diff --git "a/examples/talk-llama/llama.cpp" "b/examples/talk-llama/llama.cpp"
--- "a/examples/talk-llama/llama.cpp"
+++ "b/examples/talk-llama/llama.cpp"
@@ -7,10 +7,12 @@
 #include "ggml-alloc.h"
 #include "ggml-backend.h"
 
+#ifdef GGML_USE_RPC
+#  include "ggml-rpc.h"
+#endif
+
 #ifdef GGML_USE_CUDA
 #  include "ggml-cuda.h"
-#elif defined(GGML_USE_CLBLAST)
-#  include "ggml-opencl.h"
 #elif defined(GGML_USE_VULKAN)
 #  include "ggml-vulkan.h"
 #elif defined(GGML_USE_SYCL)
@@ -19,19 +21,16 @@
 #   include "ggml-kompute.h"
 #endif
 
+#ifdef GGML_USE_BLAS
+#  include "ggml-blas.h"
+#endif
+
 #ifdef GGML_USE_METAL
 #  include "ggml-metal.h"
 #endif
-#ifdef GGML_USE_MPI
-#  include "ggml-mpi.h"
-#endif
-#ifndef QK_K
-#  ifdef GGML_QKK_64
-#    define QK_K 64
-#  else
-#    define QK_K 256
-#  endif
-#endif
+
+// TODO: replace with ggml API call
+#define QK_K 256
 
 #ifdef __has_include
     #if __has_include(<unistd.h>)
@@ -106,14 +105,14 @@
 #endif
 
 #define LLAMA_MAX_NODES   8192
-#define LLAMA_MAX_EXPERTS 60
+#define LLAMA_MAX_EXPERTS 160
 
 //
 // logging
 //
 
 LLAMA_ATTRIBUTE_FORMAT(2, 3)
-static void llama_log_internal        (ggml_log_level level, const char* format, ...);
+static void llama_log_internal        (ggml_log_level level, const char * format, ...);
 static void llama_log_callback_default(ggml_log_level level, const char * text, void * user_data);
 
 #define LLAMA_LOG_INFO(...)  llama_log_internal(GGML_LOG_LEVEL_INFO , __VA_ARGS__)
@@ -201,7 +200,6 @@ enum llm_arch {
     LLM_ARCH_GPTNEOX,
     LLM_ARCH_MPT,
     LLM_ARCH_STARCODER,
-    LLM_ARCH_PERSIMMON,
     LLM_ARCH_REFACT,
     LLM_ARCH_BERT,
     LLM_ARCH_NOMIC_BERT,
@@ -225,6 +223,8 @@ enum llm_arch {
     LLM_ARCH_COMMAND_R,
     LLM_ARCH_DBRX,
     LLM_ARCH_OLMO,
+    LLM_ARCH_ARCTIC,
+    LLM_ARCH_DEEPSEEK2,
     LLM_ARCH_UNKNOWN,
 };
 
@@ -238,7 +238,6 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_MPT,             "mpt"          },
     { LLM_ARCH_BAICHUAN,        "baichuan"     },
     { LLM_ARCH_STARCODER,       "starcoder"    },
-    { LLM_ARCH_PERSIMMON,       "persimmon"    },
     { LLM_ARCH_REFACT,          "refact"       },
     { LLM_ARCH_BERT,            "bert"         },
     { LLM_ARCH_NOMIC_BERT,      "nomic-bert"   },
@@ -262,6 +261,8 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_COMMAND_R,       "command-r"    },
     { LLM_ARCH_DBRX,            "dbrx"         },
     { LLM_ARCH_OLMO,            "olmo"         },
+    { LLM_ARCH_ARCTIC,          "arctic"       },
+    { LLM_ARCH_DEEPSEEK2,       "deepseek2"    },
     { LLM_ARCH_UNKNOWN,         "(unknown)"    },
 };
 
@@ -282,11 +283,15 @@ enum llm_kv {
     LLM_KV_CONTEXT_LENGTH,
     LLM_KV_EMBEDDING_LENGTH,
     LLM_KV_BLOCK_COUNT,
+    LLM_KV_LEADING_DENSE_BLOCK_COUNT,
     LLM_KV_FEED_FORWARD_LENGTH,
+    LLM_KV_EXPERT_FEED_FORWARD_LENGTH,
     LLM_KV_USE_PARALLEL_RESIDUAL,
     LLM_KV_TENSOR_DATA_LAYOUT,
     LLM_KV_EXPERT_COUNT,
     LLM_KV_EXPERT_USED_COUNT,
+    LLM_KV_EXPERT_SHARED_COUNT,
+    LLM_KV_EXPERT_WEIGHTS_SCALE,
     LLM_KV_POOLING_TYPE,
     LLM_KV_LOGIT_SCALE,
 
@@ -299,14 +304,18 @@ enum llm_kv {
     LLM_KV_ATTENTION_LAYERNORM_EPS,
     LLM_KV_ATTENTION_LAYERNORM_RMS_EPS,
     LLM_KV_ATTENTION_CAUSAL,
+    LLM_KV_ATTENTION_Q_LORA_RANK,
+    LLM_KV_ATTENTION_KV_LORA_RANK,
 
     LLM_KV_ROPE_DIMENSION_COUNT,
     LLM_KV_ROPE_FREQ_BASE,
     LLM_KV_ROPE_SCALE_LINEAR,
     LLM_KV_ROPE_SCALING_TYPE,
     LLM_KV_ROPE_SCALING_FACTOR,
+    LLM_KV_ROPE_SCALING_ATTN_FACTOR,
     LLM_KV_ROPE_SCALING_ORIG_CTX_LEN,
     LLM_KV_ROPE_SCALING_FINETUNED,
+    LLM_KV_ROPE_SCALING_YARN_LOG_MUL,
 
     LLM_KV_SPLIT_NO,
     LLM_KV_SPLIT_COUNT,
@@ -355,17 +364,21 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_GENERAL_SOURCE_URL,            "general.source.url"                    },
     { LLM_KV_GENERAL_SOURCE_HF_REPO,        "general.source.huggingface.repository" },
 
-    { LLM_KV_VOCAB_SIZE,                    "%s.vocab_size"            },
-    { LLM_KV_CONTEXT_LENGTH,                "%s.context_length"        },
-    { LLM_KV_EMBEDDING_LENGTH,              "%s.embedding_length"      },
-    { LLM_KV_BLOCK_COUNT,                   "%s.block_count"           },
-    { LLM_KV_FEED_FORWARD_LENGTH,           "%s.feed_forward_length"   },
-    { LLM_KV_USE_PARALLEL_RESIDUAL,         "%s.use_parallel_residual" },
-    { LLM_KV_TENSOR_DATA_LAYOUT,            "%s.tensor_data_layout"    },
-    { LLM_KV_EXPERT_COUNT,                  "%s.expert_count"          },
-    { LLM_KV_EXPERT_USED_COUNT,             "%s.expert_used_count"     },
-    { LLM_KV_POOLING_TYPE ,                 "%s.pooling_type"          },
-    { LLM_KV_LOGIT_SCALE,                   "%s.logit_scale"           },
+    { LLM_KV_VOCAB_SIZE,                    "%s.vocab_size"                 },
+    { LLM_KV_CONTEXT_LENGTH,                "%s.context_length"             },
+    { LLM_KV_EMBEDDING_LENGTH,              "%s.embedding_length"           },
+    { LLM_KV_BLOCK_COUNT,                   "%s.block_count"                },
+    { LLM_KV_LEADING_DENSE_BLOCK_COUNT,     "%s.leading_dense_block_count"  },
+    { LLM_KV_FEED_FORWARD_LENGTH,           "%s.feed_forward_length"        },
+    { LLM_KV_EXPERT_FEED_FORWARD_LENGTH,    "%s.expert_feed_forward_length" },
+    { LLM_KV_USE_PARALLEL_RESIDUAL,         "%s.use_parallel_residual"      },
+    { LLM_KV_TENSOR_DATA_LAYOUT,            "%s.tensor_data_layout"         },
+    { 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_WEIGHTS_SCALE,          "%s.expert_weights_scale"       },
+    { LLM_KV_POOLING_TYPE ,                 "%s.pooling_type"               },
+    { LLM_KV_LOGIT_SCALE,                   "%s.logit_scale"                },
 
     { LLM_KV_ATTENTION_HEAD_COUNT,          "%s.attention.head_count"             },
     { LLM_KV_ATTENTION_HEAD_COUNT_KV,       "%s.attention.head_count_kv"          },
@@ -376,14 +389,18 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_ATTENTION_LAYERNORM_EPS,       "%s.attention.layer_norm_epsilon"     },
     { LLM_KV_ATTENTION_LAYERNORM_RMS_EPS,   "%s.attention.layer_norm_rms_epsilon" },
     { LLM_KV_ATTENTION_CAUSAL,              "%s.attention.causal"                 },
+    { LLM_KV_ATTENTION_Q_LORA_RANK,         "%s.attention.q_lora_rank"            },
+    { LLM_KV_ATTENTION_KV_LORA_RANK,        "%s.attention.kv_lora_rank"           },
 
     { LLM_KV_ROPE_DIMENSION_COUNT,          "%s.rope.dimension_count"                 },
     { LLM_KV_ROPE_FREQ_BASE,                "%s.rope.freq_base"                       },
     { LLM_KV_ROPE_SCALE_LINEAR,             "%s.rope.scale_linear"                    },
     { LLM_KV_ROPE_SCALING_TYPE,             "%s.rope.scaling.type"                    },
     { LLM_KV_ROPE_SCALING_FACTOR,           "%s.rope.scaling.factor"                  },
+    { LLM_KV_ROPE_SCALING_ATTN_FACTOR,      "%s.rope.scaling.attn_factor"             },
     { LLM_KV_ROPE_SCALING_ORIG_CTX_LEN,     "%s.rope.scaling.original_context_length" },
     { LLM_KV_ROPE_SCALING_FINETUNED,        "%s.rope.scaling.finetuned"               },
+    { LLM_KV_ROPE_SCALING_YARN_LOG_MUL,     "%s.rope.scaling.yarn_log_multiplier"     },
 
     { LLM_KV_SPLIT_NO,                      "split.no"            },
     { LLM_KV_SPLIT_COUNT,                   "split.count"         },
@@ -437,6 +454,8 @@ enum llm_tensor {
     LLM_TENSOR_OUTPUT,
     LLM_TENSOR_OUTPUT_NORM,
     LLM_TENSOR_ROPE_FREQS,
+    LLM_TENSOR_ROPE_FACTORS_LONG,
+    LLM_TENSOR_ROPE_FACTORS_SHORT,
     LLM_TENSOR_ATTN_Q,
     LLM_TENSOR_ATTN_K,
     LLM_TENSOR_ATTN_V,
@@ -456,6 +475,7 @@ enum llm_tensor {
     LLM_TENSOR_FFN_DOWN_EXP,  // split experts for backward compatibility
     LLM_TENSOR_FFN_GATE_EXP,
     LLM_TENSOR_FFN_UP_EXP,
+    LLM_TENSOR_FFN_NORM_EXPS,
     LLM_TENSOR_FFN_DOWN_EXPS, // merged experts
     LLM_TENSOR_FFN_GATE_EXPS,
     LLM_TENSOR_FFN_UP_EXPS,
@@ -472,6 +492,12 @@ enum llm_tensor {
     LLM_TENSOR_SSM_A,
     LLM_TENSOR_SSM_D,
     LLM_TENSOR_SSM_OUT,
+    LLM_TENSOR_ATTN_Q_A,
+    LLM_TENSOR_ATTN_Q_B,
+    LLM_TENSOR_ATTN_KV_A_MQA,
+    LLM_TENSOR_ATTN_KV_B,
+    LLM_TENSOR_ATTN_Q_A_NORM,
+    LLM_TENSOR_ATTN_KV_A_NORM,
 };
 
 static const std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NAMES = {
@@ -594,23 +620,6 @@ static const std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NA
             { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
         },
     },
-    {
-        LLM_ARCH_PERSIMMON,
-        {
-            { 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_QKV,        "blk.%d.attn_qkv"},
-            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output"},
-            { LLM_TENSOR_ATTN_Q_NORM,     "blk.%d.attn_q_norm"},
-            { LLM_TENSOR_ATTN_K_NORM,     "blk.%d.attn_k_norm"},
-            { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm"},
-            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down"},
-            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up"},
-            { LLM_TENSOR_ATTN_ROT_EMBD,   "blk.%d.attn_rot_embd"},
-        },
-    },
     {
         LLM_ARCH_MPT,
         {
@@ -699,6 +708,7 @@ static const std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NA
             { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
             { LLM_TENSOR_TOKEN_EMBD_NORM, "token_embd_norm" },
             { LLM_TENSOR_TOKEN_TYPES,     "token_types" },
+            { LLM_TENSOR_ATTN_NORM_2,     "blk.%d.attn_norm_2" },
             { LLM_TENSOR_ATTN_OUT_NORM,   "blk.%d.attn_output_norm" },
             { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
             { LLM_TENSOR_ATTN_Q_NORM,     "blk.%d.attn_q_norm" },
@@ -821,18 +831,20 @@ static const std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NA
     {
         LLM_ARCH_PHI3,
         {
-            { 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_QKV,        "blk.%d.attn_qkv" },
-            { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
-            { LLM_TENSOR_ATTN_K,          "blk.%d.attn_k" },
-            { LLM_TENSOR_ATTN_V,          "blk.%d.attn_v" },
-            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
-            { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm" },
-            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
-            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+            { LLM_TENSOR_TOKEN_EMBD,         "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,        "output_norm" },
+            { LLM_TENSOR_OUTPUT,             "output" },
+            { LLM_TENSOR_ROPE_FACTORS_LONG,  "rope_factors_long" },
+            { LLM_TENSOR_ROPE_FACTORS_SHORT, "rope_factors_short" },
+            { LLM_TENSOR_ATTN_NORM,          "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_QKV,           "blk.%d.attn_qkv" },
+            { LLM_TENSOR_ATTN_Q,             "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_K,             "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_V,             "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,           "blk.%d.attn_output" },
+            { LLM_TENSOR_FFN_NORM,           "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_DOWN,           "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,             "blk.%d.ffn_up" },
         },
     },
     {
@@ -1048,6 +1060,57 @@ static const std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NA
             { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
         },
     },
+    {
+        LLM_ARCH_ARCTIC,
+        {
+            { 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,          "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_K,          "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_V,          "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_FFN_GATE_INP,    "blk.%d.ffn_gate_inp" },
+            { 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_NORM_EXPS,   "blk.%d.ffn_norm_exps" },
+            { 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_ARCH_DEEPSEEK2,
+        {
+            { 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_A_NORM,      "blk.%d.attn_q_a_norm" },
+            { LLM_TENSOR_ATTN_KV_A_NORM,     "blk.%d.attn_kv_a_norm" },
+            { LLM_TENSOR_ATTN_Q,             "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_Q_A,           "blk.%d.attn_q_a" },
+            { LLM_TENSOR_ATTN_Q_B,           "blk.%d.attn_q_b" },
+            { LLM_TENSOR_ATTN_KV_A_MQA,      "blk.%d.attn_kv_a_mqa" },
+            { LLM_TENSOR_ATTN_KV_B,          "blk.%d.attn_kv_b" },
+            { LLM_TENSOR_ATTN_OUT,           "blk.%d.attn_output" },
+            { LLM_TENSOR_FFN_NORM,           "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE,           "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_UP,             "blk.%d.ffn_up" },
+            { LLM_TENSOR_FFN_DOWN,           "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_GATE_INP,       "blk.%d.ffn_gate_inp" },
+            { 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_INP_SHEXP, "blk.%d.ffn_gate_inp_shexp" },
+            { 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_ARCH_UNKNOWN,
         {
@@ -1642,12 +1705,13 @@ struct llama_mlock {
 };
 using llama_mlocks = std::vector<std::unique_ptr<llama_mlock>>;
 
-static std::string llama_token_to_piece(const struct llama_context * ctx, llama_token token, bool special) {
+// NOTE: avoid ever using this except for building the token_to_piece caches
+static std::string llama_token_to_piece(const struct llama_model * model, llama_token token, bool special) {
     std::vector<char> result(8, 0);
-    const int n_tokens = llama_token_to_piece(llama_get_model(ctx), token, result.data(), result.size(), special);
+    const int n_tokens = llama_token_to_piece(model, token, result.data(), result.size(), special);
     if (n_tokens < 0) {
         result.resize(-n_tokens);
-        int check = llama_token_to_piece(llama_get_model(ctx), token, result.data(), result.size(), special);
+        int check = llama_token_to_piece(model, token, result.data(), result.size(), special);
         GGML_ASSERT(check == -n_tokens);
     }
     else {
@@ -1685,91 +1749,6 @@ static ggml_backend_buffer_type_t llama_default_buffer_type_cpu(bool host_buffer
     GGML_UNUSED(host_buffer);
 }
 
-static ggml_backend_buffer_type_t llama_default_buffer_type_offload(int gpu) {
-    ggml_backend_buffer_type_t buft = nullptr;
-
-#ifdef GGML_USE_METAL
-    buft = ggml_backend_metal_buffer_type();
-#elif defined(GGML_USE_CUDA)
-    buft = ggml_backend_cuda_buffer_type(gpu);
-#elif defined(GGML_USE_VULKAN)
-    buft = ggml_backend_vk_buffer_type(gpu);
-#elif defined(GGML_USE_SYCL)
-    buft = ggml_backend_sycl_buffer_type(gpu);
-#elif defined(GGML_USE_CLBLAST)
-    buft = ggml_backend_opencl_buffer_type();
-#elif defined(GGML_USE_KOMPUTE)
-    buft = ggml_backend_kompute_buffer_type(gpu);
-    if (buft == nullptr) {
-        LLAMA_LOG_WARN("%s: cannot use GPU %d, check `vulkaninfo --summary`\n", __func__, gpu);
-    }
-#endif
-
-    if (buft == nullptr) {
-        buft = llama_default_buffer_type_cpu(true);
-    }
-    return buft;
-
-    GGML_UNUSED(gpu);
-}
-
-static ggml_backend_buffer_type_t llama_default_buffer_type_split(int fallback_gpu, const float * tensor_split) {
-    ggml_backend_buffer_type_t buft = nullptr;
-
-#ifdef GGML_USE_CUDA
-    if (ggml_backend_cuda_get_device_count() > 1) {
-        buft = ggml_backend_cuda_split_buffer_type(tensor_split);
-    }
-#endif
-
-#ifdef GGML_USE_SYCL
-    if (ggml_backend_sycl_get_device_count() > 1) {
-        buft = ggml_backend_sycl_split_buffer_type(tensor_split);
-    }
-#endif
-
-    if (buft == nullptr) {
-        buft = llama_default_buffer_type_offload(fallback_gpu);
-    }
-    return buft;
-
-    GGML_UNUSED(tensor_split);
-}
-
-static size_t llama_get_device_count() {
-#if defined(GGML_USE_CUDA)
-    return ggml_backend_cuda_get_device_count();
-#elif defined(GGML_USE_SYCL)
-    return ggml_backend_sycl_get_device_count();
-#elif defined(GGML_USE_VULKAN)
-    return ggml_backend_vk_get_device_count();
-#else
-    return 1;
-#endif
-}
-
-static size_t llama_get_device_memory(int device) {
-#if defined(GGML_USE_CUDA)
-    size_t total;
-    size_t free;
-    ggml_backend_cuda_get_device_memory(device, &free, &total);
-    return free;
-#elif defined(GGML_USE_SYCL)
-    size_t total;
-    size_t free;
-    ggml_backend_sycl_get_device_memory(device, &free, &total);
-    return free;
-#elif defined(GGML_USE_VULKAN)
-    size_t total;
-    size_t free;
-    ggml_backend_vk_get_device_memory(device, &free, &total);
-    return free;
-#else
-    return 1;
-    GGML_UNUSED(device);
-#endif
-}
-
 //
 // globals
 //
@@ -1778,6 +1757,8 @@ struct llama_state {
     llama_state() {
 #ifdef GGML_USE_METAL
         ggml_backend_metal_log_set_callback(log_callback, log_callback_user_data);
+#elif defined(GGML_USE_CUDA)
+        ggml_backend_cuda_log_set_callback(log_callback, log_callback_user_data);
 #endif
     }
 
@@ -1791,23 +1772,31 @@ static llama_state g_state;
 // available llama models
 enum e_model {
     MODEL_UNKNOWN,
+    MODEL_14M,
     MODEL_17M,
     MODEL_22M,
     MODEL_33M,
+    MODEL_70M,
     MODEL_109M,
     MODEL_137M,
+    MODEL_160M,
     MODEL_335M,
+    MODEL_410M,
     MODEL_0_5B,
     MODEL_1B,
+    MODEL_1_4B,
     MODEL_2B,
+    MODEL_2_8B,
     MODEL_3B,
     MODEL_4B,
+    MODEL_6_9B,
     MODEL_7B,
     MODEL_8B,
     MODEL_12B,
     MODEL_13B,
     MODEL_14B,
     MODEL_15B,
+    MODEL_16B,
     MODEL_20B,
     MODEL_30B,
     MODEL_34B,
@@ -1815,6 +1804,7 @@ enum e_model {
     MODEL_40B,
     MODEL_65B,
     MODEL_70B,
+    MODEL_236B,
     MODEL_314B,
     MODEL_SMALL,
     MODEL_MEDIUM,
@@ -1824,6 +1814,7 @@ enum e_model {
     MODEL_8x7B,
     MODEL_8x22B,
     MODEL_16x12B,
+    MODEL_10B_128x3_66B,
 };
 
 static const size_t kiB = 1024;
@@ -1833,6 +1824,7 @@ static const size_t GiB = 1024*MiB;
 struct llama_hparams {
     bool vocab_only;
     bool rope_finetuned;
+    bool use_par_res;
 
     uint32_t n_vocab;
     uint32_t n_ctx_train; // context size the model was trained on
@@ -1848,12 +1840,21 @@ struct llama_hparams {
     uint32_t n_expert_used = 0;
     uint32_t n_vocab_type = 0; // for BERT-style token types
 
+    uint32_t n_layer_dense_lead = 0;
+    uint32_t n_lora_q = 0;
+    uint32_t n_lora_kv = 0;
+    uint32_t n_ff_exp = 0;
+    uint32_t n_expert_shared = 0;
+    float    expert_weights_scale = 0.0;
+
     float f_norm_eps;
     float f_norm_rms_eps;
 
+    float    rope_attn_factor = 1.0f;
     float    rope_freq_base_train;
     float    rope_freq_scale_train;
-    uint32_t n_yarn_orig_ctx;
+    uint32_t n_ctx_orig_yarn;
+    float    rope_yarn_log_mul;
 
     // for State Space Models
     uint32_t ssm_d_conv  = 0;
@@ -1887,8 +1888,14 @@ struct llama_hparams {
         if (this->n_expert      != other.n_expert)      return true;
         if (this->n_expert_used != other.n_expert_used) return true;
 
