE_AFMOE             = 42,
+    LLAMA_VOCAB_PRE_TYPE_SOLAR_OPEN        = 43,
+    LLAMA_VOCAB_PRE_TYPE_YOUTU             = 44,
+    LLAMA_VOCAB_PRE_TYPE_EXAONE_MOE        = 45,
+    LLAMA_VOCAB_PRE_TYPE_QWEN35            = 46,
+    LLAMA_VOCAB_PRE_TYPE_TINY_AYA          = 47,
+    LLAMA_VOCAB_PRE_TYPE_JOYAI_LLM         = 48,
+    LLAMA_VOCAB_PRE_TYPE_JAIS2             = 49,
+    LLAMA_VOCAB_PRE_TYPE_GEMMA4            = 50,
+    LLAMA_VOCAB_PRE_TYPE_SARVAM_MOE        = 51,
+    LLAMA_VOCAB_PRE_TYPE_MINICPM5          = 52,
+    LLAMA_VOCAB_PRE_TYPE_WHITESPACE        = 53,
+    LLAMA_VOCAB_PRE_TYPE_GRANITE_EMB_MULTI = 54,
+    LLAMA_VOCAB_PRE_TYPE_MELLUM2           = 55,
 };
 
 struct LLM_KV;