+        if (this->n_layer_dense_lead != other.n_layer_dense_lead) return true;
+        if (this->n_lora_q           != other.n_lora_q)           return true;
+        if (this->n_lora_kv          != other.n_lora_kv)          return true;
+        if (this->n_ff_exp           != other.n_ff_exp)           return true;
+        if (this->n_expert_shared    != other.n_expert_shared)    return true;
+
         if (this->rope_finetuned  != other.rope_finetuned)  return true;
-        if (this->n_yarn_orig_ctx != other.n_yarn_orig_ctx) return true;
+        if (this->n_ctx_orig_yarn != other.n_ctx_orig_yarn) return true;
 
         if (this->ssm_d_conv  != other.ssm_d_conv)  return true;
         if (this->ssm_d_inner != other.ssm_d_inner) return true;
@@ -1899,8 +1906,11 @@ struct llama_hparams {
 
         if (!is_float_close(this->f_norm_eps,            other.f_norm_eps,            EPSILON)) return true;
         if (!is_float_close(this->f_norm_rms_eps,        other.f_norm_rms_eps,        EPSILON)) return true;
+        if (!is_float_close(this->rope_attn_factor,      other.rope_attn_factor,      EPSILON)) return true;
         if (!is_float_close(this->rope_freq_base_train,  other.rope_freq_base_train,  EPSILON)) return true;
         if (!is_float_close(this->rope_freq_scale_train, other.rope_freq_scale_train, EPSILON)) return true;
+        if (!is_float_close(this->expert_weights_scale,  other.expert_weights_scale,  EPSILON)) return true;
+        if (!is_float_close(this->rope_yarn_log_mul,     other.rope_yarn_log_mul,     EPSILON)) return true;
 
         return false;
     }
@@ -1944,7 +1954,7 @@ struct llama_cparams {
     float rope_freq_base;
     float rope_freq_scale;
 
-    uint32_t n_yarn_orig_ctx;
+    uint32_t n_ctx_orig_yarn;
     // These hyperparameters are not exposed in GGUF, because all
     // existing YaRN models use the same values for them.
     float yarn_ext_factor;
@@ -1976,6 +1986,8 @@ struct llama_layer {
     struct ggml_tensor * attn_k_norm_b;
     struct ggml_tensor * attn_out_norm;
     struct ggml_tensor * attn_out_norm_b;
+    struct ggml_tensor * attn_q_a_norm;
+    struct ggml_tensor * attn_kv_a_norm;
 
     // attention
     struct ggml_tensor * wq;
@@ -1983,6 +1995,10 @@ struct llama_layer {
     struct ggml_tensor * wv;
     struct ggml_tensor * wo;
     struct ggml_tensor * wqkv;
+    struct ggml_tensor * wq_a;
+    struct ggml_tensor * wq_b;
+    struct ggml_tensor * wkv_a_mqa;
+    struct ggml_tensor * wkv_b;
 
     // attention bias
     struct ggml_tensor * bq;
@@ -1996,6 +2012,7 @@ struct llama_layer {
     struct ggml_tensor * ffn_norm_b;
     struct ggml_tensor * layer_out_norm;
     struct ggml_tensor * layer_out_norm_b;
+    struct ggml_tensor * ffn_norm_exps;
 
     // ff
     struct ggml_tensor * ffn_gate; // w1
@@ -2015,8 +2032,9 @@ struct llama_layer {
     struct ggml_tensor * ffn_up_shexp;
 
     // ff bias
-    struct ggml_tensor * ffn_down_b; // b2
-    struct ggml_tensor * ffn_up_b;   // b3
+    struct ggml_tensor * ffn_gate_b = nullptr;
+    struct ggml_tensor * ffn_down_b = nullptr; // b2
+    struct ggml_tensor * ffn_up_b   = nullptr; // b3
     struct ggml_tensor * ffn_act;
 
     // mamba proj
@@ -2033,6 +2051,10 @@ struct llama_layer {
     // mamba bias
     struct ggml_tensor * ssm_conv1d_b;
     struct ggml_tensor * ssm_dt_b;
+
+    // long rope factors
+    struct ggml_tensor * rope_long  = nullptr;
+    struct ggml_tensor * rope_short = nullptr;
 };
 
 struct llama_kv_cell {
@@ -2130,12 +2152,12 @@ struct llama_control_vector {
 struct llama_vocab {
     using id    = int32_t;
     using token = std::string;
-    using ttype = llama_token_type;
+    using tattr = llama_token_attr;
 
     struct token_data {
         token text;
         float score;
-        ttype type;
+        tattr attr;
     };
 
     enum llama_vocab_type     type     = LLAMA_VOCAB_TYPE_SPM;
@@ -2144,7 +2166,8 @@ struct llama_vocab {
     std::unordered_map<token, id> token_to_id;
     std::vector<token_data>       id_to_token;
 
-    std::unordered_map<token, id> special_tokens_cache;
+    std::vector<id>    cache_special_tokens;
+    std::vector<token> cache_token_to_piece; // llama_token_to_piece(special = true);
 
     std::map<std::pair<std::string, std::string>, int> bpe_ranks;
 
@@ -2210,6 +2233,8 @@ struct llama_model {
     int main_gpu;
     int n_gpu_layers;
 
+    std::vector<std::string> rpc_servers;
+
     // gguf metadata
     std::unordered_map<std::string, std::string> gguf_kv;
 
@@ -2278,9 +2303,13 @@ struct llama_context {
     std::vector<ggml_backend_t> backends;
 #ifdef GGML_USE_METAL
     ggml_backend_t backend_metal = nullptr;
+#endif
+#ifdef GGML_USE_BLAS
+    ggml_backend_t backend_blas = nullptr;
 #endif
     ggml_backend_t backend_cpu = nullptr;
 
+
     const llama_model & model;
 
     // key + value cache for the self attention
@@ -2347,11 +2376,114 @@ struct llama_context {
 
     // control vectors
     struct llama_control_vector cvec;
+};
+
+static size_t llama_get_device_count(const llama_model & model) {
+    size_t count = 1;
+#if defined(GGML_USE_CUDA)
+    count = ggml_backend_cuda_get_device_count();
+#elif defined(GGML_USE_SYCL)
+    count = ggml_backend_sycl_get_device_count();
+#elif defined(GGML_USE_VULKAN)
+    count = ggml_backend_vk_get_device_count();
+#endif
+#if defined(GGML_USE_RPC)
+    count += model.rpc_servers.size();
+#endif
+    return count;
+    GGML_UNUSED(model);
+}
+
+static ggml_backend_buffer_type_t llama_default_buffer_type_offload(const llama_model & model, int gpu) {
+    ggml_backend_buffer_type_t buft = nullptr;
 
-#ifdef GGML_USE_MPI
-    ggml_mpi_context * ctx_mpi = NULL;
+#if defined(GGML_USE_RPC)
+    int dev_count = (int)llama_get_device_count(model);
+    int rpc_count = (int)model.rpc_servers.size();
+    if (gpu >= dev_count - rpc_count) {
+        const char * endpoint = model.rpc_servers[gpu - dev_count + rpc_count].c_str();
+        return ggml_backend_rpc_buffer_type(endpoint);
+    }
 #endif
-};
+#if defined(GGML_USE_METAL)
+    buft = ggml_backend_metal_buffer_type();
+#elif defined(GGML_USE_CUDA)
+    buft = ggml_backend_cuda_buffer_type(gpu);
+#elif defined(GGML_USE_VULKAN)
+    buft = ggml_backend_vk_buffer_type(gpu);
+#elif defined(GGML_USE_SYCL)
+    buft = ggml_backend_sycl_buffer_type(gpu);
+#elif defined(GGML_USE_KOMPUTE)
+    buft = ggml_backend_kompute_buffer_type(gpu);
+    if (buft == nullptr) {
+        LLAMA_LOG_WARN("%s: cannot use GPU %d, check `vulkaninfo --summary`\n", __func__, gpu);
+    }
+#endif
+
+    if (buft == nullptr) {
+        buft = llama_default_buffer_type_cpu(true);
+    }
+    return buft;
+    GGML_UNUSED(model);
+    GGML_UNUSED(gpu);
+}
+
+static ggml_backend_buffer_type_t llama_default_buffer_type_split(const llama_model & model, int fallback_gpu, const float * tensor_split) {
+    ggml_backend_buffer_type_t buft = nullptr;
+
+#ifdef GGML_USE_CUDA
+    if (ggml_backend_cuda_get_device_count() > 1) {
+        buft = ggml_backend_cuda_split_buffer_type(tensor_split);
+    }
+#endif
+
+#ifdef GGML_USE_SYCL
+    if (ggml_backend_sycl_get_device_count() > 1) {
+        buft = ggml_backend_sycl_split_buffer_type(tensor_split);
+    }
+#endif
+
+    if (buft == nullptr) {
+        buft = llama_default_buffer_type_offload(model, fallback_gpu);
+    }
+    return buft;
+
+    GGML_UNUSED(tensor_split);
+}
+
+static size_t llama_get_device_memory(const llama_model & model, int device) {
+#if defined(GGML_USE_RPC)
+    int dev_count = (int)llama_get_device_count(model);
+    int rpc_count = (int)model.rpc_servers.size();
+    if (device >= dev_count - rpc_count) {
+        size_t total;
+        size_t free;
+        const char * endpoint = model.rpc_servers[device - dev_count + rpc_count].c_str();
+        ggml_backend_rpc_get_device_memory(endpoint, &free, &total);
+        return free;
+    }
+#endif
+#if defined(GGML_USE_CUDA)
+    size_t total;
+    size_t free;
+    ggml_backend_cuda_get_device_memory(device, &free, &total);
+    return free;
+#elif defined(GGML_USE_SYCL)
+    size_t total;
+    size_t free;
+    ggml_backend_sycl_get_device_memory(device, &free, &total);
+    return free;
+#elif defined(GGML_USE_VULKAN)
+    size_t total;
+    size_t free;
+    ggml_backend_vk_get_device_memory(device, &free, &total);
+    return free;
+#else
+    return 1;
+#endif
+    GGML_UNUSED(model);
+    GGML_UNUSED(device);
+}
 
 //
 // kv cache helpers
@@ -2403,10 +2535,6 @@ static bool llama_kv_cache_init(
         }
     }
 
-#ifdef GGML_USE_CLBLAST
-    offload = false;
-#endif
-
     // count used buffer types
     std::map<ggml_backend_buffer_type_t, int> buft_layer_count;
     if (offload) {
@@ -2472,7 +2600,6 @@ static bool llama_kv_cache_init(
 static bool llama_kv_cache_find_slot(
            struct llama_kv_cache & cache,
         const struct llama_batch & batch) {
-    const uint32_t n_ctx    = cache.size;
     const uint32_t n_tokens = batch.n_tokens;
 
     if (cache.recurrent) {
@@ -2523,16 +2650,16 @@ static bool llama_kv_cache_find_slot(
     }
     // otherwise, one cell per token.
 
-    if (n_tokens > n_ctx) {
-        LLAMA_LOG_ERROR("%s: n_tokens=%d > n_ctx=%d\n", __func__, n_tokens, n_ctx);
+    if (n_tokens > cache.size) {
+        LLAMA_LOG_ERROR("%s: n_tokens=%d > cache.size=%d\n", __func__, n_tokens, cache.size);
         return false;
     }
 
     uint32_t n_tested = 0;
 
     while (true) {
-        if (cache.head + n_tokens > n_ctx) {
-            n_tested += n_ctx - cache.head;
+        if (cache.head + n_tokens > cache.size) {
+            n_tested += cache.size - cache.head;
             cache.head = 0;
             continue;
         }
@@ -2551,7 +2678,7 @@ static bool llama_kv_cache_find_slot(
             break;
         }
 
-        if (n_tested >= n_ctx) {
+        if (n_tested >= cache.size) {
             //LLAMA_LOG_ERROR("%s: failed to find a slot for %d tokens\n", __func__, n_tokens);
             return false;
         }
@@ -2805,6 +2932,11 @@ static void llama_kv_cache_defrag(struct llama_kv_cache & cache) {
     cache.do_defrag = true;
 }
 
+static uint32_t llama_kv_cache_get_padding(const struct llama_cparams & cparams) {
+    // the FA kernels require padding to avoid extra runtime boundary checks
+    return cparams.flash_attn ? 256u : 32u;
+}
+
 //
 // model loading and saving
 //
@@ -3307,20 +3439,53 @@ struct llama_model_loader {
     }
 
     template<typename T>
-    bool get_key(const std::string & key, T & result, const bool required = true) {
-        auto it = kv_overrides.find(key);
+    bool get_arr(const std::string & key, std::vector<T> & result, const bool required = true) {
+        const int kid = gguf_find_key(meta, key.c_str());
 
-        const struct llama_model_kv_override * override =
-            it != kv_overrides.end() ? &it->second : nullptr;
+        if (kid < 0) {
+            if (required) {
+                throw std::runtime_error(format("key not found in model: %s", key.c_str()));
+            }
+            return false;
+        }
 
-        const bool found = GGUFMeta::GKV<T>::set(meta, key, result, override);
+        struct GGUFMeta::ArrayInfo arr_info =
+            GGUFMeta::GKV<GGUFMeta::ArrayInfo>::get_kv(meta, kid);
 
-        if (required && !found) {
-            throw std::runtime_error(format("key not found in model: %s", key.c_str()));
+        if (arr_info.gt != GGUF_TYPE_FLOAT32 && arr_info.gt != GGUF_TYPE_INT32) {
+            throw std::runtime_error(format("%s is not a float32 or int32 array", key.c_str()));
         }
 
-        return found;
-    }
+        // GGML_ASSERT(gguf_type_size(arr_info.gt) == sizeof(T));
+        GGML_ASSERT((arr_info.gt != GGUF_TYPE_FLOAT32 || std::is_same<T, float>::value));
+        GGML_ASSERT((arr_info.gt != GGUF_TYPE_INT32   || std::is_same<T, int>::value));
+
+        result.resize(arr_info.length);
+        result.assign((const T*)arr_info.data, (const T *)arr_info.data + arr_info.length);
+
+        return true;
+    }
+
+    template<typename T>
+    bool get_arr(const enum llm_kv kid, T& result, const bool required = true) {
+        return get_arr(llm_kv(kid), result, required);
+    }
+
+    template<typename T>
+    bool get_key(const std::string & key, T & result, const bool required = true) {
+        auto it = kv_overrides.find(key);
+
+        const struct llama_model_kv_override * override =
+            it != kv_overrides.end() ? &it->second : nullptr;
+
+        const bool found = GGUFMeta::GKV<T>::set(meta, key, result, override);
+
+        if (required && !found) {
+            throw std::runtime_error(format("key not found in model: %s", key.c_str()));
+        }
+
+        return found;
+    }
 
     template<typename T>
     bool get_key(const enum llm_kv kid, T & result, const bool required = true) {
@@ -3380,11 +3545,15 @@ struct llama_model_loader {
         return get_tensor_meta(get_tensor_name(i));
     }
 
-    struct ggml_tensor * create_tensor_for(struct ggml_context * ctx, const struct ggml_tensor * cur) {
+    struct ggml_tensor * create_tensor_for(struct ggml_context * ctx, const struct ggml_tensor * cur, bool duplicated) {
         struct ggml_tensor * tensor = ggml_dup_tensor(ctx, cur);
         ggml_set_name(tensor, ggml_get_name(cur));
 
-        n_created++;
+        if (duplicated) {
+            size_data += ggml_nbytes(cur);
+        } else {
+            n_created++;
+        }
 
         return tensor;
     }
@@ -3419,14 +3588,17 @@ struct llama_model_loader {
         return cur;
     }
 
-    struct ggml_tensor * create_tensor(struct ggml_context * ctx, const std::string & name, const std::vector<int64_t> & ne, bool required = true) {
-        const struct ggml_tensor * cur = check_tensor_dims(name, ne, required);
+    static const int TENSOR_NOT_REQUIRED = 1;
+    static const int TENSOR_DUPLICATED   = 2;
+
+    struct ggml_tensor * create_tensor(struct ggml_context * ctx, const std::string & name, const std::vector<int64_t> & ne, int flags = 0) {
+        const struct ggml_tensor * cur = check_tensor_dims(name, ne, !(flags & TENSOR_NOT_REQUIRED));
 
         if (cur == NULL) {
             return NULL;
         }
 
-        return create_tensor_for(ctx, cur);
+        return create_tensor_for(ctx, cur, flags & TENSOR_DUPLICATED);
     }
 
     struct ggml_tensor * create_tensor_as_view(struct ggml_context * ctx, struct ggml_tensor * base, const std::string & name, const std::vector<int64_t> & ne, size_t offset, bool required = true) {
@@ -3726,37 +3898,50 @@ static std::string llama_model_ftype_name(llama_ftype ftype) {
 
 static const char * llama_model_type_name(e_model type) {
     switch (type) {
-        case MODEL_22M:    return "22M";
-        case MODEL_33M:    return "33M";
-        case MODEL_109M:   return "109M";
-        case MODEL_137M:   return "137M";
-        case MODEL_0_5B:   return "0.5B";
-        case MODEL_1B:     return "1B";
-        case MODEL_2B:     return "2B";
-        case MODEL_3B:     return "3B";
-        case MODEL_7B:     return "7B";
-        case MODEL_8B:     return "8B";
-        case MODEL_12B:    return "12B";
-        case MODEL_13B:    return "13B";
-        case MODEL_14B:    return "14B";
-        case MODEL_15B:    return "15B";
-        case MODEL_20B:    return "20B";
-        case MODEL_30B:    return "30B";
-        case MODEL_34B:    return "34B";
-        case MODEL_35B:    return "35B";
-        case MODEL_40B:    return "40B";
-        case MODEL_65B:    return "65B";
-        case MODEL_70B:    return "70B";
-        case MODEL_314B:   return "314B";
-        case MODEL_SMALL:  return "0.1B";
-        case MODEL_MEDIUM: return "0.4B";
-        case MODEL_LARGE:  return "0.8B";
-        case MODEL_XL:     return "1.5B";
-        case MODEL_A2_7B:  return "A2.7B";
-        case MODEL_8x7B:   return "8x7B";
-        case MODEL_8x22B:  return "8x22B";
-        case MODEL_16x12B: return "16x12B";
-        default:           return "?B";
+        case MODEL_14M:           return "14M";
+        case MODEL_17M:           return "17M";
+        case MODEL_22M:           return "22M";
+        case MODEL_33M:           return "33M";
+        case MODEL_70M:           return "70M";
+        case MODEL_109M:          return "109M";
+        case MODEL_137M:          return "137M";
+        case MODEL_160M:          return "160M";
+        case MODEL_335M:          return "335M";
+        case MODEL_410M:          return "410M";
+        case MODEL_0_5B:          return "0.5B";
+        case MODEL_1B:            return "1B";
+        case MODEL_1_4B:          return "1.4B";
+        case MODEL_2B:            return "2B";
+        case MODEL_2_8B:          return "2.8B";
+        case MODEL_3B:            return "3B";
+        case MODEL_4B:            return "4B";
+        case MODEL_6_9B:          return "6.9B";
+        case MODEL_7B:            return "7B";
+        case MODEL_8B:            return "8B";
+        case MODEL_12B:           return "12B";
+        case MODEL_13B:           return "13B";
+        case MODEL_14B:           return "14B";
+        case MODEL_15B:           return "15B";
+        case MODEL_16B:           return "16B";
+        case MODEL_20B:           return "20B";
+        case MODEL_30B:           return "30B";
+        case MODEL_34B:           return "34B";
+        case MODEL_35B:           return "35B";
+        case MODEL_40B:           return "40B";
+        case MODEL_65B:           return "65B";
+        case MODEL_70B:           return "70B";
+        case MODEL_236B:          return "236B";
+        case MODEL_314B:          return "314B";
+        case MODEL_SMALL:         return "0.1B";
+        case MODEL_MEDIUM:        return "0.4B";
+        case MODEL_LARGE:         return "0.8B";
+        case MODEL_XL:            return "1.5B";
+        case MODEL_A2_7B:         return "A2.7B";
+        case MODEL_8x7B:          return "8x7B";
+        case MODEL_8x22B:         return "8x22B";
+        case MODEL_16x12B:        return "16x12B";
+        case MODEL_10B_128x3_66B: return "10B+128x3.66B";
+        default:                  return "?B";
     }
 }
 
@@ -3829,8 +4014,8 @@ static void llm_load_hparams(
     ml.get_key(LLM_KV_ROPE_SCALING_FINETUNED, rope_finetuned, false);
     hparams.rope_finetuned = rope_finetuned;
 
-    hparams.n_yarn_orig_ctx = hparams.n_ctx_train;
-    ml.get_key(LLM_KV_ROPE_SCALING_ORIG_CTX_LEN, hparams.n_yarn_orig_ctx, false);
+    hparams.n_ctx_orig_yarn = hparams.n_ctx_train;
+    ml.get_key(LLM_KV_ROPE_SCALING_ORIG_CTX_LEN, hparams.n_ctx_orig_yarn, false);
 
     // rope_freq_base (optional)
     hparams.rope_freq_base_train = 10000.0f;
@@ -3849,6 +4034,8 @@ static void llm_load_hparams(
     }
     hparams.rope_freq_scale_train = ropescale == 0.0f ? 1.0f : 1.0f/ropescale;
 
+    ml.get_key(LLM_KV_ROPE_SCALING_ATTN_FACTOR, hparams.rope_attn_factor, false);
+
     // sanity check for n_rot (optional)
     {
         hparams.n_rot = (hparams.n_head == 0) ? 0 : hparams.n_embd / hparams.n_head;
@@ -3886,7 +4073,9 @@ static void llm_load_hparams(
                     switch (hparams.n_layer) {
                         case 22: model.type = e_model::MODEL_1B; break;
                         case 26: model.type = e_model::MODEL_3B; break;
-                        case 32: model.type = hparams.n_vocab < 40000 ? e_model::MODEL_7B : e_model::MODEL_8B; break;
+                        // granite uses a vocab with len 49152
+                        case 32: model.type = hparams.n_vocab == 49152 ? e_model::MODEL_3B : (hparams.n_vocab < 40000 ? e_model::MODEL_7B : e_model::MODEL_8B); break;
+                        case 36: model.type = e_model::MODEL_8B; break; // granite
                         case 40: model.type = e_model::MODEL_13B; break;
                         case 48: model.type = e_model::MODEL_34B; break;
                         case 60: model.type = e_model::MODEL_30B; break;
@@ -3948,14 +4137,6 @@ static void llm_load_hparams(
                     default: model.type = e_model::MODEL_UNKNOWN;
                 }
             } break;
-        case LLM_ARCH_PERSIMMON:
-            {
-                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps);
-                switch (hparams.n_layer) {
-                    case 36: model.type = e_model::MODEL_8B; break;
-                    default: model.type = e_model::MODEL_UNKNOWN;
-                }
-            } break;
         case LLM_ARCH_REFACT:
             {
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
@@ -4097,6 +4278,7 @@ static void llm_load_hparams(
                 switch (hparams.n_layer) {
                     case 24: model.type = e_model::MODEL_1B; break;
                     case 32: model.type = e_model::MODEL_3B; break;
+                    case 40: model.type = e_model::MODEL_14B; break;
                     default: model.type = e_model::MODEL_UNKNOWN;
                 }
             } break;
@@ -4163,6 +4345,8 @@ static void llm_load_hparams(
                     case 30: model.type = e_model::MODEL_3B; break;
                     case 32: model.type = e_model::MODEL_7B; break;
                     case 40: model.type = e_model::MODEL_15B; break;
+                    case 52: model.type = e_model::MODEL_20B; break; // granite
+                    case 88: model.type = e_model::MODEL_34B; break; // granite
                     default: model.type = e_model::MODEL_UNKNOWN;
                 }
             } break;
@@ -4237,6 +4421,85 @@ static void llm_load_hparams(
                     default: model.type = e_model::MODEL_UNKNOWN;
                 }
             } break;
+        case LLM_ARCH_GPTNEOX:
+            {
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps);
+                ml.get_key(LLM_KV_USE_PARALLEL_RESIDUAL, hparams.use_par_res);
+                switch (hparams.n_layer) {
+                    case 6:
+                        switch (hparams.n_ff) {
+                            case 512: model.type = e_model::MODEL_14M; break;
+                            case 2048: model.type = e_model::MODEL_70M; break;
+                            default: model.type = e_model::MODEL_UNKNOWN;
+                        } break;
+                    case 12:
+                        switch (hparams.n_ff) {
+                            case 3072: model.type = e_model::MODEL_160M; break;
+                            default: model.type = e_model::MODEL_UNKNOWN;
+                        } break;
+                    case 16:
+                        switch (hparams.n_ff) {
+                            case 8192: model.type = e_model::MODEL_1B; break;
+                            default: model.type = e_model::MODEL_UNKNOWN;
+                        } break;
+                    case 24:
+                        switch (hparams.n_ff) {
+                            case 4096: model.type = e_model::MODEL_410M; break;
+                            case 8192: model.type = e_model::MODEL_1_4B; break;
+                            default: model.type = e_model::MODEL_UNKNOWN;
+                        } break;
+                    case 32:
+                        switch (hparams.n_ff) {
+                            case 10240: model.type = e_model::MODEL_2_8B; break;
+                            case 16384: model.type = e_model::MODEL_6_9B; break;
+                            default: model.type = e_model::MODEL_UNKNOWN;
+                        } break;
+                    case 36:
+                        switch (hparams.n_ff) {
+                            case 20480: model.type = e_model::MODEL_12B; break;
+                            default: model.type = e_model::MODEL_UNKNOWN;
+                        } break;
+                    case 44:
+                        switch (hparams.n_ff) {
+                            case 24576: model.type = e_model::MODEL_20B; break;
+                            default: model.type = e_model::MODEL_UNKNOWN;
+                        } break;
+                    default: model.type = e_model::MODEL_UNKNOWN;
+                }
+            } break;
+        case LLM_ARCH_ARCTIC:
+            {
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+
+                if (hparams.n_expert == 128) {
+                    switch (hparams.n_layer) {
+                        case 35: model.type = e_model::MODEL_10B_128x3_66B; break;
+                        default: model.type = e_model::MODEL_UNKNOWN;
+                    }
+                } else {
+                    model.type = e_model::MODEL_UNKNOWN;
+                }
+            } break;
+        case LLM_ARCH_DEEPSEEK2:
+            {
+                bool is_lite = (hparams.n_layer == 27);
+                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);
+                if (!is_lite) {
+                    ml.get_key(LLM_KV_ATTENTION_Q_LORA_RANK, hparams.n_lora_q);
+                }
+                ml.get_key(LLM_KV_ATTENTION_KV_LORA_RANK, hparams.n_lora_kv);
+                ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp);
+                ml.get_key(LLM_KV_EXPERT_SHARED_COUNT, hparams.n_expert_shared);
+                ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE, hparams.expert_weights_scale);
+                ml.get_key(LLM_KV_ROPE_SCALING_YARN_LOG_MUL, hparams.rope_yarn_log_mul);
+
+                switch (hparams.n_layer) {
+                    case 27: model.type = e_model::MODEL_16B; break;
+                    case 60: model.type = e_model::MODEL_236B; break;
+                    default: model.type = e_model::MODEL_UNKNOWN;
+                }
+            } break;
         default: (void)0;
     }
 
@@ -4298,35 +4561,6 @@ static void llm_load_vocab(
             vocab.special_cls_id  = -1;
             vocab.special_mask_id = -1;
 
-            // For Fill-In-the-Middle (FIM)/infill models which where converted
-            // prior to support of FIM special tokens in GGUF, the following
-            // will allow those models to continue to work. The general names
-            // of the known models are currently CodeLlama (LLM_ARCH_LLAMA) and
-            // CodeGemma (LLM_ARCH_GEMMA). This can potentially be removed once
-            // new versions of these models have been published.
-            std::string gen_name;
-            ml.get_key(LLM_KV_GENERAL_NAME, gen_name, false);
-
-            std::transform(gen_name.begin(), gen_name.end(), gen_name.begin(),
-                [](unsigned char c){ return std::tolower(c); });
-
-            if (gen_name.find("code") != std::string::npos) {
-                if (model.arch == LLM_ARCH_LLAMA) {
-                    vocab.special_prefix_id = 32007;
-                    vocab.special_suffix_id = 32008;
-                    vocab.special_middle_id = 32009;
-                    vocab.special_eot_id    = 32010;
-                } else if (model.arch == LLM_ARCH_GEMMA) {
-                    vocab.special_prefix_id = 67;
-                    vocab.special_suffix_id = 69;
-                    vocab.special_middle_id = 68;
-                    // TODO: this is not EOT, it is "file separator" token, needs fix
-                    //       https://huggingface.co/google/codegemma-7b-it/blob/9b1d9231388358c04d90bd003458f5070d97db44/tokenizer_config.json#L565-L572
-                    //vocab.special_eot_id    = 70;
-                    vocab.special_eot_id    = 107;
-                }
-            }
-
             const int add_space_prefix_keyidx = gguf_find_key(ctx, kv(LLM_KV_TOKENIZER_ADD_PREFIX).c_str());
             if (add_space_prefix_keyidx != -1) {
                 vocab.add_space_prefix = gguf_get_val_bool(ctx, add_space_prefix_keyidx);
@@ -4343,15 +4577,14 @@ static void llm_load_vocab(
             vocab.special_cls_id  = 101;
             vocab.special_mask_id = 103;
             vocab.add_space_prefix = false;
-        } else {
-            if (tokenizer_model == "gpt2") {
-                vocab.type = LLAMA_VOCAB_TYPE_BPE;
-            } else {
-                LLAMA_LOG_WARN("%s: unknown tokenizer: '%s'", __func__, tokenizer_model.c_str());
-                LLAMA_LOG_WARN("%s: using default tokenizer: 'llama'", __func__);
-                vocab.type = LLAMA_VOCAB_TYPE_SPM;
-                return;
+        } else if (tokenizer_model == "gpt2") {
+            vocab.type = LLAMA_VOCAB_TYPE_BPE;
+
+            const int add_space_prefix_keyidx = gguf_find_key(ctx, kv(LLM_KV_TOKENIZER_ADD_PREFIX).c_str());
+            if (add_space_prefix_keyidx != -1) {
+                vocab.add_space_prefix = gguf_get_val_bool(ctx, add_space_prefix_keyidx);
             }
+
             // read bpe merges and populate bpe ranks
             const int merges_keyidx = gguf_find_key(ctx, kv(LLM_KV_TOKENIZER_MERGES).c_str());
             if (merges_keyidx == -1) {
@@ -4385,6 +4618,8 @@ static void llm_load_vocab(
             vocab.special_pad_id  = -1;
             vocab.special_cls_id  = -1;
             vocab.special_mask_id = -1;
+        } else {
+            throw std::runtime_error(format("unknown tokenizer: '%s'", tokenizer_model.c_str()));
         }
 
         // for now, only BPE models have pre-tokenizers
@@ -4398,8 +4633,7 @@ static void llm_load_vocab(
                 LLAMA_LOG_WARN("%s: ************************************        \n", __func__);
                 LLAMA_LOG_WARN("%s:                                             \n", __func__);
                 vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_DEFAULT;
-            } else if (
-                    tokenizer_pre == "default") {
+            } else if (tokenizer_pre == "default") {
                 vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_DEFAULT;
             } else if (
                     tokenizer_pre == "llama3"   ||
@@ -4424,7 +4658,10 @@ static void llm_load_vocab(
             } else if (
                     tokenizer_pre == "gpt-2"   ||
                     tokenizer_pre == "jina-es" ||
-                    tokenizer_pre == "jina-de") {
+                    tokenizer_pre == "jina-de" ||
+                    tokenizer_pre == "jina-v2-es" ||
+                    tokenizer_pre == "jina-v2-de" ||
+                    tokenizer_pre == "jina-v2-code") {
                 vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_GPT2;
             } else if (
                     tokenizer_pre == "refact") {
@@ -4435,12 +4672,21 @@ static void llm_load_vocab(
             } else if (
                 tokenizer_pre == "qwen2") {
                 vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_QWEN2;
+            } else if (
+                tokenizer_pre == "stablelm2") {
+                vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_STABLELM2;
             } else if (
                 tokenizer_pre == "olmo") {
                 vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_OLMO;
             } else if (
                 tokenizer_pre == "dbrx") {
                 vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_DBRX;
+            } else if (
+                tokenizer_pre == "smaug-bpe") {
+                vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_SMAUG;
+            } else if (
+                tokenizer_pre == "poro-chat") {
+                vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_PORO;
             } else {
                 throw std::runtime_error(format("unknown pre-tokenizer type: '%s'", tokenizer_pre.c_str()));
             }
@@ -4479,12 +4725,64 @@ static void llm_load_vocab(
         auto & token_data = vocab.id_to_token[i];
         token_data.text  = std::move(word);
         token_data.score = scores ? scores[i] : 0.0f;
-        token_data.type  = toktypes ? (llama_token_type) toktypes[i] : LLAMA_TOKEN_TYPE_NORMAL;
+        token_data.attr  = LLAMA_TOKEN_ATTR_NORMAL;
+
+        if (toktypes) {  //TODO: remove, required until per token attributes are available from GGUF file
+            switch(toktypes[i]) {
+                case LLAMA_TOKEN_TYPE_UNKNOWN:      token_data.attr = LLAMA_TOKEN_ATTR_UNKNOWN;      break;
+                case LLAMA_TOKEN_TYPE_UNUSED:       token_data.attr = LLAMA_TOKEN_ATTR_UNUSED;       break;
+                case LLAMA_TOKEN_TYPE_NORMAL:       token_data.attr = LLAMA_TOKEN_ATTR_NORMAL;       break;
+                case LLAMA_TOKEN_TYPE_CONTROL:      token_data.attr = LLAMA_TOKEN_ATTR_CONTROL;      break;
+                case LLAMA_TOKEN_TYPE_USER_DEFINED: token_data.attr = LLAMA_TOKEN_ATTR_USER_DEFINED; break;
+                case LLAMA_TOKEN_TYPE_BYTE:         token_data.attr = LLAMA_TOKEN_ATTR_BYTE;         break;
+                case LLAMA_TOKEN_TYPE_UNDEFINED:    token_data.attr = LLAMA_TOKEN_ATTR_UNDEFINED;    break;
+                default:                            token_data.attr = LLAMA_TOKEN_ATTR_UNDEFINED;    break;
+            }
+        }
     }
     GGML_ASSERT(vocab.id_to_token.size() == vocab.token_to_id.size());
 
     // determine the newline token: LLaMA "<0x0A>" == 10 == '\n', Falcon 193 == '\n'
     if (vocab.type == LLAMA_VOCAB_TYPE_SPM) {
+        // For Fill-In-the-Middle (FIM)/infill models which where converted
+        // prior to support of FIM special tokens in GGUF, the following
+        // will allow those models to continue to work. The general names
+        // of the known models are currently CodeLlama (LLM_ARCH_LLAMA) and
+        // CodeGemma (LLM_ARCH_GEMMA). This can potentially be removed once
+        // new versions of these models have been published.
+        std::string gen_name;
+        ml.get_key(LLM_KV_GENERAL_NAME, gen_name, false);
+
+        std::transform(gen_name.begin(), gen_name.end(), gen_name.begin(),
+            [](unsigned char c){ return std::tolower(c); });
+
+        if (gen_name.find("code") != std::string::npos) {
+            if (model.arch == LLM_ARCH_LLAMA
+              && 32010 < vocab.id_to_token.size()
+              && vocab.id_to_token[32007].text == "<PRE>"
+              && vocab.id_to_token[32008].text == "<SUF>"
+              && vocab.id_to_token[32009].text == "<MID>"
+              && vocab.id_to_token[32010].text == "<EOT>") {
+                vocab.special_prefix_id = 32007;
+                vocab.special_suffix_id = 32008;
+                vocab.special_middle_id = 32009;
+                vocab.special_eot_id    = 32010;
+            } else if (model.arch == LLM_ARCH_GEMMA
+              && 107 < vocab.id_to_token.size()
+              && vocab.id_to_token[67].text == "<|fim_prefix|>"
+              && vocab.id_to_token[69].text == "<|fim_suffix|>"
+              && vocab.id_to_token[68].text == "<|fim_middle|>"
+              && vocab.id_to_token[107].text == "<end_of_turn>") {
+                vocab.special_prefix_id = 67;
+                vocab.special_suffix_id = 69;
+                vocab.special_middle_id = 68;
+                // TODO: this is not EOT, it is "file separator" token, needs fix
+                //       https://huggingface.co/google/codegemma-7b-it/blob/9b1d9231388358c04d90bd003458f5070d97db44/tokenizer_config.json#L565-L572
+                //vocab.special_eot_id    = 70;
+                vocab.special_eot_id    = 107;
+            }
+        }
+
         try {
             vocab.linefeed_id = llama_byte_to_token(vocab, '\n');
         } catch (const std::exception & e) {
@@ -4556,7 +4854,8 @@ static void llm_load_vocab(
                         (t.first == "<|eot_id|>" ||
                          t.first == "<|im_end|>" ||
                          t.first == "<|end|>" ||
-                         t.first == "<end_of_turn>"
+                         t.first == "<end_of_turn>" ||
+                         t.first == "<|endoftext|>"
                         )
                    ) {
                     vocab.special_eot_id = t.second;
@@ -4568,96 +4867,88 @@ static void llm_load_vocab(
 
     // build special tokens cache
     {
-        // TODO: It is unclear (to me) at this point, whether special tokes are guaranteed to be of a deterministic type,
-        //  and will always be correctly labeled in 'added_tokens.json' etc.
-        // The assumption is, since special tokens aren't meant to be exposed to end user, they are designed
-        //  to be unmatchable by the tokenizer, therefore tokens from the vocab, which are unmatchable by the tokenizer
-        //  are special tokens.
-        // From testing, this appears to correlate 1:1 with special tokens.
-        //
+        for (llama_vocab::id id = 0; id < (llama_vocab::id)n_vocab; ++id) {
+            if (!(vocab.id_to_token[id].attr & LLAMA_TOKEN_ATTR_NORMAL)) {
+                vocab.cache_special_tokens.push_back(id);
+            }
+        }
 
-        // Counting special tokens and verifying in only one direction
-        //  is sufficient to detect difference in those two sets.
-        //
-        uint32_t special_tokens_count_by_type = 0;
-        uint32_t special_tokens_count_from_verification = 0;
+        std::sort( vocab.cache_special_tokens.begin(), vocab.cache_special_tokens.end(),
+            [&] (const llama_vocab::id a, const llama_vocab::id b) {
+                return vocab.id_to_token[a].text.size() > vocab.id_to_token[b].text.size();
+            }
+        );
 
-        bool special_tokens_definition_mismatch = false;
+        LLAMA_LOG_INFO("%s: special tokens cache size = %u\n", __func__, (uint32_t)vocab.cache_special_tokens.size());
+    }
 
-        for (const auto & t : vocab.token_to_id) {
-            const auto & token = t.first;
-            const auto & id    = t.second;
+    // build token to piece cache
+    {
+        size_t size_cache = 0;
 
-            // Count all non-normal tokens in the vocab while iterating
-            if (vocab.id_to_token[id].type != LLAMA_TOKEN_TYPE_NORMAL) {
-                special_tokens_count_by_type++;
-            }
+        std::vector<llama_vocab::token> cache_token_to_piece(n_vocab);
 
-            // Skip single character tokens
-            if (token.length() > 1) {
-                bool is_tokenizable = false;
+        for (uint32_t id = 0; id < n_vocab; ++id) {
+            cache_token_to_piece[id] = llama_token_to_piece(&model, id, true);
 
-                // Split token string representation in two, in all possible ways
-                //  and check if both halves can be matched to a valid token
-                for (unsigned i = 1; i < token.length();) {
-                    const auto left  = token.substr(0, i);
-                    const auto right = token.substr(i);
+            size_cache += cache_token_to_piece[id].size();
+        }
 
-                    // check if we didnt partition in the middle of a utf sequence
-                    auto utf = utf8_len(left.at(left.length() - 1));
+        std::swap(vocab.cache_token_to_piece, cache_token_to_piece);
 
-                    if (utf == 1) {
-                        if (vocab.token_to_id.find(left)  != vocab.token_to_id.end() &&
-                            vocab.token_to_id.find(right) != vocab.token_to_id.end() ) {
-                            is_tokenizable = true;
-                            break;
-                        }
-                        i++;
-                    } else {
-                        // skip over the rest of multibyte utf sequence
-                        i += utf - 1;
-                    }
+        LLAMA_LOG_INFO("%s: token to piece cache size = %.4f MB\n", __func__, size_cache / 1024.0 / 1024.0);
+    }
+
+    // Handle per token attributes
+    //NOTE: Each model customizes per token attributes.
+    //NOTE: Per token attributes are missing from the GGUF file.
+    //TODO: Extract attributes from GGUF file.
+    {
+        auto _contains_any = [] (const std::string &str, const std::vector<std::string> &substrs) -> bool {
+            for (auto substr : substrs) {
+                if (str.find(substr) < std::string::npos) {
+                    return true;
                 }
+            }
+            return false;
+        };
 
-                if (!is_tokenizable) {
-                    // Some tokens are multibyte, but they are utf sequences with equivalent text length of 1
-                    //  it's faster to re-filter them here, since there are way less candidates now
+        auto _set_tokenid_attr = [&] (const llama_vocab::id id, llama_token_attr attr, bool value) {
+            uint32_t current = vocab.id_to_token.at(id).attr;
+            current = value ? (current | attr) : (current & ~attr);
+            vocab.id_to_token[id].attr = (llama_token_attr) current;
+        };
 
-                    // Calculate a total "utf" length of a token string representation
-                    size_t utf8_str_len = 0;
-                    for (unsigned i = 0; i < token.length();) {
-                        utf8_str_len++;
-                        i += utf8_len(token.at(i));
-                    }
+        auto _set_token_attr = [&] (const std::string & token, llama_token_attr attr, bool value) {
+            _set_tokenid_attr(vocab.token_to_id.at(token), attr, value);
+        };
 
-                    // And skip the ones which are one character
-                    if (utf8_str_len > 1) {
-                        // At this point what we have left are special tokens only
-                        vocab.special_tokens_cache[token] = id;
+        std::string model_name;
+        std::string tokenizer_pre;
 
-                        // Count manually found special tokens
-                        special_tokens_count_from_verification++;
+        ml.get_key(LLM_KV_GENERAL_NAME, model_name, false);
+        ml.get_key(LLM_KV_TOKENIZER_PRE, tokenizer_pre, false);
 
-                        // If this manually found special token is not marked as such, flag a mismatch
-                        if (vocab.id_to_token[id].type == LLAMA_TOKEN_TYPE_NORMAL) {
-                            special_tokens_definition_mismatch = true;
-                        }
-                    }
-                }
+        // model name to lowercase
+        std::transform(model_name.begin(), model_name.end(), model_name.begin(),
+            [] (const std::string::value_type x) {
+                return std::tolower(x);
             }
-        }
+        );
 
-        if (special_tokens_definition_mismatch || special_tokens_count_from_verification != special_tokens_count_by_type) {
-            LLAMA_LOG_WARN("%s: mismatch in special tokens definition ( %u/%zu vs %u/%zu ).\n",
-                __func__,
-                special_tokens_count_from_verification, vocab.id_to_token.size(),
-                special_tokens_count_by_type, vocab.id_to_token.size()
-            );
-        } else {
-            LLAMA_LOG_INFO("%s: special tokens definition check successful ( %u/%zu ).\n",
-                __func__,
-                special_tokens_count_from_verification, vocab.id_to_token.size()
-            );
+        // set attributes by model/tokenizer name
+        if (_contains_any(tokenizer_pre, {"jina-v2-es", "jina-v2-de"})) {
+            _set_token_attr("<mask>", LLAMA_TOKEN_ATTR_LSTRIP, true);
+        } else if (_contains_any(model_name, {"phi-3", "phi3"})) {
+            for (auto id : vocab.cache_special_tokens) {
+                _set_tokenid_attr(id, LLAMA_TOKEN_ATTR_RSTRIP, true);
+            }
+            for (auto token : {"</s>"}) {
+                _set_token_attr(token, LLAMA_TOKEN_ATTR_RSTRIP, true);
+            }
+            for (auto token : {"<unk>", "<s>", "<|endoftext|>"}) {
+                _set_token_attr(token, LLAMA_TOKEN_ATTR_RSTRIP, false);
+            }
         }
     }
 }
@@ -4699,7 +4990,7 @@ static void llm_load_print_meta(llama_model_loader & ml, llama_model & model) {
     LLAMA_LOG_INFO("%s: rope scaling     = %s\n",     __func__, rope_scaling_type);
     LLAMA_LOG_INFO("%s: freq_base_train  = %.1f\n",   __func__, hparams.rope_freq_base_train);
     LLAMA_LOG_INFO("%s: freq_scale_train = %g\n",     __func__, hparams.rope_freq_scale_train);
-    LLAMA_LOG_INFO("%s: n_yarn_orig_ctx  = %u\n",     __func__, hparams.n_yarn_orig_ctx);
+    LLAMA_LOG_INFO("%s: n_ctx_orig_yarn  = %u\n",     __func__, hparams.n_ctx_orig_yarn);
     LLAMA_LOG_INFO("%s: rope_finetuned   = %s\n",     __func__, hparams.rope_finetuned ? "yes" : "unknown");
     LLAMA_LOG_INFO("%s: ssm_d_conv       = %u\n",     __func__, hparams.ssm_d_conv);
     LLAMA_LOG_INFO("%s: ssm_d_inner      = %u\n",     __func__, hparams.ssm_d_inner);
@@ -4739,6 +5030,16 @@ static void llm_load_print_meta(llama_model_loader & ml, llama_model & model) {
     if (vocab.special_suffix_id != -1) { LLAMA_LOG_INFO( "%s: SUF token        = %d '%s'\n", __func__, vocab.special_suffix_id, vocab.id_to_token[vocab.special_suffix_id].text.c_str() ); }
     if (vocab.special_middle_id != -1) { LLAMA_LOG_INFO( "%s: MID token        = %d '%s'\n", __func__, vocab.special_middle_id, vocab.id_to_token[vocab.special_middle_id].text.c_str() ); }
     if (vocab.special_eot_id    != -1) { LLAMA_LOG_INFO( "%s: EOT token        = %d '%s'\n", __func__, vocab.special_eot_id,    vocab.id_to_token[vocab.special_eot_id].text.c_str() );    }
+
+    if (model.arch == LLM_ARCH_DEEPSEEK2) {
+        LLAMA_LOG_INFO("%s: n_layer_dense_lead   = %d\n",     __func__, hparams.n_layer_dense_lead);
+        LLAMA_LOG_INFO("%s: n_lora_q             = %d\n",     __func__, hparams.n_lora_q);
+        LLAMA_LOG_INFO("%s: n_lora_kv            = %d\n",     __func__, hparams.n_lora_kv);
+        LLAMA_LOG_INFO("%s: n_ff_exp             = %d\n",     __func__, hparams.n_ff_exp);
+        LLAMA_LOG_INFO("%s: n_expert_shared      = %d\n",     __func__, hparams.n_expert_shared);
+        LLAMA_LOG_INFO("%s: expert_weights_scale = %.1f\n",   __func__, hparams.expert_weights_scale);
+        LLAMA_LOG_INFO("%s: rope_yarn_log_mul    = %.4f\n",   __func__, hparams.rope_yarn_log_mul);
+    }
 }
 
 // Returns false if cancelled by progress_callback
@@ -4784,13 +5085,13 @@ static bool llm_load_tensors(
 
     if (split_mode == LLAMA_SPLIT_MODE_LAYER) {
         // calculate the split points
-        int device_count = llama_get_device_count();
+        int device_count = llama_get_device_count(model);
         bool all_zero = tensor_split == nullptr || std::all_of(tensor_split, tensor_split + device_count, [](float x) { return x == 0.0f; });
         std::vector<float> splits(device_count);
         if (all_zero) {
             // default split, by free memory
             for (int i = 0; i < device_count; ++i) {
-                splits[i] = llama_get_device_memory(i);
+                splits[i] = llama_get_device_memory(model, i);
             }
         } else {
             std::copy(tensor_split, tensor_split + device_count, splits.begin());
@@ -4810,35 +5111,35 @@ static bool llm_load_tensors(
         int act_gpu_layers = std::min(n_gpu_layers, (int)n_layer + 1);
         for (int64_t i = i_gpu_start; i < n_layer; ++i) {
             int layer_gpu = std::upper_bound(splits.begin(), splits.begin() + device_count, float(i - i_gpu_start)/act_gpu_layers) - splits.begin();
-            model.buft_layer[i] = llama_default_buffer_type_offload(layer_gpu);
+            model.buft_layer[i] = llama_default_buffer_type_offload(model, layer_gpu);
         }
         // assign the output layer
         if (n_gpu_layers > n_layer) {
             int layer_gpu = std::upper_bound(splits.begin(), splits.begin() + device_count, float(act_gpu_layers - 1)/act_gpu_layers) - splits.begin();
-            model.buft_output = llama_default_buffer_type_offload(layer_gpu);
+            model.buft_output = llama_default_buffer_type_offload(model, layer_gpu);
         } else {
             model.buft_output = llama_default_buffer_type_cpu(true);
         }
     } else {
         ggml_backend_buffer_type_t split_buft;
         if (split_mode == LLAMA_SPLIT_MODE_ROW) {
-            split_buft = llama_default_buffer_type_split(main_gpu, tensor_split);
+            split_buft = llama_default_buffer_type_split(model, main_gpu, tensor_split);
         } else {
             // LLAMA_SPLIT_MODE_NONE or LLAMA_SPLIT_MODE_LAYER in backends where it is not supported
-            split_buft = llama_default_buffer_type_offload(main_gpu);
+            split_buft = llama_default_buffer_type_offload(model, main_gpu);
         }
         // assign the repeating layers
         for (int64_t i = i_gpu_start; i < n_layer; ++i) {
             model.buft_layer[i] = {
                 split_buft,
-                llama_default_buffer_type_offload(main_gpu)
+                llama_default_buffer_type_offload(model, main_gpu)
             };
         }
         // assign the output layer
         if (n_gpu_layers > n_layer) {
             model.buft_output = {
                 split_buft,
-                llama_default_buffer_type_offload(main_gpu)
+                llama_default_buffer_type_offload(model, main_gpu)
             };
         } else {
             model.buft_output = llama_default_buffer_type_cpu(true);
@@ -4882,6 +5183,7 @@ static bool llm_load_tensors(
     // create tensors for the weights
     {
         const int64_t n_embd       = hparams.n_embd;
+        const int64_t n_embd_head  = n_embd / hparams.n_head;
         const int64_t n_embd_k_gqa = hparams.n_embd_k_gqa();
         const int64_t n_embd_v_gqa = hparams.n_embd_v_gqa();
         const int64_t n_embd_gqa   = n_embd_v_gqa;
@@ -4894,8 +5196,6 @@ static bool llm_load_tensors(
             throw std::runtime_error("model has expert layers but no expert layers are used");
         }
 
-        GGML_ASSERT(n_embd_gqa == n_embd_k_gqa);
-
         ggml_context * ctx_input        = ctx_map.at(model.buft_input.buft);
         ggml_context * ctx_output       = ctx_map.at(model.buft_output.buft);
         ggml_context * ctx_output_split = ctx_map.at(model.buft_output.buft_matrix);
@@ -4915,14 +5215,10 @@ static bool llm_load_tensors(
                     // output
                     {
                         model.output_norm = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
-                        if (model.arch != LLM_ARCH_MINICPM){
-                            model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, false);
-                            // if output is NULL, init from the input tok embed
-                            if (model.output == NULL) {
-                                model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
-                                ml.n_created--; // artificial tensor
-                                ml.size_data += ggml_nbytes(model.output);
-                            }
+                        model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        // if output is NULL, init from the input tok embed
+                        if (model.output == NULL) {
+                            model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED);
                         }
                     }
 
@@ -4940,10 +5236,10 @@ static bool llm_load_tensors(
                         layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd});
 
                         // optional bias tensors
-                        layer.bq = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q,   "bias", i), {n_embd},     false);
-                        layer.bk = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K,   "bias", i), {n_embd_gqa}, false);
-                        layer.bv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_V,   "bias", i), {n_embd_gqa}, false);
-                        layer.bo = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd},     false);
+                        layer.bq = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q,   "bias", i), {n_embd},     llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.bk = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K,   "bias", i), {n_embd_gqa}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.bv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_V,   "bias", i), {n_embd_gqa}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.bo = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd},     llama_model_loader::TENSOR_NOT_REQUIRED);
 
                         layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
 
@@ -4951,10 +5247,15 @@ static bool llm_load_tensors(
                             layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff});
                             layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd});
                             layer.ffn_up   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff});
+
+                            // optional MLP bias
+                            layer.ffn_gate_b = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "bias", i), {n_ff}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                            layer.ffn_down_b = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                            layer.ffn_up_b   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP,   "bias", i), {n_ff}, llama_model_loader::TENSOR_NOT_REQUIRED);
                         } else {
                             layer.ffn_gate_inp = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert});
 
-                            layer.ffn_gate_exps = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd,   n_ff, n_expert}, false);
+                            layer.ffn_gate_exps = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd,   n_ff, n_expert}, llama_model_loader::TENSOR_NOT_REQUIRED);
                             if (layer.ffn_gate_exps) {
                                 layer.ffn_down_exps = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {  n_ff, n_embd, n_expert});
                                 layer.ffn_up_exps   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {n_embd,   n_ff, n_expert});
@@ -4996,12 +5297,10 @@ static bool llm_load_tensors(
                     // output
                     {
                         model.output_norm = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
-                        model.output      = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, false);
+                        model.output      = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_NOT_REQUIRED);
                         // if output is NULL, init from the input tok embed
                         if (model.output == NULL) {
-                            model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
-                            ml.n_created--; // artificial tensor
-                            ml.size_data += ggml_nbytes(model.output);
+                            model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED);
                         }
                     }
 
@@ -5024,7 +5323,7 @@ static bool llm_load_tensors(
 
                         layer.ffn_gate_inp = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert});
 
-                        layer.ffn_gate_exps = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd, n_ff, n_expert}, false);
+                        layer.ffn_gate_exps = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd, n_ff, n_expert}, llama_model_loader::TENSOR_NOT_REQUIRED);
                         if (layer.ffn_gate_exps) {
                             layer.ffn_down_exps = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {  n_ff, n_embd, n_expert});
                             layer.ffn_up_exps   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {n_embd,   n_ff, n_expert});
@@ -5126,11 +5425,9 @@ static bool llm_load_tensors(
                         model.output_norm   = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
                         model.output_norm_b = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "bias"),   {n_embd});
 
-                        model.output        = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, false);
+                        model.output        = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_NOT_REQUIRED);
                         if (!model.output) {
-                            model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}); // needs to be on GPU
-                            ml.n_created--; // artificial tensor
-                            ml.size_data += ggml_nbytes(model.output);
+                            model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED); // needs to be on GPU
                         }
                     }
 
@@ -5143,8 +5440,8 @@ static bool llm_load_tensors(
                         layer.attn_norm   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
                         layer.attn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "bias", i),   {n_embd});
 
-                        layer.attn_norm_2   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM_2, "weight", i), {n_embd}, false);
-                        layer.attn_norm_2_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM_2, "bias", i),   {n_embd}, false);
+                        layer.attn_norm_2   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM_2, "weight", i), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.attn_norm_2_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM_2, "bias", i),   {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
 
                         layer.wqkv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa});
                         layer.wo   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd});
@@ -5162,7 +5459,12 @@ static bool llm_load_tensors(
                     {
                         model.output_norm   = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
                         model.output_norm_b = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "bias"),   {n_embd});
-                        model.output        = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab});
+                        model.output        = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        if (!model.output) {
+                            // needs to be on GPU
+                            model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED);
+                        }
+
                     }
 
                     for (int i = 0; i < n_layer; ++i) {
@@ -5190,47 +5492,6 @@ static bool llm_load_tensors(
                         layer.ffn_up_b   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_UP, "bias", i),     {n_ff});
                     }
                 } break;
-            case LLM_ARCH_PERSIMMON:
-                {
-                    model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"),  {n_embd, n_vocab});
-
-                    {
-                        model.output_norm    = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
-                        model.output_norm_b  = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "bias"),   {n_embd});
-                        model.output         = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab});
-                    }
-
-                    for (int i = 0; i < n_layer; ++i) {
-                        ggml_context * ctx_layer = ctx_for_layer(i);
-                        ggml_context * ctx_split = ctx_for_layer_split(i);
-
-                        auto & layer = model.layers[i];
-
-                        layer.attn_norm     = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM,   "weight", i), {n_embd});
-                        layer.attn_norm_b   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM,   "bias",   i), {n_embd});
-
-                        layer.wqkv          = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_QKV,    "weight", i), {n_embd, n_embd + 2*n_embd_gqa});
-                        layer.bqkv          = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_QKV,    "bias",   i), {n_embd + 2*n_embd_gqa});
-
-                        layer.wo            = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT,    "weight", i), {n_embd, n_embd});
-                        layer.bo            = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT,    "bias",   i), {n_embd});
-
-                        layer.ffn_down      = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN,    "weight", i), {n_ff, n_embd});
-                        layer.ffn_down_b    = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_DOWN,    "bias",   i), {n_embd});
-
-                        layer.ffn_up        = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP,      "weight", i), {n_embd, n_ff});
-                        layer.ffn_up_b      = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_UP,      "bias",   i), {n_ff});
-
-                        layer.ffn_norm      = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM,    "weight", i), {n_embd});
-                        layer.ffn_norm_b    = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM,    "bias",   i), {n_embd});
-
-                        layer.attn_q_norm   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {64});
-                        layer.attn_q_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q_NORM, "bias",   i), {64});
-
-                        layer.attn_k_norm   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {64});
-                        layer.attn_k_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K_NORM, "bias",   i), {64});
-                    }
-                } break;
             case LLM_ARCH_BERT:
             case LLM_ARCH_NOMIC_BERT:
                 {
@@ -5276,7 +5537,7 @@ static bool llm_load_tensors(
 
                             layer.ffn_down_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_DOWN, "bias", i),   {n_embd});
                         } else {
-                            layer.ffn_gate   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff});
+                            layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff});
                         }
 
                         layer.layer_out_norm   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_LAYER_OUT_NORM, "weight", i), {n_embd});
@@ -5299,14 +5560,14 @@ static bool llm_load_tensors(
                         layer.wq   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd});
                         layer.bq   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q,   "bias", i),   {n_embd});
 
-                        layer.attn_q_norm   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd}, false);
-                        layer.attn_q_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q_NORM, "bias", i), {n_embd}, false);
+                        layer.attn_q_norm   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.attn_q_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q_NORM, "bias", i), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
 
                         layer.wk   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_gqa});
                         layer.bk   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K,   "bias", i),   {n_embd_gqa});
 
-                        layer.attn_k_norm   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd}, false);
-                        layer.attn_k_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K_NORM, "bias", i), {n_embd}, false);
+                        layer.attn_k_norm   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.attn_k_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K_NORM, "bias", i), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
 
                         layer.wv   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_gqa});
                         layer.bv   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_V,   "bias", i),   {n_embd_gqa});
@@ -5317,6 +5578,9 @@ static bool llm_load_tensors(
                         layer.attn_out_norm   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT_NORM, "weight", i), {n_embd}); //output_norm
                         layer.attn_out_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT_NORM, "bias", i),   {n_embd});
 
+                        layer.attn_norm_2   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM_2, "weight", i), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.attn_norm_2_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM_2, "bias", i),   {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
+
                         layer.ffn_up = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP,        "weight", i), {n_embd, n_ff});
                         layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE,    "weight", i), {n_embd, n_ff});
 
@@ -5368,18 +5632,16 @@ static bool llm_load_tensors(
             case LLM_ARCH_MPT:
                 {
                     model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
-                    model.pos_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_POS_EMBD,   "weight"), {n_embd, hparams.n_ctx_train}, false);
+                    model.pos_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_POS_EMBD,   "weight"), {n_embd, hparams.n_ctx_train}, llama_model_loader::TENSOR_NOT_REQUIRED);
 
                     // output
                     {
                         model.output_norm   = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
-                        model.output_norm_b = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "bias"),   {n_embd}, false);
+                        model.output_norm_b = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "bias"),   {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
 
-                        model.output        = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, false);
+                        model.output        = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_NOT_REQUIRED);
                         if (!model.output) {
-                            model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}); // needs to be on GPU
-                            ml.n_created--; // artificial tensor
-                            ml.size_data += ggml_nbytes(model.output);
+                            model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED); // needs to be on GPU
                         }
                     }
 
@@ -5390,31 +5652,31 @@ static bool llm_load_tensors(
                         auto & layer = model.layers[i];
 
                         layer.attn_norm   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
-                        layer.attn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "bias", i),   {n_embd}, false);
+                        layer.attn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "bias", i),   {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
 
                         layer.wqkv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa});
-                        layer.bqkv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_QKV, "bias", i),   {n_embd + 2*n_embd_gqa}, false);
+                        layer.bqkv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_QKV, "bias", i),   {n_embd + 2*n_embd_gqa}, llama_model_loader::TENSOR_NOT_REQUIRED);
 
                         layer.wo   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd});
-                        layer.bo   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT, "bias", i),   {n_embd}, false);
+                        layer.bo   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT, "bias", i),   {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
 
                         layer.ffn_norm   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
-                        layer.ffn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "bias", i),   {n_embd}, false);
+                        layer.ffn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "bias", i),   {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
 
                         layer.ffn_down   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd});
-                        layer.ffn_down_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_DOWN, "bias", i),   {n_embd}, false);
+                        layer.ffn_down_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_DOWN, "bias", i),   {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
 
                         layer.ffn_up     = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff});
-                        layer.ffn_up_b   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_UP,   "bias", i),   {n_ff}, false);
+                        layer.ffn_up_b   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_UP,   "bias", i),   {n_ff}, llama_model_loader::TENSOR_NOT_REQUIRED);
 
-                        layer.attn_q_norm   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd}, false);
-                        layer.attn_q_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q_NORM, "bias",   i), {n_embd}, false);
+                        layer.attn_q_norm   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.attn_q_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q_NORM, "bias",   i), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
 
-                        layer.attn_k_norm   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd}, false);
-                        layer.attn_k_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K_NORM, "bias",   i), {n_embd}, false);
+                        layer.attn_k_norm   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.attn_k_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K_NORM, "bias",   i), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
 
                         // AWQ ScaleActivation layer
-                        layer.ffn_act = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_ACT, "scales", i), {n_ff}, false);
+                        layer.ffn_act = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_ACT, "scales", i), {n_ff}, llama_model_loader::TENSOR_NOT_REQUIRED);
                     }
                 } break;
             case LLM_ARCH_STABLELM:
@@ -5443,17 +5705,17 @@ static bool llm_load_tensors(
                         layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd});
 
                         // optional bias tensors, present in Stable LM 2 1.6B
-                        layer.bq = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q,   "bias", i), {n_embd},     false);
-                        layer.bk = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K,   "bias", i), {n_embd_gqa}, false);
-                        layer.bv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_V,   "bias", i), {n_embd_gqa}, false);
+                        layer.bq = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q,   "bias", i), {n_embd},     llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.bk = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K,   "bias", i), {n_embd_gqa}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.bv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_V,   "bias", i), {n_embd_gqa}, llama_model_loader::TENSOR_NOT_REQUIRED);
 
                         // optional q and k layernorms, present in StableLM 2 12B
-                        layer.attn_q_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {hparams.n_embd_head_k, hparams.n_head}, false);
-                        layer.attn_k_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {hparams.n_embd_head_k, hparams.n_head_kv}, false);
+                        layer.attn_q_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {hparams.n_embd_head_k, hparams.n_head}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.attn_k_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {hparams.n_embd_head_k, hparams.n_head_kv}, llama_model_loader::TENSOR_NOT_REQUIRED);
 
                         // optional FFN norm, not present in StableLM 2 12B which uses parallel residual
-                        layer.ffn_norm   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, false);
-                        layer.ffn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "bias", i),   {n_embd}, false);
+                        layer.ffn_norm   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.ffn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "bias", i),   {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
 
                         layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff});
                         layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd});
@@ -5496,12 +5758,10 @@ static bool llm_load_tensors(
                     // output
                     {
                         model.output_norm = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
-                        model.output      = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, false);
+                        model.output      = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_NOT_REQUIRED);
                         // if output is NULL, init from the input tok embed
                         if (model.output == NULL) {
-                            model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
-                            ml.n_created--; // artificial tensor
-                            ml.size_data += ggml_nbytes(model.output);
+                            model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED);
                         }
                     }
 
@@ -5599,8 +5859,8 @@ static bool llm_load_tensors(
                         layer.attn_norm   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
                         layer.attn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "bias", i),   {n_embd});
 
-                        layer.wqkv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, false);
-                        layer.bqkv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_QKV, "bias", i),   {n_embd + 2*n_embd_gqa}, false);
+                        layer.wqkv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.bqkv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_QKV, "bias", i),   {n_embd + 2*n_embd_gqa}, llama_model_loader::TENSOR_NOT_REQUIRED);
 
                         if (layer.wqkv == nullptr) {
                             layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd});
@@ -5637,17 +5897,20 @@ static bool llm_load_tensors(
                         ggml_context* ctx_layer = ctx_for_layer(i);
                         ggml_context* ctx_split = ctx_for_layer_split(i);
 
-                        auto& layer = model.layers[i];
+                        auto & layer = model.layers[i];
 
                         layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), { n_embd });
 
-                        layer.wqkv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_QKV, "weight", i), { n_embd, n_embd + 2 * n_embd_gqa }, false);
-                        layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), { n_embd, n_embd });
+                        layer.wqkv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_QKV, "weight", i), { n_embd, n_embd + 2 * n_embd_gqa }, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.wo   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), { n_embd, n_embd });
 
                         layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), { n_embd });
 
                         layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd });
                         layer.ffn_up = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP, "weight", i), { n_embd, 2 * n_ff });
+
+                        layer.rope_long  = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ROPE_FACTORS_LONG,  "weight"), { n_embd_head/2 }, llama_model_loader::TENSOR_NOT_REQUIRED | (i != 0 ? llama_model_loader::TENSOR_DUPLICATED : 0));
+                        layer.rope_short = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ROPE_FACTORS_SHORT, "weight"), { n_embd_head/2 }, llama_model_loader::TENSOR_NOT_REQUIRED | (i != 0 ? llama_model_loader::TENSOR_DUPLICATED : 0));
                     }
                 } break;
             case LLM_ARCH_PLAMO:
@@ -5816,9 +6079,7 @@ static bool llm_load_tensors(
 
                     // output
                     model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
-                    model.output      = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD,  "weight"), {n_embd, n_vocab}); // same as tok_embd, duplicated to allow offloading
-                    ml.n_created--; // artificial tensor
-                    ml.size_data += ggml_nbytes(model.output);
+                    model.output      = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD,  "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED); // same as tok_embd, duplicated to allow offloading
 
                     const int64_t n_ff          = hparams.n_ff;
                     const int64_t n_embd_head_k = hparams.n_embd_head_k;
@@ -5853,12 +6114,10 @@ static bool llm_load_tensors(
                         model.output_norm   = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
                         model.output_norm_b = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "bias"),   {n_embd});
 
-                        model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, false);
+                        model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_NOT_REQUIRED);
                         // if output is NULL, init from the input tok embed
                         if (model.output == NULL) {
-                            model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
-                            ml.n_created--; // artificial tensor
-                            ml.size_data += ggml_nbytes(model.output);
+                            model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED);
                         }
 
                     }
@@ -5909,12 +6168,10 @@ static bool llm_load_tensors(
                     {
                         model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
 
-                        model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, false);
+                        model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_NOT_REQUIRED);
                         // if output is NULL, init from the input tok embed, duplicated to allow offloading
                         if (model.output == NULL) {
-                            model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
-                            ml.n_created--; // artificial tensor
-                            ml.size_data += ggml_nbytes(model.output);
+                            model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED);
                         }
                     }
 
@@ -5975,9 +6232,7 @@ static bool llm_load_tensors(
                     {
                         model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
                         // init output from the input tok embed
-                        model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
-                        ml.n_created--; // artificial tensor
-                        ml.size_data += ggml_nbytes(model.output);
+                        model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED);
                     }
 
                     for (int i = 0; i < n_layer; ++i) {
@@ -6009,12 +6264,10 @@ static bool llm_load_tensors(
 
                     // output
                     {
-                        model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, false);
+                        model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_NOT_REQUIRED);
                         // if output is NULL, init from the input tok embed
                         if (model.output == NULL) {
-                            model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
-                            ml.n_created--; // artificial tensor
-                            ml.size_data += ggml_nbytes(model.output);
+                            model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED);
                         }
                     }
 
@@ -6034,84 +6287,223 @@ static bool llm_load_tensors(
                         layer.ffn_up   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff});
                     }
                 } break;
-            default:
-                throw std::runtime_error("unknown architecture");
-        }
-    }
+            case LLM_ARCH_GPTNEOX:
+                {
+                    model.tok_embd   = ml.create_tensor(ctx_input,  tn(LLM_TENSOR_TOKEN_EMBD,      "weight"), {n_embd, n_vocab});
+                    // output
+                    {
+                        model.output_norm   = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
+                        model.output_norm_b = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "bias"),   {n_embd});
+                        model.output        = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab});
+                    }
 
-    ml.done_getting_tensors();
+                    for (int i = 0; i < n_layer; ++i) {
+                        ggml_context * ctx_layer = ctx_for_layer(i);
+                        ggml_context * ctx_split = ctx_for_layer_split(i);
 
-    ml.init_mappings(true, use_mlock ? &model.mlock_mmaps : nullptr);
-    model.mappings.reserve(ml.mappings.size());
+                        auto & layer = model.layers[i];
 
-    // create the backend buffers
-    std::vector<std::pair<ggml_context *, llama_buf_map>> ctx_bufs;
-    ctx_bufs.reserve(ctx_map.size());
+                        layer.attn_norm   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
+                        layer.attn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "bias", i),   {n_embd});
 
-    // Ensure we have enough capacity for the maximum backend buffer we will potentially create
-    size_t n_max_backend_buffer = ctx_map.size() * ml.files.size();
-    model.bufs.reserve(n_max_backend_buffer);
+                        layer.wqkv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa});
+                        layer.bqkv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_QKV, "bias", i),   {n_embd + 2*n_embd_gqa});
 
-    for (auto & it : ctx_map) {
-        ggml_backend_buffer_type_t buft = it.first;
-        ggml_context * ctx              = it.second;
+                        layer.wo   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd});
+                        layer.bo   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT, "bias", i),   {n_embd});
 
-        llama_buf_map bufs;
-        bufs.reserve(n_max_backend_buffer);
+                        layer.ffn_norm   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
+                        layer.ffn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "bias", i),   {n_embd});
 
-        // only the mmap region containing the tensors in the model is mapped to the backend buffer
-        // this is important for metal with apple silicon: if the entire model could be mapped to a metal buffer, then we could just use metal for all layers
-        // this allows using partial offloading when the model size exceeds the metal buffer size, but not the RAM size
-        if (ml.use_mmap && use_mmap_buffer && buft == llama_default_buffer_type_cpu(true)) {
-            for (uint32_t idx = 0; idx < ml.files.size(); idx++) {
-                void * addr = nullptr;
-                size_t first, last;
-                ml.get_mapping_range(&first, &last, &addr, idx, ctx);
-                if (first >= last) {
-                    continue;
-                }
-                ggml_backend_buffer_t buf = ggml_backend_cpu_buffer_from_ptr((char *) addr + first, last - first);
-                if (buf == nullptr) {
-                    throw std::runtime_error("unable to allocate backend CPU buffer");
-                }
-                model.bufs.push_back(buf);
-                bufs.emplace(idx, buf);
-#ifdef GGML_USE_CUDA
-                if (n_layer >= n_gpu_layers) {
-                    ggml_backend_cuda_register_host_buffer(
-                        ggml_backend_buffer_get_base(buf),
-                        ggml_backend_buffer_get_size(buf));
-                }
-#endif
-            }
-        }
-#ifdef GGML_USE_METAL
-        else if (ml.use_mmap && use_mmap_buffer && buft == ggml_backend_metal_buffer_type()) {
-            for (uint32_t idx = 0; idx < ml.files.size(); idx++) {
-                const size_t max_size = ggml_get_max_tensor_size(ctx);
-                void * addr = nullptr;
-                size_t first, last;
-                ml.get_mapping_range(&first, &last, &addr, idx, ctx);
-                if (first >= last) {
-                    continue;
-                }
-                ggml_backend_buffer_t buf = ggml_backend_metal_buffer_from_ptr((char *) addr + first, last - first, max_size);
-                if (buf == nullptr) {
-                    throw std::runtime_error("unable to allocate backend metal buffer");
-                }
-                model.bufs.push_back(buf);
-                bufs.emplace(idx, buf);
-            }
-        }
-#endif
-        else {
-            ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx, buft);
-            if (buf == nullptr) {
-                throw std::runtime_error("unable to allocate backend buffer");
-            }
-            model.bufs.push_back(buf);
-            if (use_mlock && ggml_backend_buffer_is_host(buf)) {
-                model.mlock_bufs.emplace_back(new llama_mlock);
+                        layer.ffn_down   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd});
+                        layer.ffn_down_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_DOWN, "bias", i),   {n_embd});
+
+                        layer.ffn_up     = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_ff});
+                        layer.ffn_up_b   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_UP,   "bias", i),   {n_ff});
+                    }
+                } break;
+            case LLM_ARCH_ARCTIC:
+                {
+                    model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
+
+                    // output
+                    {
+                        model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
+                        model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        // if output is NULL, init from the input tok embed
+                        if (model.output == NULL) {
+                            model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED);
+                        }
+                    }
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        ggml_context * ctx_layer = ctx_for_layer(i);
+                        ggml_context * ctx_split = ctx_for_layer_split(i);
+
+                        auto & layer = model.layers[i];
+
+                        layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
+
+                        layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd});
+                        layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_gqa});
+                        layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_gqa});
+                        layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd});
+
+                        layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
+
+                        layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_embd});
+                        layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_embd, n_embd});
+                        layer.ffn_up   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_embd});
+
+                        layer.ffn_gate_inp = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert});
+                        layer.ffn_norm_exps = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM_EXPS, "weight", i), {n_embd});
+                        layer.ffn_gate_exps = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd,   n_ff, n_expert}, false);
+                        layer.ffn_down_exps = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {  n_ff, n_embd, n_expert});
+                        layer.ffn_up_exps   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {n_embd,   n_ff, n_expert});
+                    }
+                } break;
+            case LLM_ARCH_DEEPSEEK2:
+                {
+                    bool is_lite = (hparams.n_layer == 27);
+
+                    const uint32_t n_embd_head_qk_rope = hparams.n_rot;
+                    const uint32_t n_embd_head_qk_nope = hparams.n_embd_head_k - hparams.n_rot;
+                    const uint32_t q_lora_rank = hparams.n_lora_q;
+                    const uint32_t kv_lora_rank = hparams.n_lora_kv;
+                    const uint32_t n_ff_exp = hparams.n_ff_exp;
+
+                    model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
+
+                    // output
+                    {
+                        model.output_norm = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
+                        model.output      = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab});
+                    }
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        ggml_context * ctx_layer = ctx_for_layer(i);
+                        ggml_context * ctx_split = ctx_for_layer_split(i);
+
+                        auto & layer = model.layers[i];
+
+                        layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
+                        if (!is_lite) {
+                            layer.attn_q_a_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q_A_NORM, "weight", i), {q_lora_rank});
+                        }
+                        layer.attn_kv_a_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_KV_A_NORM, "weight", i), {kv_lora_rank});
+
+                        if (!is_lite) {
+                            layer.wq_a = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q_A,   "weight", i), {n_embd, q_lora_rank});
+                            layer.wq_b = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q_B,   "weight", i), {q_lora_rank, hparams.n_head * hparams.n_embd_head_k});
+                        } else {
+                            layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd_k_gqa});
+                        }
+                        layer.wkv_a_mqa = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_KV_A_MQA,   "weight", i), {n_embd, kv_lora_rank + n_embd_head_qk_rope});
+                        layer.wkv_b = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_KV_B,   "weight", i), {kv_lora_rank, hparams.n_head * (n_embd_head_qk_nope + hparams.n_embd_head_v)});
+                        layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {hparams.n_head * hparams.n_embd_head_v, n_embd});
+
+                        layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
+
+                        if ((uint32_t) i < hparams.n_layer_dense_lead) {
+                            layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff});
+                            layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd});
+                            layer.ffn_up   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff});
+                        } else {
+                            layer.ffn_gate_inp = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert});
+
+                            GGML_ASSERT(hparams.n_expert      > 0);
+                            GGML_ASSERT(hparams.n_expert_used > 0);
+
+                            // MoE branch
+                            layer.ffn_gate_exps = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {  n_embd, n_ff_exp, n_expert});
+                            layer.ffn_down_exps = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp,   n_embd, n_expert});
+                            layer.ffn_up_exps   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {  n_embd, n_ff_exp, n_expert});
+
+                            // Shared expert branch
+                            layer.ffn_gate_shexp = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd,   n_ff_exp * hparams.n_expert_shared});
+                            layer.ffn_down_shexp = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {  n_ff_exp * hparams.n_expert_shared, n_embd});
+                            layer.ffn_up_shexp   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {n_embd,   n_ff_exp * hparams.n_expert_shared});
+                        }
+                    }
+                } break;
+            default:
+                throw std::runtime_error("unknown architecture");
+        }
+    }
+
+    ml.done_getting_tensors();
+
+    ml.init_mappings(true, use_mlock ? &model.mlock_mmaps : nullptr);
+    model.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());
+
+    // Ensure we have enough capacity for the maximum backend buffer we will potentially create
+    size_t n_max_backend_buffer = ctx_map.size() * ml.files.size();
+    model.bufs.reserve(n_max_backend_buffer);
+
+    for (auto & it : ctx_map) {
+        ggml_backend_buffer_type_t buft = it.first;
+        ggml_context * ctx              = it.second;
+
+        llama_buf_map bufs;
+        bufs.reserve(n_max_backend_buffer);
+
+        // only the mmap region containing the tensors in the model is mapped to the backend buffer
+        // this is important for metal with apple silicon: if the entire model could be mapped to a metal buffer, then we could just use metal for all layers
+        // this allows using partial offloading when the model size exceeds the metal buffer size, but not the RAM size
+        if (ml.use_mmap && use_mmap_buffer && buft == llama_default_buffer_type_cpu(true)) {
+            for (uint32_t idx = 0; idx < ml.files.size(); idx++) {
+                void * addr = nullptr;
+                size_t first, last;
+                ml.get_mapping_range(&first, &last, &addr, idx, ctx);
+                if (first >= last) {
+                    continue;
+                }
+                ggml_backend_buffer_t buf = ggml_backend_cpu_buffer_from_ptr((char *) addr + first, last - first);
+                if (buf == nullptr) {
+                    throw std::runtime_error("unable to allocate backend CPU buffer");
+                }
+                model.bufs.push_back(buf);
+                bufs.emplace(idx, buf);
+#ifdef GGML_USE_CUDA
+                if (n_layer >= n_gpu_layers) {
+                    ggml_backend_cuda_register_host_buffer(
+                        ggml_backend_buffer_get_base(buf),
+                        ggml_backend_buffer_get_size(buf));
+                }
+#endif
+            }
+        }
+#ifdef GGML_USE_METAL
+        else if (ml.use_mmap && use_mmap_buffer && buft == ggml_backend_metal_buffer_type()) {
+            for (uint32_t idx = 0; idx < ml.files.size(); idx++) {
+                const size_t max_size = ggml_get_max_tensor_size(ctx);
+                void * addr = nullptr;
+                size_t first, last;
+                ml.get_mapping_range(&first, &last, &addr, idx, ctx);
+                if (first >= last) {
+                    continue;
+                }
+                ggml_backend_buffer_t buf = ggml_backend_metal_buffer_from_ptr((char *) addr + first, last - first, max_size);
+                if (buf == nullptr) {
+                    throw std::runtime_error("unable to allocate backend metal buffer");
+                }
+                model.bufs.push_back(buf);
+                bufs.emplace(idx, buf);
+            }
+        }
+#endif
+        else {
+            ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx, buft);
+            if (buf == nullptr) {
+                throw std::runtime_error("unable to allocate backend buffer");
+            }
+            model.bufs.push_back(buf);
+            if (use_mlock && ggml_backend_buffer_is_host(buf)) {
+                model.mlock_bufs.emplace_back(new llama_mlock);
                 auto & mlock_buf = model.mlock_bufs.back();
                 mlock_buf->init   (ggml_backend_buffer_get_base(buf));
                 mlock_buf->grow_to(ggml_backend_buffer_get_size(buf));
@@ -6233,16 +6625,6 @@ static int llama_model_load(const std::string & fname, llama_model & model, llam
         }
 #endif
 
-#ifdef GGML_USE_SYCL
-        if (params.split_mode == LLAMA_SPLIT_MODE_NONE) {
-            ggml_backend_sycl_set_single_device_mode(params.main_gpu);
-            //SYCL use device index (0, 1, 2) directly, uer input device id, then convert to device index.
-            params.main_gpu = ggml_backend_sycl_get_device_index(params.main_gpu);
-        } else {
-            ggml_backend_sycl_set_mul_device_mode();
-        }
-#endif
-
         if (!llm_load_tensors(
             ml, model, params.n_gpu_layers, params.split_mode,  params.main_gpu, params.tensor_split, params.use_mlock,
             params.progress_callback, params.progress_callback_user_data
@@ -6298,10 +6680,7 @@ static struct ggml_tensor * llm_build_inp_embd(
 
         inpL = ggml_get_rows(ctx, tok_embd, lctx.inp_tokens);
     } else {
-#ifdef GGML_USE_MPI
-        GGML_ASSERT(false && "not implemented");
-#endif
-        lctx.inp_embd = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, batch.n_tokens);
+       lctx.inp_embd = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, batch.n_tokens);
         inpL = lctx.inp_embd;
         ggml_set_input(lctx.inp_embd);
     }
@@ -6491,6 +6870,8 @@ static struct ggml_tensor * llm_build_moe_ffn(
                     int64_t   n_expert_used,
             llm_ffn_op_type   type_op,
                        bool   norm_w,
+                       bool   scale_w,
+                      float   w_scale,
          const llm_build_cb & cb,
                         int   il) {
     int64_t n_embd = cur->ne[0];
@@ -6522,6 +6903,10 @@ static struct ggml_tensor * llm_build_moe_ffn(
 
         weights = ggml_reshape_3d(ctx, weights, 1, n_expert_used, n_tokens);
     }
+    if (scale_w) {
+        weights = ggml_scale(ctx, weights, w_scale);
+        cb(weights, "ffn_moe_weights_scaled", il);
+    }
 
     cur = ggml_reshape_3d(ctx, cur, n_embd, 1, n_tokens);
     ggml_tensor * up = ggml_mul_mat_id(ctx, up_exps, cur, selected_experts); // [n_ff, n_expert_used, n_tokens]
@@ -6596,6 +6981,7 @@ static struct ggml_tensor * llm_build_kqv(
     const int64_t n_embd_head_k = hparams.n_embd_head_k;
     const int64_t n_embd_k_gqa  = hparams.n_embd_k_gqa();
     const int64_t n_embd_head_v = hparams.n_embd_head_v;
+    const int64_t n_embd_v_gqa  = hparams.n_embd_v_gqa();
 
     struct ggml_tensor * q = ggml_permute(ctx, q_cur, 0, 2, 1, 3);
     cb(q, "q", il);
@@ -6618,23 +7004,23 @@ static struct ggml_tensor * llm_build_kqv(
         struct ggml_tensor * v =
             ggml_view_3d(ctx, kv.v_l[il],
                     n_embd_head_v, n_kv, n_head_kv,
-                    ggml_row_size(kv.v_l[il]->type, n_embd_k_gqa),
-                    ggml_row_size(kv.v_l[il]->type, n_embd_head_k),
+                    ggml_row_size(kv.v_l[il]->type, n_embd_v_gqa),
+                    ggml_row_size(kv.v_l[il]->type, n_embd_head_v),
                     0);
         cb(v, "v", il);
 
         cur = ggml_flash_attn_ext(ctx, q, k, v, kq_mask, kq_scale, hparams.f_max_alibi_bias);
 
-        if (model.arch == LLM_ARCH_PHI2 || model.arch == LLM_ARCH_PHI3) {
+        if (model.arch == LLM_ARCH_PHI2 || model.arch == LLM_ARCH_PHI3 || model.arch == LLM_ARCH_GPTNEOX) {
             ggml_flash_attn_ext_set_prec(cur, GGML_PREC_F32);
         }
 
-        cur = ggml_reshape_2d(ctx, cur, n_embd_head_k*n_head, n_tokens);
+        cur = ggml_reshape_2d(ctx, cur, n_embd_head_v*n_head, n_tokens);
     } else {
         struct ggml_tensor * kq = ggml_mul_mat(ctx, k, q);
         cb(kq, "kq", il);
 
-        if (model.arch == LLM_ARCH_PHI2 || model.arch == LLM_ARCH_PHI3) {
+        if (model.arch == LLM_ARCH_PHI2 || model.arch == LLM_ARCH_PHI3 || model.arch == LLM_ARCH_GPTNEOX) {
             // for this arch, we need to perform the KQ multiplication with F32 precision, otherwise we get NaNs
             // ref: https://github.com/ggerganov/llama.cpp/pull/4490#issuecomment-1859055847
             ggml_mul_mat_set_prec(kq, GGML_PREC_F32);
@@ -6674,7 +7060,7 @@ static struct ggml_tensor * llm_build_kqv(
         struct ggml_tensor * kqv_merged = ggml_permute(ctx, kqv, 0, 2, 1, 3);
         cb(kqv_merged, "kqv_merged", il);
 
-        cur = ggml_cont_2d(ctx, kqv_merged, n_embd_head_k*n_head, n_tokens);
+        cur = ggml_cont_2d(ctx, kqv_merged, n_embd_head_v*n_head, n_tokens);
         cb(cur, "kqv_merged_cont", il);
     }
 
@@ -6763,7 +7149,7 @@ struct llm_build_context {
     const int32_t n_kv;     // size of KV cache to consider (n_kv <= kv_self.size)
     const int32_t n_outputs;
     const int32_t kv_head;  // index of where we store new KV data in the cache
-    const int32_t n_orig_ctx;
+    const int32_t n_ctx_orig;
 
     const bool flash_attn;
 
@@ -6812,7 +7198,7 @@ struct llm_build_context {
         n_kv             (worst_case ? kv_self.size : kv_self.n),
         n_outputs        (worst_case ? n_tokens : lctx.n_outputs),
         kv_head          (worst_case ? (kv_self.recurrent ? 0 : kv_self.size - n_tokens) : kv_self.head),
-        n_orig_ctx       (cparams.n_yarn_orig_ctx),
+        n_ctx_orig       (cparams.n_ctx_orig_yarn),
         flash_attn       (cparams.flash_attn),
         pooling_type     (cparams.pooling_type),
         rope_type        (hparams.rope_type),
@@ -6859,17 +7245,20 @@ struct llm_build_context {
         cb(lctx.inp_K_shift, "K_shift", -1);
         ggml_set_input(lctx.inp_K_shift);
 
+
         for (int il = 0; il < n_layer; ++il) {
+            struct ggml_tensor * rope_factors = build_rope_factors(il);
             struct ggml_tensor * tmp =
                 // we rotate only the first n_rot dimensions
-                ggml_rope_custom_inplace(ctx0,
+                ggml_rope_ext_inplace(ctx0,
                         ggml_view_3d(ctx0, kv_self.k_l[il],
                             n_embd_head_k, n_head_kv, n_ctx,
                             ggml_row_size(kv_self.k_l[il]->type, n_embd_head_k),
                             ggml_row_size(kv_self.k_l[il]->type, n_embd_k_gqa),
                             0),
-                        lctx.inp_K_shift, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
+                        lctx.inp_K_shift, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
                         ext_factor, attn_factor, beta_fast, beta_slow);
+
             cb(tmp, "K_shifted", il);
             ggml_build_forward_expand(gf, tmp);
         }
@@ -6972,6 +7361,17 @@ struct llm_build_context {
         return lctx.inp_pos;
     }
 
+    struct ggml_tensor * build_rope_factors(int il) {
+        // choose long/short freq factors based on the context size
+        const auto n_ctx_pre_seq = cparams.n_ctx / cparams.n_seq_max;
+
+        if (n_ctx_pre_seq > hparams.n_ctx_orig_yarn) {
+            return model.layers[il].rope_long;
+        }
+
+        return model.layers[il].rope_short;
+    }
+
     struct ggml_tensor * build_inp_out_ids() {
         lctx.inp_out_ids = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_outputs);
         cb(lctx.inp_out_ids, "inp_out_ids", -1);
@@ -7079,16 +7479,16 @@ struct llm_build_context {
                     cb(Vcur, "Vcur", il);
                 }
 
-                Qcur = ggml_rope_custom(
-                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos,
-                    n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
+                Qcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), 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);
 
-                Kcur = ggml_rope_custom(
-                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos,
-                    n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
+                Kcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
                     ext_factor, attn_factor, beta_fast, beta_slow
                 );
                 cb(Kcur, "Kcur", il);
@@ -7117,9 +7517,9 @@ struct llm_build_context {
                 cb(cur, "ffn_norm", il);
 
                 cur = llm_build_ffn(ctx0, cur,
-                        model.layers[il].ffn_up,   NULL,
-                        model.layers[il].ffn_gate, NULL,
-                        model.layers[il].ffn_down, NULL,
+                        model.layers[il].ffn_up,   model.layers[il].ffn_up_b,
+                        model.layers[il].ffn_gate, model.layers[il].ffn_gate_b,
+                        model.layers[il].ffn_down, model.layers[il].ffn_down_b,
                         NULL,
                         LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
                 cb(cur, "ffn_out", il);
@@ -7137,6 +7537,7 @@ struct llm_build_context {
                         model.layers[il].ffn_down_exps,
                         n_expert, n_expert_used,
                         LLM_FFN_SILU, true,
+                        false, 0.0,
                         cb, il);
                 cb(cur, "ffn_moe_out", il);
             }
@@ -7209,14 +7610,14 @@ struct llm_build_context {
 
                 switch (model.type) {
                     case MODEL_7B:
-                        Qcur = ggml_rope_custom(
-                            ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos,
-                            n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
+                        Qcur = ggml_rope_ext(
+                            ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr,
+                            n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
                             ext_factor, attn_factor, beta_fast, beta_slow
                         );
-                        Kcur = ggml_rope_custom(
-                            ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos,
-                            n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
+                        Kcur = ggml_rope_ext(
+                            ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
+                            n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
                             ext_factor, attn_factor, beta_fast, beta_slow
                         );
                         break;
@@ -7321,16 +7722,16 @@ struct llm_build_context {
                 struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv, cur);
                 cb(Vcur, "Vcur", il);
 
-                Qcur = ggml_rope_custom(
-                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos,
-                    n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
+                Qcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), 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);
 
-                Kcur = ggml_rope_custom(
-                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos,
-                    n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
+                Kcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
                     ext_factor, attn_factor, beta_fast, beta_slow
                 );
                 cb(Kcur, "Kcur", il);
@@ -7442,14 +7843,14 @@ struct llm_build_context {
                 Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
 
                 // using mode = 2 for neox mode
-                Qcur = ggml_rope_custom(
-                    ctx0, Qcur, inp_pos, n_rot, rope_type, 0, n_orig_ctx,
+                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
                 );
                 cb(Qcur, "Qcur", il);
 
-                Kcur = ggml_rope_custom(
-                    ctx0, Kcur, inp_pos, n_rot, rope_type, 0, n_orig_ctx,
+                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(Kcur, "Kcur", il);
@@ -7565,16 +7966,16 @@ struct llm_build_context {
                     cb(Vcur, "Vcur", il);
                 }
 
-                Qcur = ggml_rope_custom(
-                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos,
-                    n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
+                Qcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), 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);
 
-                Kcur = ggml_rope_custom(
-                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos,
-                    n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
+                Kcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
                     ext_factor, attn_factor, beta_fast, beta_slow
                 );
                 cb(Kcur, "Kcur", il);
@@ -7618,6 +8019,7 @@ struct llm_build_context {
                     model.layers[il].ffn_down_exps,
                     n_expert, n_expert_used,
                     LLM_FFN_GELU, true,
+                    false, 0.0,
                     cb, il);
             cb(cur, "ffn_moe_out", il);
 
@@ -7717,16 +8119,16 @@ struct llm_build_context {
                 cb(Kcur, "Kcur", il);
                 cb(Vcur, "Vcur", il);
 
-                Qcur = ggml_rope_custom(
-                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos,
-                    n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
+                Qcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), 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);
 
-                Kcur = ggml_rope_custom(
-                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos,
-                    n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
+                Kcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
                     ext_factor, attn_factor, beta_fast, beta_slow
                 );
                 cb(Kcur, "Kcur", il);
@@ -7761,6 +8163,7 @@ struct llm_build_context {
                     model.layers[il].ffn_down_exps,
                     n_expert, n_expert_used,
                     LLM_FFN_SILU, true,
+                    false, 0.0,
                     cb, il);
             cb(cur, "ffn_moe_out", il);
 
@@ -7894,298 +8297,91 @@ struct llm_build_context {
         return gf;
     }
 
-    struct ggml_cgraph * build_persimmon() {
+    struct ggml_cgraph * build_refact() {
         struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
 
         const int64_t n_embd_head = hparams.n_embd_head_v;
-        GGML_ASSERT(n_embd_head   == hparams.n_embd_head_k);
-        GGML_ASSERT(n_embd_head/2 == hparams.n_rot);
+        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
 
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
         inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
 
-        // inp_pos - contains the positions
-        struct ggml_tensor * inp_pos = build_inp_pos();
-
         // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
         struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
 
         for (int il = 0; il < n_layer; ++il) {
-            struct ggml_tensor * residual = inpL;
+            struct ggml_tensor * inpSA = inpL;
 
             cur = llm_build_norm(ctx0, inpL, hparams,
-                    model.layers[il].attn_norm,
-                    model.layers[il].attn_norm_b,
-                    LLM_NORM, cb, il);
+                    model.layers[il].attn_norm, NULL,
+                    LLM_NORM_RMS, cb, il);
             cb(cur, "attn_norm", il);
 
-            // self attention
+            // self-attention
             {
-                cur = ggml_mul_mat(ctx0, model.layers[il].wqkv, cur);
-                cb(cur, "wqkv", il);
-
-                cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
-                cb(cur, "bqkv", il);
-
-                // split qkv
-                GGML_ASSERT(n_head_kv == n_head);
-
-                struct ggml_tensor * tmpqkv = ggml_reshape_4d(ctx0, cur, n_embd_head, 3, n_head, n_tokens);
-                cb(tmpqkv, "tmpqkv", il);
-
-                struct ggml_tensor * tmpqkv_perm = ggml_cont(ctx0, ggml_permute(ctx0, tmpqkv, 0, 3, 1, 2));
-                cb(tmpqkv_perm, "tmpqkv", il);
-
-                struct ggml_tensor * tmpq = ggml_view_3d(
-                        ctx0, tmpqkv_perm, n_embd_head, n_head, n_tokens,
-                        ggml_element_size(tmpqkv_perm) * n_embd_head,
-                        ggml_element_size(tmpqkv_perm) * n_embd_head * n_head,
-                        0
-                        );
-                cb(tmpq, "tmpq", il);
-
-                struct ggml_tensor * tmpk = ggml_view_3d(
-                        ctx0, tmpqkv_perm, n_embd_head, n_head, n_tokens,
-                        ggml_element_size(tmpqkv_perm) * n_embd_head,
-                        ggml_element_size(tmpqkv_perm) * n_embd_head * n_head,
-                        ggml_element_size(tmpqkv_perm) * n_embd_head * n_head * n_tokens
-                        );
-                cb(tmpk, "tmpk", il);
-
-                // Q/K Layernorm
-                tmpq = llm_build_norm(ctx0, tmpq, hparams,
-                        model.layers[il].attn_q_norm,
-                        model.layers[il].attn_q_norm_b,
-                        LLM_NORM, cb, il);
-                cb(tmpq, "tmpq", il);
-
-                tmpk = llm_build_norm(ctx0, tmpk, hparams,
-                        model.layers[il].attn_k_norm,
-                        model.layers[il].attn_k_norm_b,
-                        LLM_NORM, cb, il);
-                cb(tmpk, "tmpk", il);
-
-                // RoPE the first n_rot of q/k, pass the other half, and concat.
-                struct ggml_tensor * qrot = ggml_view_3d(
-                        ctx0, tmpq, n_rot, n_head, n_tokens,
-                        ggml_element_size(tmpq) * n_embd_head,
-                        ggml_element_size(tmpq) * n_embd_head * n_head,
-                        0
-                        );
-                cb(qrot, "qrot", il);
-
-                struct ggml_tensor * krot = ggml_view_3d(
-                        ctx0, tmpk, n_rot, n_head, n_tokens,
-                        ggml_element_size(tmpk) * n_embd_head,
-                        ggml_element_size(tmpk) * n_embd_head * n_head,
-                        0
-                        );
-                cb(krot, "krot", il);
-
-                // get the second half of tmpq, e.g tmpq[n_rot:, :, :]
-                struct ggml_tensor * qpass = ggml_view_3d(
-                        ctx0, tmpq, n_rot, n_head, n_tokens,
-                        ggml_element_size(tmpq) * n_embd_head,
-                        ggml_element_size(tmpq) * n_embd_head * n_head,
-                        ggml_element_size(tmpq) * n_rot
-                        );
-                cb(qpass, "qpass", il);
-
-                struct ggml_tensor * kpass = ggml_view_3d(
-                        ctx0, tmpk, n_rot, n_head, n_tokens,
-                        ggml_element_size(tmpk) * n_embd_head,
-                        ggml_element_size(tmpk) * n_embd_head * n_head,
-                        ggml_element_size(tmpk) * n_rot
-                        );
-                cb(kpass, "kpass", il);
-
-                struct ggml_tensor * qrotated = ggml_rope_custom(
-                    ctx0, qrot, inp_pos, n_rot, rope_type, 0, n_orig_ctx,
-                    freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow
-                );
-                cb(qrotated, "qrotated", il);
-
-                struct ggml_tensor * krotated = ggml_rope_custom(
-                    ctx0, krot, inp_pos, n_rot, rope_type, 0, n_orig_ctx,
-                    freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow
-                );
-                cb(krotated, "krotated", il);
-
-                // ggml currently only supports concatenation on dim=2
-                // so we need to permute qrot, qpass, concat, then permute back.
-                qrotated = ggml_cont(ctx0, ggml_permute(ctx0, qrotated, 2, 1, 0, 3));
-                cb(qrotated, "qrotated", il);
-
-                krotated = ggml_cont(ctx0, ggml_permute(ctx0, krotated, 2, 1, 0, 3));
-                cb(krotated, "krotated", il);
-
-                qpass = ggml_cont(ctx0, ggml_permute(ctx0, qpass, 2, 1, 0, 3));
-                cb(qpass, "qpass", il);
-
-                kpass = ggml_cont(ctx0, ggml_permute(ctx0, kpass, 2, 1, 0, 3));
-                cb(kpass, "kpass", il);
-
-                struct ggml_tensor * Qcur = ggml_concat(ctx0, qrotated, qpass);
+                struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].wq, cur);
                 cb(Qcur, "Qcur", il);
 
-                struct ggml_tensor * Kcur = ggml_concat(ctx0, krotated, kpass);
+                struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, model.layers[il].wk, cur);
                 cb(Kcur, "Kcur", il);
 
-                struct ggml_tensor * Q = ggml_cont(ctx0, ggml_permute(ctx0, Qcur, 2, 1, 0, 3));
-                cb(Q, "Q", il);
+                struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv, cur);
+                cb(Vcur, "Vcur", il);
 
-                Kcur = ggml_cont(ctx0, ggml_permute(ctx0, Kcur, 2, 1, 0, 3));
+                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
                 cb(Kcur, "Kcur", il);
 
-                struct ggml_tensor * Vcur = ggml_view_3d(
-                        ctx0, tmpqkv_perm, n_embd_head, n_head, n_tokens,
-                        ggml_element_size(tmpqkv_perm) * n_embd_head,
-                        ggml_element_size(tmpqkv_perm) * n_embd_head * n_head,
-                        ggml_element_size(tmpqkv_perm) * n_embd_head * n_head * n_tokens * 2
-                        );
-                cb(Vcur, "Vcur", il);
+                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+                cb(Qcur, "Qcur", il);
 
                 cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf,
-                        model.layers[il].wo, model.layers[il].bo,
-                        Kcur, Vcur, Q, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+                        model.layers[il].wo, NULL,
+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
             }
 
             if (il == n_layer - 1) {
                 // skip computing output for unused tokens
                 struct ggml_tensor * inp_out_ids = build_inp_out_ids();
-                cur      = ggml_get_rows(ctx0,      cur, inp_out_ids);
-                residual = ggml_get_rows(ctx0, residual, inp_out_ids);
+                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
             }
 
-            struct ggml_tensor * ffn_inp = ggml_add(ctx0, residual, cur);
+            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
             cb(ffn_inp, "ffn_inp", il);
 
             // feed-forward network
             {
                 cur = llm_build_norm(ctx0, ffn_inp, hparams,
-                        model.layers[il].ffn_norm,
-                        model.layers[il].ffn_norm_b,
-                        LLM_NORM, cb, il);
+                        model.layers[il].ffn_norm, NULL,
+                        LLM_NORM_RMS, cb, il);
                 cb(cur, "ffn_norm", il);
 
                 cur = llm_build_ffn(ctx0, cur,
-                        model.layers[il].ffn_up,   model.layers[il].ffn_up_b,
-                        NULL,                      NULL,
-                        model.layers[il].ffn_down, model.layers[il].ffn_down_b,
+                        model.layers[il].ffn_up,   NULL,
+                        model.layers[il].ffn_gate, NULL,
+                        model.layers[il].ffn_down, NULL,
                         NULL,
-                        LLM_FFN_RELU_SQR, LLM_FFN_SEQ, cb, il);
+                        LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
                 cb(cur, "ffn_out", il);
             }
 
             cur = ggml_add(ctx0, cur, ffn_inp);
             cb(cur, "l_out", il);
 
+            // input for next layer
             inpL = cur;
         }
 
         cur = inpL;
 
         cur = llm_build_norm(ctx0, cur, hparams,
-                model.output_norm,
-                model.output_norm_b,
-                LLM_NORM, cb, -1);
+                model.output_norm, NULL,
+                LLM_NORM_RMS, cb, -1);
         cb(cur, "result_norm", -1);
 
-        cur = ggml_mul_mat(ctx0, model.output, cur);
-        cb(cur, "result_output", -1);
-
-        ggml_build_forward_expand(gf, cur);
-
-        return gf;
-    }
-
-    struct ggml_cgraph * build_refact() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
-
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-
-        struct ggml_tensor * cur;
-        struct ggml_tensor * inpL;
-
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
-
-        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
-        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
-
-        for (int il = 0; il < n_layer; ++il) {
-            struct ggml_tensor * inpSA = inpL;
-
-            cur = llm_build_norm(ctx0, inpL, hparams,
-                    model.layers[il].attn_norm, NULL,
-                    LLM_NORM_RMS, cb, il);
-            cb(cur, "attn_norm", il);
-
-            // self-attention
-            {
-                struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].wq, cur);
-                cb(Qcur, "Qcur", il);
-
-                struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, model.layers[il].wk, cur);
-                cb(Kcur, "Kcur", il);
-
-                struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv, cur);
-                cb(Vcur, "Vcur", il);
-
-                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
-                cb(Kcur, "Kcur", il);
-
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
-                cb(Qcur, "Qcur", il);
-
-                cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf,
-                        model.layers[il].wo, NULL,
-                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
-            }
-
-            if (il == n_layer - 1) {
-                // skip computing output for unused tokens
-                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
-                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // feed-forward network
-            {
-                cur = llm_build_norm(ctx0, ffn_inp, hparams,
-                        model.layers[il].ffn_norm, NULL,
-                        LLM_NORM_RMS, cb, il);
-                cb(cur, "ffn_norm", il);
-
-                cur = llm_build_ffn(ctx0, cur,
-                        model.layers[il].ffn_up,   NULL,
-                        model.layers[il].ffn_gate, NULL,
-                        model.layers[il].ffn_down, NULL,
-                        NULL,
-                        LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
-                cb(cur, "ffn_out", il);
-            }
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = llm_build_norm(ctx0, cur, hparams,
-                model.output_norm, NULL,
-                LLM_NORM_RMS, cb, -1);
-        cb(cur, "result_norm", -1);
-
-        // lm_head
+        // lm_head
         cur = ggml_mul_mat(ctx0, model.output, cur);
         cb(cur, "result_output", -1);
 
@@ -8277,16 +8473,16 @@ struct llm_build_context {
                 cb(Kcur, "Kcur", il);
                 cb(Vcur, "Vcur", il);
 
-                Qcur = ggml_rope_custom(
-                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens), inp_pos,
-                    n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
+                Qcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens), 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);
 
-                Kcur = ggml_rope_custom(
-                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos,
-                    n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
+                Kcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
                     ext_factor, attn_factor, beta_fast, beta_slow
                 );
                 cb(Kcur, "Kcur", il);
@@ -8338,6 +8534,11 @@ struct llm_build_context {
             // attention layer norm
             cur = llm_build_norm(ctx0, cur, hparams, model.layers[il].attn_out_norm, model.layers[il].attn_out_norm_b, LLM_NORM, cb, il);
 
+            if (model.layers[il].attn_norm_2 != nullptr) {
+                cur = ggml_add(ctx0, cur, inpL); // re-add the layer input
+                cur = llm_build_norm(ctx0, cur, hparams, model.layers[il].attn_norm_2, model.layers[il].attn_norm_2_b, LLM_NORM, cb, il);
+            }
+
             struct ggml_tensor * ffn_inp = cur;
             cb(ffn_inp, "ffn_inp", il);
 
@@ -8717,16 +8918,16 @@ struct llm_build_context {
                 }
 
 
-                Qcur = ggml_rope_custom(
-                    ctx0, Qcur, inp_pos,
-                    n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
+                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
                 );
                 cb(Qcur, "Qcur", il);
 
-                Kcur = ggml_rope_custom(
-                    ctx0, Kcur, inp_pos,
-                    n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
+                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(Kcur, "Kcur", il);
@@ -8837,14 +9038,14 @@ struct llm_build_context {
                 Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
 
                 // using mode = 2 for neox mode
-                Qcur = ggml_rope_custom(
-                    ctx0, Qcur, inp_pos, n_rot, rope_type, 0, n_orig_ctx,
+                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
                 );
                 cb(Qcur, "Qcur", il);
 
-                Kcur = ggml_rope_custom(
-                    ctx0, Kcur, inp_pos, n_rot, rope_type, 0, n_orig_ctx,
+                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(Kcur, "Kcur", il);
@@ -8948,16 +9149,16 @@ struct llm_build_context {
                 Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
                 cb(Vcur, "Vcur", il);
 
-                Qcur = ggml_rope_custom(
-                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens), inp_pos,
-                    n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
+                Qcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens), 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);
 
-                Kcur = ggml_rope_custom(
-                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos,
-                    n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
+                Kcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
                     ext_factor, attn_factor, beta_fast, beta_slow
                 );
                 cb(Kcur, "Kcur", il);
@@ -9062,16 +9263,16 @@ struct llm_build_context {
                 Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
                 cb(Vcur, "Vcur", il);
 
-                Qcur = ggml_rope_custom(
-                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos,
-                    n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
+                Qcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), 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);
 
-                Kcur = ggml_rope_custom(
-                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos,
-                    n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
+                Kcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
                     ext_factor, attn_factor, beta_fast, beta_slow
                 );
                 cb(Kcur, "Kcur", il);
@@ -9106,6 +9307,7 @@ struct llm_build_context {
                         model.layers[il].ffn_down_exps,
                         n_expert, n_expert_used,
                         LLM_FFN_SILU, false,
+                        false, 0.0,
                         cb, il);
             cb(cur, "ffn_moe_out", il);
 
@@ -9214,8 +9416,8 @@ struct llm_build_context {
                 Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
                 Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
 
-                Qcur = ggml_rope_custom(
-                    ctx0, Qcur, inp_pos, n_rot, rope_type, 0, n_orig_ctx,
+                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
                 );
                 cb(Qcur, "Qcur", il);
@@ -9225,8 +9427,8 @@ struct llm_build_context {
                 Qcur = ggml_scale(ctx0, Qcur, 1.0f/sqrtf(float(n_embd_head)));
                 cb(Qcur, "Qcur", il);
 
-                Kcur = ggml_rope_custom(
-                    ctx0, Kcur, inp_pos, n_rot, rope_type, 0, n_orig_ctx,
+                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(Kcur, "Kcur", il);
@@ -9302,6 +9504,9 @@ struct llm_build_context {
 
             // self-attention
             {
+                // rope freq factors for 128k context
+                struct ggml_tensor * rope_factors = build_rope_factors(il);
+
                 struct ggml_tensor* attn_norm_output = llm_build_norm(ctx0, inpL, hparams,
                     model.layers[il].attn_norm,
                     NULL,
@@ -9333,8 +9538,8 @@ struct llm_build_context {
                 Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
                 Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
 
-                Qcur = ggml_rope_custom(
-                    ctx0, Qcur, inp_pos, n_rot, rope_type, 0, n_orig_ctx,
+                Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig,
                     freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow
                 );
                 cb(Qcur, "Qcur", il);
@@ -9342,8 +9547,8 @@ struct llm_build_context {
                 Qcur = ggml_scale(ctx0, Qcur, 1.0f / sqrtf(float(n_embd_head)));
                 cb(Qcur, "Qcur", il);
 
-                Kcur = ggml_rope_custom(
-                    ctx0, Kcur, inp_pos, n_rot, rope_type, 0, n_orig_ctx,
+                Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig,
                     freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow
                 );
                 cb(Kcur, "Kcur", il);
@@ -9449,15 +9654,15 @@ struct llm_build_context {
                 struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv, cur);
                 cb(Vcur, "Vcur", il);
 
-                Qcur = ggml_rope_custom(
-                        ctx0, ggml_reshape_3d(ctx0, Qcur, n_rot, n_head,    n_tokens), inp_pos,
-                        n_embd_head, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
+                Qcur = ggml_rope_ext(
+                        ctx0, ggml_reshape_3d(ctx0, Qcur, n_rot, n_head,    n_tokens), inp_pos, nullptr,
+                        n_embd_head, rope_type, n_ctx_orig, freq_base, freq_scale,
                         ext_factor, attn_factor, beta_fast, beta_slow);
                 cb(Qcur, "Qcur", il);
 
-                Kcur = ggml_rope_custom(
-                        ctx0, ggml_reshape_3d(ctx0, Kcur, n_rot, n_head_kv, n_tokens), inp_pos,
-                        n_embd_head, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
+                Kcur = ggml_rope_ext(
+                        ctx0, ggml_reshape_3d(ctx0, Kcur, n_rot, n_head_kv, n_tokens), inp_pos, nullptr,
+                        n_embd_head, rope_type, n_ctx_orig, freq_base, freq_scale,
                         ext_factor, attn_factor, beta_fast, beta_slow);
                 cb(Kcur, "Kcur", il);
 
@@ -9657,16 +9862,16 @@ struct llm_build_context {
                 cb(tmpk, "tmpk", il);
                 cb(Vcur, "Vcur", il);
 
-                struct ggml_tensor * Qcur = ggml_rope_custom(
-                    ctx0, ggml_reshape_3d(ctx0, tmpq, n_embd_head, n_head,    n_tokens), inp_pos,
-                    n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
+                struct ggml_tensor * Qcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, tmpq, n_embd_head, n_head,    n_tokens), 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);
 
-                struct ggml_tensor * Kcur = ggml_rope_custom(
-                    ctx0, ggml_reshape_3d(ctx0, tmpk, n_embd_head, n_head_kv, n_tokens), inp_pos,
-                    n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
+                struct ggml_tensor * Kcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, tmpk, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
                     ext_factor, attn_factor, beta_fast, beta_slow
                 );
                 cb(Kcur, "Kcur", il);
@@ -9773,16 +9978,16 @@ struct llm_build_context {
                 //     cb(Vcur, "Vcur", il);
                 // }
 
-                Qcur = ggml_rope_custom(
-                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens), inp_pos,
-                    n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
+                Qcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens), 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);
 
-                Kcur = ggml_rope_custom(
-                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos,
-                    n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
+                Kcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
                     ext_factor, attn_factor, beta_fast, beta_slow
                 );
                 cb(Kcur, "Kcur", il);
@@ -9890,16 +10095,16 @@ struct llm_build_context {
                     cb(Vcur, "Vcur", il);
                 }
 
-                Qcur = ggml_rope_custom(
-                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens), inp_pos,
-                    n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
+                Qcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens), 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);
 
-                Kcur = ggml_rope_custom(
-                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos,
-                    n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
+                Kcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
                     ext_factor, attn_factor, beta_fast, beta_slow
                 );
                 cb(Kcur, "Kcur", il);
@@ -10020,16 +10225,16 @@ struct llm_build_context {
                     cb(Vcur, "Vcur", il);
                 }
 
-                Qcur = ggml_rope_custom(
-                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens), inp_pos,
-                    n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
+                Qcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens), 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);
 
-                Kcur = ggml_rope_custom(
-                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos,
-                    n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
+                Kcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
                     ext_factor, attn_factor, beta_fast, beta_slow
                 );
                 cb(Kcur, "Kcur", il);
@@ -10094,7 +10299,7 @@ struct llm_build_context {
         cb(cur, "lmhead_scaling", -1);
 
         // lm_head
-        cur = ggml_mul_mat(ctx0, model.tok_embd, cur);
+        cur = ggml_mul_mat(ctx0, model.output, cur);
         cb(cur, "result_output", -1);
 
         ggml_build_forward_expand(gf, cur);
@@ -10140,18 +10345,18 @@ struct llm_build_context {
                 struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv, cur);
                 cb(Vcur, "Vcur", il);
 
-                Qcur = ggml_rope_custom(
-                        ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head_k, n_head,    n_tokens), inp_pos,
-                        n_embd_head_k, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
+                Qcur = ggml_rope_ext(
+                        ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head_k, n_head,    n_tokens), inp_pos, nullptr,
+                        n_embd_head_k, rope_type, n_ctx_orig, freq_base, freq_scale,
                         ext_factor, attn_factor, beta_fast, beta_slow);
                 cb(Qcur, "Qcur", il);
 
                 Qcur = ggml_scale(ctx0, Qcur, 1.0f / sqrtf(float(n_embd_head_k)));
                 cb(Qcur, "Qcur_scaled", il);
 
-                Kcur = ggml_rope_custom(
-                        ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head_k, n_head_kv, n_tokens), inp_pos,
-                        n_embd_head_k, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
+                Kcur = ggml_rope_ext(
+                        ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head_k, n_head_kv, n_tokens), inp_pos, nullptr,
+                        n_embd_head_k, rope_type, n_ctx_orig, freq_base, freq_scale,
                         ext_factor, attn_factor, beta_fast, beta_slow);
                 cb(Kcur, "Kcur", il);
 
@@ -10260,16 +10465,16 @@ struct llm_build_context {
                     cb(Vcur, "Vcur", il);
                 }
 
-                Qcur = ggml_rope_custom(
-                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos,
-                    n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
+                Qcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), 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);
 
-                Kcur = ggml_rope_custom(
-                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos,
-                    n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
+                Kcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
                     ext_factor, attn_factor, beta_fast, beta_slow
                 );
                 cb(Kcur, "Kcur", il);
@@ -10550,16 +10755,16 @@ struct llm_build_context {
                     cb(Kcur, "Kcur", il);
                 }
 
-                Qcur = ggml_rope_custom(
-                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos,
-                    n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
+                Qcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), 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);
 
-                Kcur = ggml_rope_custom(
-                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos,
-                    n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
+                Kcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
                     ext_factor, attn_factor, beta_fast, beta_slow
                 );
                 cb(Kcur, "Kcur", il);
@@ -10681,16 +10886,16 @@ struct llm_build_context {
                     cb(Vcur, "Vcur", il);
                 }
 
-                Qcur = ggml_rope_custom(
-                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos,
-                    n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
+                Qcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), 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);
 
-                Kcur = ggml_rope_custom(
-                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos,
-                    n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
+                Kcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
                     ext_factor, attn_factor, beta_fast, beta_slow
                 );
                 cb(Kcur, "Kcur", il);
@@ -10753,6 +10958,508 @@ struct llm_build_context {
 
         return gf;
     }
+
+    struct ggml_cgraph * build_gptneox() {
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+
+        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);
+
+        struct ggml_tensor * cur;
+        struct ggml_tensor * inpL;
+
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+
+        // inp_pos - contains the positions
+        struct ggml_tensor * inp_pos = build_inp_pos();
+
+        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
+
+        for (int il = 0; il < n_layer; ++il) {
+            cur = llm_build_norm(ctx0, inpL, hparams,
+                    model.layers[il].attn_norm,
+                    model.layers[il].attn_norm_b,
+                    LLM_NORM, cb, il);
+            cb(cur, "attn_norm", il);
+
+            // self-attention
+            {
+                cur = ggml_mul_mat(ctx0, model.layers[il].wqkv, cur);
+                cb(cur, "wqkv", il);
+
+                cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
+                cb(cur, "bqkv", il);
+
+                struct ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd,     n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd)));
+                struct ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd)));
+                struct ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)));
+
+                cb(Qcur, "Qcur", il);
+                cb(Kcur, "Kcur", il);
+                cb(Vcur, "Vcur", il);
+
+                Qcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), 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);
+
+                Kcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Kcur, "Kcur", il);
+
+                cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf,
+                        model.layers[il].wo, model.layers[il].bo,
+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+            }
+
+            if (il == n_layer - 1) {
+                // skip computing output for unused tokens
+                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
+                cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
+                inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
+            }
+
+            // ffn
+            if (hparams.use_par_res) {
+                // attention and ffn are computed in parallel
+                // x = x + attn(ln1(x)) + ffn(ln2(x))
+
+                struct ggml_tensor * attn_out = cur;
+
+                cur = llm_build_norm(ctx0, inpL, hparams,
+                        model.layers[il].ffn_norm,
+                        model.layers[il].ffn_norm_b,
+                        LLM_NORM, cb, il);
+                cb(cur, "ffn_norm", il);
+
+                cur = llm_build_ffn(ctx0, cur,
+                        model.layers[il].ffn_up,   model.layers[il].ffn_up_b,
+                        NULL,                      NULL,
+                        model.layers[il].ffn_down, model.layers[il].ffn_down_b,
+                        NULL,
+                        LLM_FFN_GELU, LLM_FFN_SEQ, cb, il);
+                cb(cur, "ffn_out", il);
+
+                cur = ggml_add(ctx0, cur, inpL);
+                cb(cur, "ffn_out", il);
+
+                inpL = ggml_add(ctx0, cur, attn_out);
+                cb(inpL, "l_out", il);
+            } else {
+                // attention and ffn are computed sequentially
+                // x = x + attn(ln1(x))
+                // x = x + ffn(ln2(x))
+
+                struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
+                cb(ffn_inp, "ffn_inp", il);
+
+                cur = llm_build_norm(ctx0, ffn_inp, hparams,
+                        model.layers[il].ffn_norm,
+                        model.layers[il].ffn_norm_b,
+                        LLM_NORM, cb, il);
+                cb(cur, "ffn_norm", il);
+
+                cur = llm_build_ffn(ctx0, cur,
+                        model.layers[il].ffn_up,   model.layers[il].ffn_up_b,
+                        NULL,                      NULL,
+                        model.layers[il].ffn_down, model.layers[il].ffn_down_b,
+                        NULL,
+                        LLM_FFN_GELU, LLM_FFN_SEQ, cb, il);
+                cb(cur, "ffn_out", il);
+
+                inpL = ggml_add(ctx0, cur, ffn_inp);
+                cb(inpL, "l_out", il);
+            }
+        }
+
+        cur = llm_build_norm(ctx0, inpL, hparams,
+                model.output_norm,
+                model.output_norm_b,
+                LLM_NORM, cb, -1);
+        cb(cur, "result_norm", -1);
+
+        cur = ggml_mul_mat(ctx0, model.output, cur);
+        cb(cur, "result_output", -1);
+
+        ggml_build_forward_expand(gf, cur);
+
+        return gf;
+    }
+
+    struct ggml_cgraph * build_arctic() {
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+
+        // mutable variable, needed during the last layer of the computation to skip unused tokens
+        int32_t n_tokens = this->n_tokens;
+
+        const int64_t n_embd_head = hparams.n_embd_head_v;
+        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+        GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+        struct ggml_tensor * cur;
+        struct ggml_tensor * inpL;
+
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+
+        // inp_pos - contains the positions
+        struct ggml_tensor * inp_pos = build_inp_pos();
+
+        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
+
+        for (int il = 0; il < n_layer; ++il) {
+            struct ggml_tensor * inpSA = inpL;
+
+            // norm
+            cur = llm_build_norm(ctx0, inpL, hparams,
+                    model.layers[il].attn_norm, NULL,
+                    LLM_NORM_RMS, cb, il);
+            cb(cur, "attn_norm", il);
+
+            // self-attention
+            {
+                // compute Q and K and RoPE them
+                struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].wq, cur);
+                cb(Qcur, "Qcur", il);
+
+                struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, model.layers[il].wk, cur);
+                cb(Kcur, "Kcur", il);
+
+                struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv, cur);
+                cb(Vcur, "Vcur", il);
+
+                Qcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), 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);
+
+                Kcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Kcur, "Kcur", il);
+
+                cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf,
+                        model.layers[il].wo, NULL,
+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+            }
+
+            if (il == n_layer - 1) {
+                // skip computing output for unused tokens
+                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
+                n_tokens = n_outputs;
+                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+            }
+
+            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+            cb(ffn_inp, "ffn_inp", il);
+
+            // feed-forward network
+            cur = llm_build_norm(ctx0, ffn_inp, hparams,
+                    model.layers[il].ffn_norm, NULL,
+                    LLM_NORM_RMS, cb, il);
+            cb(cur, "ffn_norm", il);
+
+            cur = llm_build_ffn(ctx0, cur,
+                    model.layers[il].ffn_up,   NULL,
+                    model.layers[il].ffn_gate, NULL,
+                    model.layers[il].ffn_down, NULL,
+                    NULL,
+                    LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
+            cb(cur, "ffn_out", il);
+
+            struct ggml_tensor * ffn_out = ggml_add(ctx0, cur, ffn_inp);
+            cb(ffn_out, "ffn_out", il);
+
+            // MoE
+            cur = llm_build_norm(ctx0, inpSA, hparams,
+                    model.layers[il].ffn_norm_exps, NULL,
+                    LLM_NORM_RMS, cb, il);
+            cb(cur, "ffn_norm_exps", il);
+
+            cur = llm_build_moe_ffn(ctx0, 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,
+                    n_expert, n_expert_used,
+                    LLM_FFN_SILU, true,
+                    false, 0.0,
+                    cb, il);
+            cb(cur, "ffn_moe_out", il);
+
+            cur = ggml_add(ctx0, cur, ffn_out);
+            cb(cur, "ffn_out", il);
+
+            ggml_tensor * layer_dir = lctx.cvec.tensor_for(il);
+            if (layer_dir != nullptr) {
+                cur = ggml_add(ctx0, cur, layer_dir);
+            }
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+
+        cur = inpL;
+
+        cur = llm_build_norm(ctx0, cur, hparams,
+                model.output_norm, NULL,
+                LLM_NORM_RMS, cb, -1);
+        cb(cur, "result_norm", -1);
+
+        // lm_head
+        cur = ggml_mul_mat(ctx0, model.output, cur);
+        cb(cur, "result_output", -1);
+
+        ggml_build_forward_expand(gf, cur);
+
+        return gf;
+    }
+
+    struct ggml_cgraph * build_deepseek2() {
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+
+        // mutable variable, needed during the last layer of the computation to skip unused tokens
+        int32_t n_tokens = this->n_tokens;
+
+        bool is_lite = (hparams.n_layer == 27);
+
+        // We have to pre-scale kq_scale and attn_factor to make the YaRN RoPE work correctly.
+        // See https://github.com/ggerganov/llama.cpp/discussions/7416 for detailed explanation.
+        const float mscale = attn_factor * (1.0f + hparams.rope_yarn_log_mul * logf(1.0f / freq_scale));
+        const float kq_scale = 1.0f*mscale*mscale/sqrtf(float(hparams.n_embd_head_k));
+        const float attn_factor_scaled = 1.0f / (1.0f + 0.1f * logf(1.0f / freq_scale));
+
+        const uint32_t n_embd_head_qk_rope = hparams.n_rot;
+        const uint32_t n_embd_head_qk_nope = hparams.n_embd_head_k - hparams.n_rot;
+        const uint32_t kv_lora_rank = hparams.n_lora_kv;
+
+        struct ggml_tensor * cur;
+        struct ggml_tensor * inpL;
+
+        // {n_embd, n_tokens}
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+
+        // inp_pos - contains the positions
+        struct ggml_tensor * inp_pos = build_inp_pos();
+
+        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
+
+        for (int il = 0; il < n_layer; ++il) {
+            struct ggml_tensor * inpSA = inpL;
+
+            // norm
+            cur = llm_build_norm(ctx0, inpL, hparams,
+                    model.layers[il].attn_norm, NULL,
+                    LLM_NORM_RMS, cb, il);
+            cb(cur, "attn_norm", il);
+
+            // self_attention
+            {
+                struct ggml_tensor * q = NULL;
+                if (!is_lite) {
+                    // {n_embd, q_lora_rank} * {n_embd, n_tokens} -> {q_lora_rank, n_tokens}
+                    q = ggml_mul_mat(ctx0, model.layers[il].wq_a, cur);
+                    cb(q, "q", il);
+
+                    q = llm_build_norm(ctx0, q, hparams,
+                            model.layers[il].attn_q_a_norm, NULL,
+                            LLM_NORM_RMS, cb, il);
+                    cb(q, "q", il);
+
+                    // {q_lora_rank, n_head * hparams.n_embd_head_k} * {q_lora_rank, n_tokens} -> {n_head * hparams.n_embd_head_k, n_tokens}
+                    q = ggml_mul_mat(ctx0, model.layers[il].wq_b, q);
+                    cb(q, "q", il);
+                } else {
+                    q = ggml_mul_mat(ctx0, model.layers[il].wq, cur);
+                    cb(q, "q", il);
+                }
+
+                // split into {n_head * n_embd_head_qk_nope, n_tokens}
+                struct ggml_tensor * q_nope = ggml_view_3d(ctx0, q, n_embd_head_qk_nope, n_head, n_tokens,
+                        ggml_row_size(q->type, hparams.n_embd_head_k),
+                        ggml_row_size(q->type, hparams.n_embd_head_k * n_head),
+                        0);
+                cb(q_nope, "q_nope", il);
+
+                // and {n_head * n_embd_head_qk_rope, n_tokens}
+                struct ggml_tensor * q_pe = ggml_view_3d(ctx0, q, n_embd_head_qk_rope, n_head, n_tokens,
+                        ggml_row_size(q->type, hparams.n_embd_head_k),
+                        ggml_row_size(q->type, hparams.n_embd_head_k * n_head),
+                        ggml_row_size(q->type, n_embd_head_qk_nope));
+                cb(q_pe, "q_pe", il);
+
+                // {n_embd, kv_lora_rank + n_embd_head_qk_rope} * {n_embd, n_tokens} -> {kv_lora_rank + n_embd_head_qk_rope, n_tokens}
+                struct ggml_tensor * kv_pe_compresseed = ggml_mul_mat(ctx0, model.layers[il].wkv_a_mqa, cur);
+                cb(kv_pe_compresseed, "kv_pe_compresseed", il);
+
+                // split into {kv_lora_rank, n_tokens}
+                struct ggml_tensor * kv_compressed = ggml_view_2d(ctx0, kv_pe_compresseed, kv_lora_rank, n_tokens,
+                        kv_pe_compresseed->nb[1],
+                        0);
+                cb(kv_compressed, "kv_compressed", il);
+
+                // and {n_embd_head_qk_rope, n_tokens}
+                struct ggml_tensor * k_pe = ggml_view_3d(ctx0, kv_pe_compresseed, n_embd_head_qk_rope, 1, n_tokens,
+                        kv_pe_compresseed->nb[1],
+                        kv_pe_compresseed->nb[1],
+                        ggml_row_size(kv_pe_compresseed->type, kv_lora_rank));
+                cb(k_pe, "k_pe", il);
+
+                kv_compressed = ggml_cont(ctx0, kv_compressed); // TODO: the CUDA backend does not support non-contiguous norm
+                kv_compressed = llm_build_norm(ctx0, kv_compressed, hparams,
+                        model.layers[il].attn_kv_a_norm, NULL,
+                        LLM_NORM_RMS, cb, il);
+                cb(kv_compressed, "kv_compressed", il);
+
+                // {kv_lora_rank, n_head * (n_embd_head_qk_nope + n_embd_head_v)} * {kv_lora_rank, n_tokens} -> {n_head * (n_embd_head_qk_nope + n_embd_head_v), n_tokens}
+                struct ggml_tensor * kv = ggml_mul_mat(ctx0, model.layers[il].wkv_b, kv_compressed);
+                cb(kv, "kv", il);
+
+                // split into {n_head * n_embd_head_qk_nope, n_tokens}
+                struct ggml_tensor * k_nope = ggml_view_3d(ctx0, kv, n_embd_head_qk_nope, n_head, n_tokens,
+                        ggml_row_size(kv->type, n_embd_head_qk_nope + hparams.n_embd_head_v),
+                        ggml_row_size(kv->type, n_head * (n_embd_head_qk_nope + hparams.n_embd_head_v)),
+                        0);
+                cb(k_nope, "k_nope", il);
+
+                // and {n_head * n_embd_head_v, n_tokens}
+                struct ggml_tensor * v_states = ggml_view_3d(ctx0, kv, hparams.n_embd_head_v, n_head, n_tokens,
+                        ggml_row_size(kv->type, (n_embd_head_qk_nope + hparams.n_embd_head_v)),
+                        ggml_row_size(kv->type, (n_embd_head_qk_nope + hparams.n_embd_head_v)*n_head),
+                        ggml_row_size(kv->type, (n_embd_head_qk_nope)));
+                cb(v_states, "v_states", il);
+
+                v_states = ggml_cont(ctx0, v_states);
+                cb(v_states, "v_states", il);
+
+                v_states = ggml_view_2d(ctx0, v_states, hparams.n_embd_head_v * n_head, n_tokens,
+                    ggml_row_size(kv->type, hparams.n_embd_head_v * n_head),
+                    0);
+                cb(v_states, "v_states", il);
+
+                q_pe = ggml_cont(ctx0, q_pe); // TODO: the CUDA backend does not support non-contiguous RoPE
+                q_pe = ggml_rope_ext(
+                    ctx0, q_pe, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor_scaled, beta_fast, beta_slow
+                );
+                cb(q_pe, "q_pe", il);
+
+                // shared RoPE key
+                k_pe = ggml_cont(ctx0, k_pe); // TODO: the CUDA backend does not support non-contiguous RoPE
+                k_pe = ggml_rope_ext(
+                    ctx0, k_pe, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor_scaled, beta_fast, beta_slow
+                );
+                cb(k_pe, "k_pe", il);
+
+                struct ggml_tensor * q_states = ggml_concat(ctx0, q_nope, q_pe, 0);
+                cb(q_states, "q_states", il);
+
+                struct ggml_tensor * k_states = ggml_concat(ctx0, k_nope, ggml_repeat(ctx0, k_pe, q_pe), 0);
+                cb(k_states, "k_states", il);
+
+                cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf,
+                        model.layers[il].wo, NULL,
+                        k_states, v_states, q_states, KQ_mask, n_tokens, kv_head, n_kv, kq_scale, cb, il);
+            }
+
+            if (il == n_layer - 1) {
+                // skip computing output for unused tokens
+                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
+                n_tokens = n_outputs;
+                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+            }
+
+            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+            cb(ffn_inp, "ffn_inp", il);
+
+            if ((uint32_t) il < hparams.n_layer_dense_lead) {
+                cur = llm_build_norm(ctx0, ffn_inp, hparams,
+                        model.layers[il].ffn_norm, NULL,
+                        LLM_NORM_RMS, cb, il);
+                cb(cur, "ffn_norm", il);
+
+                cur = llm_build_ffn(ctx0, cur,
+                        model.layers[il].ffn_up,   NULL,
+                        model.layers[il].ffn_gate, NULL,
+                        model.layers[il].ffn_down, NULL,
+                        NULL,
+                        LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
+                cb(cur, "ffn_out", il);
+            } else {
+                // MoE branch
+                cur = llm_build_norm(ctx0, ffn_inp, hparams,
+                        model.layers[il].ffn_norm, NULL,
+                        LLM_NORM_RMS, cb, il);
+                cb(cur, "ffn_norm", il);
+
+                ggml_tensor * moe_out =
+                        llm_build_moe_ffn(ctx0, 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,
+                            n_expert, n_expert_used,
+                            LLM_FFN_SILU, false,
+                            true, hparams.expert_weights_scale,
+                            cb, il);
+                cb(moe_out, "ffn_moe_out", il);
+
+                // FFN shared expert
+                {
+                    ggml_tensor * ffn_shexp = llm_build_ffn(ctx0, cur,
+                            model.layers[il].ffn_up_shexp,   NULL,
+                            model.layers[il].ffn_gate_shexp, NULL,
+                            model.layers[il].ffn_down_shexp, NULL,
+                            NULL,
+                            LLM_FFN_SILU, LLM_FFN_PAR, cb, 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, ffn_inp);
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+
+        cur = inpL;
+
+        cur = llm_build_norm(ctx0, cur, hparams,
+                model.output_norm, NULL,
+                LLM_NORM_RMS, cb, -1);
+        cb(cur, "result_norm", -1);
+
+        // lm_head
+        cur = ggml_mul_mat(ctx0, model.output, cur);
+        cb(cur, "result_output", -1);
+
+        ggml_build_forward_expand(gf, cur);
+
+        return gf;
+    }
+
 };
 
 static struct ggml_cgraph * llama_build_graph_defrag(llama_context & lctx, const std::vector<uint32_t> & ids) {
@@ -10833,7 +11540,8 @@ static struct ggml_cgraph * llama_build_graph(
         if (batch.n_tokens < 32 || full_offload) {
             if (il != -1 && strcmp(name, "norm") == 0) {
                 for (auto * backend : lctx.backends) {
-                    if (ggml_backend_buft_supports_backend(lctx.model.buft_layer[il].buft, backend)) {
+                    if (ggml_backend_supports_buft(backend, lctx.model.buft_layer[il].buft) &&
+                        (ggml_backend_supports_op(backend, cur) || ggml_backend_offload_op(backend, cur))) {
                         ggml_backend_sched_set_tensor_backend(lctx.sched, cur, backend);
                         break;
                     }
@@ -10869,10 +11577,6 @@ static struct ggml_cgraph * llama_build_graph(
             {
                 result = llm.build_starcoder();
             } break;
-        case LLM_ARCH_PERSIMMON:
-            {
-                result = llm.build_persimmon();
-            } break;
         case LLM_ARCH_REFACT:
             {
                 result = llm.build_refact();
@@ -10967,6 +11671,18 @@ static struct ggml_cgraph * llama_build_graph(
             {
                 result = llm.build_olmo();
             } break;
+        case LLM_ARCH_GPTNEOX:
+            {
+                result = llm.build_gptneox();
+            } break;
+        case LLM_ARCH_ARCTIC:
+            {
+                result = llm.build_arctic();
+            } break;
+        case LLM_ARCH_DEEPSEEK2:
+            {
+                result = llm.build_deepseek2();
+            } break;
         default:
             GGML_ASSERT(false);
     }
@@ -11312,11 +12028,6 @@ static void llama_graph_compute(
         llama_context & lctx,
           ggml_cgraph * gf,
                   int   n_threads) {
-#ifdef GGML_USE_MPI
-    const int64_t n_layer = lctx.model.hparams.n_layer;
-    ggml_mpi_graph_compute_pre(lctx.ctx_mpi, gf, n_layer);
-#endif
-
 #ifdef GGML_USE_METAL
     if (ggml_backend_is_metal(lctx.backend_metal)) {
         ggml_backend_metal_set_n_cb(lctx.backend_metal, n_threads);
@@ -11327,14 +12038,15 @@ static void llama_graph_compute(
         ggml_backend_cpu_set_n_threads(lctx.backend_cpu, n_threads);
         ggml_backend_cpu_set_abort_callback(lctx.backend_cpu, lctx.abort_callback, lctx.abort_callback_data);
     }
+#ifdef GGML_USE_BLAS
+    if (lctx.backend_blas != nullptr) {
+        ggml_backend_blas_set_n_threads(lctx.backend_blas, n_threads);
+    }
+#endif
 
     ggml_backend_sched_graph_compute_async(lctx.sched, gf);
 
     // fprintf(stderr, "splits: %d\n", ggml_backend_sched_get_n_splits(lctx.sched));
-
-#ifdef GGML_USE_MPI
-    ggml_mpi_graph_compute_post(lctx.ctx_mpi, gf, n_layer);
-#endif
 }
 
 // decode a batch of tokens by evaluating the transformer
@@ -11372,12 +12084,6 @@ static int llama_decode_internal(
     }
     lctx.n_queued_tokens += n_tokens_all;
 
-#ifdef GGML_USE_MPI
-    // TODO: needs fix after #3228
-    GGML_ASSERT(false && "not implemented");
-    //ggml_mpi_eval_init(lctx.ctx_mpi, &n_tokens, &n_past, &n_threads);
-#endif
-
     auto & kv_self = lctx.kv_self;
 
     const int64_t n_embd  = hparams.n_embd;
@@ -11508,7 +12214,8 @@ static int llama_decode_internal(
                 // a heuristic, to avoid attending the full cache if it is not yet utilized
                 // after enough generations, the benefit from this heuristic disappears
                 // if we start defragmenting the cache, the benefit from this will be more important
-                kv_self.n = std::min(kv_self.size, std::max(256u, GGML_PAD(llama_kv_cache_cell_max(kv_self), 256)));
+                const uint32_t pad = llama_kv_cache_get_padding(cparams);
+                kv_self.n = std::min(kv_self.size, std::max(pad, GGML_PAD(llama_kv_cache_cell_max(kv_self), pad)));
                 //kv_self.n = llama_kv_cache_cell_max(kv_self);
             }
         }
@@ -11558,17 +12265,6 @@ static int llama_decode_internal(
         }
         // LLAMA_LOG_INFO("graph build time: %.3f ms (%d nodes, %d leafs)\n", (ggml_time_us() - t_start_us)/1000.0, gf->n_nodes, gf->n_leafs);
 
-        // for big prompts, if BLAS is enabled, it is better to use only one thread
-        // otherwise, the threads are spin-lock waiting for the BLAS calls and are degrading the performance
-        // TODO: this is mostly important for Apple Silicon where CBLAS is still performing very well
-        //       we still need some threads to process all non-mul_mat ops, but not too much to avoid interfering
-        //       with the BLAS calls. need a better solution
-        // MoE Special Case: This logic applies when hparams.n_expert == 0, i.e. the model is NOT an MoE model. When an MoE is
-        //                   being processed then Accelerate/BLAS will not be involved, so capping would limit performance.
-        if (n_tokens >= 32 && hparams.n_expert == 0 && ggml_cpu_has_blas() && !ggml_cpu_has_gpublas()) {
-            n_threads = std::min(4, n_threads);
-        }
-
         ggml_backend_sched_alloc_graph(lctx.sched, gf);
 
         llama_set_inputs(lctx, u_batch);
@@ -11997,27 +12693,27 @@ static enum llama_vocab_type llama_vocab_get_type(const llama_vocab & vocab) {
 
 static bool llama_is_normal_token(const llama_vocab & vocab, llama_token id) {
     GGML_ASSERT(vocab.type != LLAMA_VOCAB_TYPE_NONE);
-    return vocab.id_to_token[id].type == LLAMA_TOKEN_TYPE_NORMAL;
+    return vocab.id_to_token[id].attr & LLAMA_TOKEN_ATTR_NORMAL;
 }
 
 static bool llama_is_unknown_token(const llama_vocab & vocab, llama_token id) {
     GGML_ASSERT(vocab.type != LLAMA_VOCAB_TYPE_NONE);
-    return vocab.id_to_token[id].type == LLAMA_TOKEN_TYPE_UNKNOWN;
+    return vocab.id_to_token[id].attr & LLAMA_TOKEN_ATTR_UNKNOWN;
 }
 
 static bool llama_is_control_token(const llama_vocab & vocab, llama_token id) {
     GGML_ASSERT(vocab.type != LLAMA_VOCAB_TYPE_NONE);
-    return vocab.id_to_token[id].type == LLAMA_TOKEN_TYPE_CONTROL;
+    return vocab.id_to_token[id].attr & LLAMA_TOKEN_ATTR_CONTROL;
 }
 
 static bool llama_is_byte_token(const llama_vocab & vocab, llama_token id) {
     GGML_ASSERT(vocab.type != LLAMA_VOCAB_TYPE_NONE);
-    return vocab.id_to_token[id].type == LLAMA_TOKEN_TYPE_BYTE;
+    return vocab.id_to_token[id].attr & LLAMA_TOKEN_ATTR_BYTE;
 }
 
 static bool llama_is_user_defined_token(const llama_vocab& vocab, llama_token id) {
     GGML_ASSERT(vocab.type != LLAMA_VOCAB_TYPE_NONE);
-    return vocab.id_to_token[id].type == LLAMA_TOKEN_TYPE_USER_DEFINED;
+    return vocab.id_to_token[id].attr & LLAMA_TOKEN_ATTR_USER_DEFINED;
 }
 
 static uint8_t llama_token_to_byte(const llama_vocab& vocab, llama_token id) {
@@ -12270,6 +12966,7 @@ struct llm_tokenizer_bpe {
                         });
                         break;
                     case LLAMA_VOCAB_PRE_TYPE_DBRX:
+                    case LLAMA_VOCAB_PRE_TYPE_SMAUG:
                         word_collection = unicode_regex_split(text, {
                             // same as llama3
                             "(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
@@ -12326,6 +13023,7 @@ struct llm_tokenizer_bpe {
                             "'s|'t|'re|'ve|'m|'ll|'d| ?\\p{L}+| ?\\p{N}+| ?[^\\s\\p{L}\\p{N}]+|\\s+(?!\\S)",
                         });
                         break;
+                    case LLAMA_VOCAB_PRE_TYPE_STABLELM2:
                     case LLAMA_VOCAB_PRE_TYPE_QWEN2:
                         word_collection = unicode_regex_split(text, {
                             // original regex from tokenizer.json
@@ -12333,6 +13031,11 @@ struct llm_tokenizer_bpe {
                             "(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
                         });
                         break;
+                    case LLAMA_VOCAB_PRE_TYPE_PORO:
+                        word_collection = unicode_regex_split(text, {
+                            " ?[^(\\s|.,!?…。，、।۔،)]+",
+                        });
+                        break;
                     default:
                         // default regex for BPE tokenization pre-processing
                         word_collection = unicode_regex_split(text, {
@@ -12491,7 +13194,7 @@ struct llm_tokenizer_wpm {
     llm_tokenizer_wpm(const llama_vocab & vocab): vocab(vocab) {}
 
     void tokenize(const std::string & text, std::vector<llama_vocab::id> & output) {
-        auto * token_map = &vocab.token_to_id;
+        const auto & token_map = vocab.token_to_id;
 
         // normalize and split by whitespace
         std::vector<std::string> words = preprocess(text);
@@ -12506,108 +13209,89 @@ struct llm_tokenizer_wpm {
             }
 
             // prepend phantom space
-            std::string word1 = "\xe2\x96\x81" + word;
-            int n = word1.size();
+            const std::string word1 = "\xe2\x96\x81" + word;
+            const int n = word1.size();
 
-            // we're at the start of a new word
-            int i = 0;
-            bool match_any = false;
+            const size_t current_tokens = output.size();
 
+            // we're at the start of a new word
             // move through character position in word
-            while (i < n) {
+            for (int i = 0; i < n; ++i) {
                 // loop through possible match length
                 bool match = false;
                 for (int j = n; j > i; j--) {
-                    auto it = token_map->find(word1.substr(i, j - i));
-                    if (it != token_map->end()) {
+                    auto it = token_map.find(word1.substr(i, j - i));
+                    if (it != token_map.end()) {
                         output.push_back(it->second);
                         match = true;
-                        match_any = true;
-                        i = j;
+                        i = j - 1;
                         break;
                     }
                 }
 
-                // must be an unknown character
-                if (!match) {
-                    i++;
+                if (!match) { // discard all
+                    output.resize(current_tokens);
+                    break;  // and discard next tokens
                 }
             }
 
             // we didn't find any matches for this word
-            if (!match_any) {
+            if (current_tokens == output.size()) {
                 output.push_back(vocab.special_unk_id);
             }
         }
     }
 
     std::vector<std::string> preprocess(const std::string & text) {
-        std::vector<uint32_t> cpts_nfd = unicode_cpts_normalize_nfd(unicode_cpts_from_utf8(text));
-
-        // strip accents, strip control, uniformize whitespace,
-        // to lowercase, pad chinese characters, pad punctuation
-        std::string new_str = "";
-        for (uint32_t code : cpts_nfd) {
-            int type = unicode_cpt_type(code);
-            if (type == CODEPOINT_TYPE_ACCENT_MARK || type == CODEPOINT_TYPE_CONTROL) {
+        const std::vector<uint32_t> cpts_nfd = unicode_cpts_normalize_nfd(unicode_cpts_from_utf8(text));
+        std::vector<std::string> words(1, "");
+
+        for (const char32_t cpt : cpts_nfd) {
+            const auto flags = unicode_cpt_flags(cpt);
+
+            if (flags.is_whitespace) {
+                if (words.back().size()) {  // finish previous word if any
+                    words.emplace_back();
+                }
                 continue;
             }
-            code = unicode_tolower(code);
-            if (type == CODEPOINT_TYPE_SEPARATOR) {
-                code = ' ';
-            }
-            std::string s = unicode_cpt_to_utf8(code);
-            if (type == CODEPOINT_TYPE_PUNCTUATION || is_ascii_punct(code) || is_chinese_char(code)) {
-                new_str += " ";
-                new_str += s;
-                new_str += " ";
-            } else {
-                new_str += s;
+
+            assert (!flags.is_separator);
+            if (cpt == 0 || cpt == 0xFFFD || flags.is_control) {
+                continue;
             }
-        }
 
-        // split by whitespace
-        uint64_t l = 0;
-        uint64_t r = 0;
-        std::vector<std::string> words;
-        while (r < new_str.size()) {
-            // if is whitespace
-            if (isspace(new_str[r], std::locale::classic())) {
-                if (r > l) words.push_back(new_str.substr(l, (r - l)));
-                l = r + 1;
-                r = l;
+            const std::string s = unicode_cpt_to_utf8(unicode_tolower(cpt));
+            if (flags.is_punctuation || ( cpt < 0x7F && flags.is_symbol ) || is_chinese_char(cpt)) {
+                if (words.back().size()) {  // finish previous word if any
+                    words.emplace_back();
+                }
+                words.back() = s;       // single char word
+                words.emplace_back();   // start a new word
             } else {
-                r += 1;
+                words.back() += s;  // append char to word
             }
         }
-        if (r > l) {
-            words.push_back(new_str.substr(l, (r - l)));
-        }
-        return words;
-    }
 
-    bool is_ascii_punct(uint32_t code) {
-        if (code > 0xFF) {
-            return false;
+        if (!words.back().size()) {
+            words.pop_back();
         }
-        auto c = char(static_cast<unsigned char>(code));
-        return ispunct(c, std::locale::classic());
+
+        return words;
     }
 
-    bool is_chinese_char(uint32_t cpt) {
-        if ((cpt >= 0x4E00  && cpt <= 0x9FFF)  ||
-            (cpt >= 0x3400  && cpt <= 0x4DBF)  ||
+    static bool is_chinese_char(uint32_t cpt) {
+        return
+            (cpt >= 0x04E00 && cpt <= 0x09FFF) ||
+            (cpt >= 0x03400 && cpt <= 0x04DBF) ||
             (cpt >= 0x20000 && cpt <= 0x2A6DF) ||
             (cpt >= 0x2A700 && cpt <= 0x2B73F) ||
             (cpt >= 0x2B740 && cpt <= 0x2B81F) ||
             (cpt >= 0x2B920 && cpt <= 0x2CEAF) || // this should be 0x2B820 but in hf rust code it is 0x2B920
-            (cpt >= 0xF900  && cpt <= 0xFAFF)  ||
-            (cpt >= 0x2F800 && cpt <= 0x2FA1F) ||
-            (cpt >= 0x3000  && cpt <= 0x303F)  ||
-            (cpt >= 0xFF00  && cpt <= 0xFFEF)) {
-            return true; // NOLINT
-        }
-        return false;
+            (cpt >= 0x0F900 && cpt <= 0x0FAFF) ||
+            (cpt >= 0x2F800 && cpt <= 0x2FA1F);
+            //(cpt >= 0x3000  && cpt <= 0x303F)  ||
+            //(cpt >= 0xFF00  && cpt <= 0xFFEF);
     }
 
     const llama_vocab & vocab;
@@ -12651,9 +13335,9 @@ struct fragment_buffer_variant {
 
 static void tokenizer_st_partition(const llama_vocab & vocab, std::forward_list<fragment_buffer_variant> & buffer) {
     // for each special token
-    for (const auto & st: vocab.special_tokens_cache) {
-        const auto & special_token = st.first;
-        const auto & special_id    = st.second;
+    for (const llama_vocab::id special_id : vocab.cache_special_tokens) {
+        const auto & data = vocab.id_to_token[special_id];
+        const auto & special_token = data.text;
 
         // for each text fragment
         std::forward_list<fragment_buffer_variant>::iterator it = buffer.begin();
@@ -12662,7 +13346,7 @@ static void tokenizer_st_partition(const llama_vocab & vocab, std::forward_list<
 
             // if a fragment is text ( not yet processed )
             if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_RAW_TEXT) {
-                auto * raw_text = &(fragment.raw_text);
+                auto & raw_text = fragment.raw_text;
 
                 auto raw_text_base_offset = fragment.offset;
                 auto raw_text_base_length = fragment.length;
@@ -12672,7 +13356,7 @@ static void tokenizer_st_partition(const llama_vocab & vocab, std::forward_list<
                     // find the first occurrence of a given special token in this fragment
                     //  passing offset argument only limit the "search area" but match coordinates
                     //  are still relative to the source full raw_text
-                    auto match = raw_text->find(special_token, raw_text_base_offset);
+                    auto match = raw_text.find(special_token, raw_text_base_offset);
 
                     // no occurrences found, stop processing this fragment for a given special token
                     if (match == std::string::npos) break;
@@ -12690,13 +13374,22 @@ static void tokenizer_st_partition(const llama_vocab & vocab, std::forward_list<
                     if (match > raw_text_base_offset) {
                         // left
                         const int64_t left_reminder_offset = raw_text_base_offset + 0;
-                        const int64_t left_reminder_length = match - raw_text_base_offset;
-                        buffer.emplace_after(it, (*raw_text), left_reminder_offset, left_reminder_length);
+                        int64_t left_reminder_length = match - raw_text_base_offset;
+
+                        if (data.attr & LLAMA_TOKEN_ATTR_LSTRIP) {
+                            while (left_reminder_length > 0 && isspace(raw_text[left_reminder_offset + left_reminder_length - 1])) {
+                                left_reminder_length--;
+                            }
+                        }
+
+                        if (left_reminder_length > 0) {
+                            buffer.emplace_after(it, raw_text, left_reminder_offset, left_reminder_length);
+                            it++;
+                        }
 
 #ifdef PRETOKENIZERDEBUG
                         LLAMA_LOG_WARN("FL: (%ld %ld) '%s'\n", left_reminder_offset, left_reminder_length, raw_text->substr(left_reminder_offset, left_reminder_length).c_str());
 #endif
-                        it++;
                     }
 
                     // special token
@@ -12705,16 +13398,25 @@ static void tokenizer_st_partition(const llama_vocab & vocab, std::forward_list<
 
                     // right
                     if (match + special_token.length() < raw_text_base_offset + raw_text_base_length) {
-                        const int64_t right_reminder_offset = match + special_token.length();
-                        const int64_t right_reminder_length = raw_text_base_length - ((match - raw_text_base_offset) + special_token.length());
-                        buffer.emplace_after(it, (*raw_text), right_reminder_offset, right_reminder_length);
+                        int64_t right_reminder_offset = match + special_token.length();
+                        int64_t right_reminder_length = raw_text_base_length - ((match - raw_text_base_offset) + special_token.length());
+
+                        if (data.attr & LLAMA_TOKEN_ATTR_RSTRIP) {
+                            while (right_reminder_length > 0 && isspace(raw_text[right_reminder_offset])) {
+                                right_reminder_offset++;
+                                right_reminder_length--;
+                            }
+                        }
+
+                        if (right_reminder_length > 0) {
+                            buffer.emplace_after(it, raw_text, right_reminder_offset, right_reminder_length);
+                            it++;
+                        }
 
 #ifdef PRETOKENIZERDEBUG
                         LLAMA_LOG_WARN("FR: (%ld %ld) '%s'\n", right_reminder_offset, right_reminder_length, raw_text->substr(right_reminder_offset, right_reminder_length).c_str());
 #endif
 
-                        it++;
-
                         if (source == 0) {
                             buffer.erase_after(buffer.before_begin());
                         } else {
@@ -12760,23 +13462,21 @@ static std::vector<llama_vocab::id> llama_tokenize_internal(const llama_vocab &
                 // tokenizer.encode('', add_special_tokens=True)  returns [1]
                 // tokenizer.encode('', add_special_tokens=False) returns []
 
+                bool is_prev_special = false;
+
                 if (add_special && vocab.special_add_bos != 0) {
                     GGML_ASSERT(vocab.special_bos_id != -1);
                     output.push_back(vocab.special_bos_id);
+                    is_prev_special = true;
                 }
 
                 for (const auto & fragment : fragment_buffer) {
                     if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_RAW_TEXT) {
-                        // without adding this leading whitespace, we do not get the same results as the original tokenizer
-
-                        // TODO: It's likely possible to get rid of this string copy entirely
-                        //  by modifying llm_tokenizer_x to operate with string offsets like pre-tokenizer
-                        //  and passing 'add space prefix' as bool argument
-                        //
                         auto raw_text = fragment.raw_text.substr(fragment.offset, fragment.length);
-                        if (&fragment == &fragment_buffer.front()) {
-                            if (vocab.add_space_prefix) {
-                                raw_text = " " + raw_text; // prefix with space if the first token is not special
+
+                        if (vocab.add_space_prefix) {
+                            if (!output.size() || is_prev_special) {  // prefix with space if first token
+                                raw_text = " " + raw_text;
                             }
                         }
 
@@ -12788,9 +13488,17 @@ static std::vector<llama_vocab::id> llama_tokenize_internal(const llama_vocab &
                         tokenizer.tokenize(raw_text, output);
                     } else { // if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_TOKEN)
                         output.push_back(fragment.token);
+                        is_prev_special = true;
                     }
                 }
 
+                if (add_special && vocab.special_add_bos != 0 && output.size() >= 2 && output[1] == vocab.special_bos_id) {
+                    LLAMA_LOG_WARN(
+                        "%s: Added a BOS token to the prompt as specified by the model but the prompt "
+                        "also starts with a BOS token. So now the final prompt starts with 2 BOS tokens. "
+                        "Are you sure this is what you want?\n", __FUNCTION__);
+                }
+
                 if (add_special && vocab.special_add_eos == 1) {
                     GGML_ASSERT(vocab.special_eos_id != -1);
                     output.push_back(vocab.special_eos_id);
@@ -12817,6 +13525,13 @@ static std::vector<llama_vocab::id> llama_tokenize_internal(const llama_vocab &
                     }
                 }
 
+                if (add_special && vocab.special_add_bos != 0 && output.size() >= 2 && output[1] == vocab.special_bos_id) {
+                    LLAMA_LOG_WARN(
+                        "%s: Added a BOS token to the prompt as specified by the model but the prompt "
+                        "also starts with a BOS token. So now the final prompt starts with 2 BOS tokens. "
+                        "Are you sure this is what you want?\n", __FUNCTION__);
+                }
+
                 if (add_special && vocab.special_add_eos == 1) {
                     GGML_ASSERT(vocab.special_add_eos != -1);
                     output.push_back(vocab.special_eos_id);
@@ -12936,7 +13651,7 @@ static std::pair<bool, const llama_grammar_element *> llama_grammar_match_char(
         const uint32_t                chr) {
 
     bool found            = false;
-    bool is_positive_char = pos->type == LLAMA_GRETYPE_CHAR;
+    bool is_positive_char = pos->type == LLAMA_GRETYPE_CHAR || pos->type == LLAMA_GRETYPE_CHAR_ANY;
 
     GGML_ASSERT(is_positive_char || pos->type == LLAMA_GRETYPE_CHAR_NOT); // NOLINT
 
@@ -12945,6 +13660,10 @@ static std::pair<bool, const llama_grammar_element *> llama_grammar_match_char(
             // inclusive range, e.g. [a-z]
             found = found || (pos->value <= chr && chr <= pos[1].value);
             pos += 2;
+        } else if (pos->type == LLAMA_GRETYPE_CHAR_ANY) {
+            // Any character matches "."
+            found = true;
+            pos += 1;
         } else {
             // exact char match, e.g. [a] or "a"
             found = found || pos->value == chr;
@@ -12962,7 +13681,7 @@ static bool llama_grammar_match_partial_char(
         const llama_grammar_element * pos,
         const llama_partial_utf8      partial_utf8) {
 
-    bool is_positive_char = pos->type == LLAMA_GRETYPE_CHAR;
+    bool is_positive_char = pos->type == LLAMA_GRETYPE_CHAR || pos->type == LLAMA_GRETYPE_CHAR_ANY;
     GGML_ASSERT(is_positive_char || pos->type == LLAMA_GRETYPE_CHAR_NOT);
 
     uint32_t partial_value = partial_utf8.value;
@@ -12992,6 +13711,9 @@ static bool llama_grammar_match_partial_char(
                 return is_positive_char;
             }
             pos += 2;
+        } else if (pos->type == LLAMA_GRETYPE_CHAR_ANY) {
+            // Any character matches "."
+            return true;
         } else {
             // exact char match, e.g. [a] or "a"
             if (low <= pos->value && pos->value <= high) {
@@ -13052,6 +13774,7 @@ static void llama_grammar_advance_stack(
         }
         case LLAMA_GRETYPE_CHAR:
         case LLAMA_GRETYPE_CHAR_NOT:
+        case LLAMA_GRETYPE_CHAR_ANY:
             if (std::find(new_stacks.begin(), new_stacks.end(), stack) == new_stacks.end()) {
                 // only add the stack if it's not a duplicate of one we already have
                 new_stacks.emplace_back(stack);
@@ -13174,6 +13897,58 @@ static std::vector<llama_grammar_candidate> llama_grammar_reject_candidates(
     return rejects;
 }
 
+static bool llama_grammar_detect_left_recursion(
+        const std::vector<std::vector<llama_grammar_element>> & rules,
+        size_t                                                  rule_index,
+        std::vector<bool>                                     * rules_visited,
+        std::vector<bool>                                     * rules_in_progress,
+        std::vector<bool>                                     * rules_may_be_empty) {
+    if ((*rules_in_progress)[rule_index]) {
+        return true;
+    }
+
+    (*rules_in_progress)[rule_index] = true;
+
+    const std::vector<llama_grammar_element> & rule = rules[rule_index];
+
+    // First check if the rule might produce the empty string. This could be done combined with the second
+    // step but it's more readable as two steps.
+    bool at_rule_start = true;
+    for (size_t i = 0; i < rule.size(); i++) {
+        if (llama_grammar_is_end_of_sequence(&rule[i])) {
+            if (at_rule_start) {
+                (*rules_may_be_empty)[rule_index] = true;
+                break;
+            }
+            at_rule_start = true;
+        } else {
+            at_rule_start = false;
+        }
+    }
+
+    // Second, recurse into leftmost nonterminals (or next-leftmost as long as the previous nonterminal may
+    // be empty)
+    bool recurse_into_nonterminal = true;
+    for (size_t i = 0; i < rule.size(); i++) {
+        if (rule[i].type == LLAMA_GRETYPE_RULE_REF && recurse_into_nonterminal) {
+            if (llama_grammar_detect_left_recursion(rules, (size_t)rule[i].value, rules_visited, rules_in_progress, rules_may_be_empty)) {
+                return true;
+            }
+            if (!((*rules_may_be_empty)[(size_t)rule[i].value])) {
+                recurse_into_nonterminal = false;
+            }
+        } else if (llama_grammar_is_end_of_sequence(&rule[i])) {
+            recurse_into_nonterminal = true;
+        } else {
+            recurse_into_nonterminal = false;
+        }
+    }
+
+    (*rules_in_progress)[rule_index] = false;
+    (*rules_visited)[rule_index] = true;
+    return false;
+}
+
 //
 // grammar - external
 //
@@ -13193,6 +13968,19 @@ struct llama_grammar * llama_grammar_init(
         vec_rules[i].push_back({LLAMA_GRETYPE_END, 0});
     }
 
+    // Check for left recursion
+    std::vector<bool> rules_visited(n_rules);
+    std::vector<bool> rules_in_progress(n_rules);
+    std::vector<bool> rules_may_be_empty(n_rules);
+    for (size_t i = 0; i < n_rules; i++) {
+        if (rules_visited[i]) {
+            continue;
+        }
+        if (llama_grammar_detect_left_recursion(vec_rules, i, &rules_visited, &rules_in_progress, &rules_may_be_empty)) {
+            throw std::runtime_error(format("unsupported grammar, left recursion detected for nonterminal at index %zu", i));
+        }
+    }
+
     // loop over alternates of start rule to build initial stacks
     std::vector<std::vector<const llama_grammar_element *>> stacks;
     pos = vec_rules[start_rule_index].data();
@@ -13215,6 +14003,9 @@ struct llama_grammar * llama_grammar_init(
         }
     } while (true);
 
+    // Important: vec_rules has to be moved here, not copied, because stacks contains
+    // pointers to elements of vec_rules. If vec_rules were copied into llama_grammar
+    // then the pointers would be invalidated when the local vec_rules goes out of scope.
     return new llama_grammar{ std::move(vec_rules), std::move(stacks), {} };
 }
 
@@ -13706,7 +14497,7 @@ void llama_sample_repetition_penalties(
 
 void llama_sample_grammar(struct llama_context * ctx, llama_token_data_array * candidates, const struct llama_grammar * grammar) {
     GGML_ASSERT(ctx);
-    const int64_t t_start_sample_us = ggml_time_us();
+    int64_t t_start_sample_us = ggml_time_us();
 
     bool allow_eog = false;
     for (const auto & stack : grammar->stacks) {
@@ -13718,12 +14509,13 @@ void llama_sample_grammar(struct llama_context * ctx, llama_token_data_array * c
 
     std::vector<std::pair<std::vector<uint32_t>, llama_partial_utf8>> candidates_decoded;
     candidates_decoded.reserve(candidates->size);
-    std::vector<llama_grammar_candidate>                              candidates_grammar;
+
+    std::vector<llama_grammar_candidate> candidates_grammar;
     candidates_grammar.reserve(candidates->size);
 
     for (size_t i = 0; i < candidates->size; ++i) {
-        const llama_token id    = candidates->data[i].id;
-        const std::string piece = llama_token_to_piece(ctx, id, false);
+        const llama_token id      = candidates->data[i].id;
+        const std::string & piece = ctx->model.vocab.cache_token_to_piece.at(id);
 
         if (llama_token_is_eog(&ctx->model, id)) {
             if (!allow_eog) {
@@ -13809,9 +14601,7 @@ llama_token llama_sample_token_mirostat(struct llama_context * ctx, llama_token_
 
     // Sample the next word X using top-k sampling
     llama_sample_top_k(nullptr, candidates, int(k), 1);
-    if (ctx) {
-        ctx->t_sample_us += ggml_time_us() - t_start_sample_us;
-    }
+    ctx->t_sample_us += ggml_time_us() - t_start_sample_us;
     llama_token X = llama_sample_token(ctx, candidates);
     t_start_sample_us = ggml_time_us();
 
@@ -13825,9 +14615,7 @@ llama_token llama_sample_token_mirostat(struct llama_context * ctx, llama_token_
     // Update mu using the learning rate and error
     *mu = *mu - eta * e;
 
-    if (ctx) {
-        ctx->t_sample_us += ggml_time_us() - t_start_sample_us;
-    }
+    ctx->t_sample_us += ggml_time_us() - t_start_sample_us;
     return X;
 }
 
@@ -13927,7 +14715,7 @@ void llama_grammar_accept_token(struct llama_context * ctx, struct llama_grammar
         GGML_ASSERT(false);
     }
 
-    const std::string piece = llama_token_to_piece(ctx, token, false);
+    const std::string & piece = ctx->model.vocab.cache_token_to_piece.at(token);
 
     // Note terminating 0 in decoded string
     const auto   decoded     = decode_utf8(piece, grammar->partial_utf8);
@@ -13943,260 +14731,6 @@ void llama_grammar_accept_token(struct llama_context * ctx, struct llama_grammar
     ctx->t_sample_us += ggml_time_us() - t_start_sample_us;
 }
 
-//
-// Beam search
-//
-
-struct llama_beam {
-    std::vector<llama_token> tokens;
-    float p;  // Cumulative beam probability (renormalized relative to all beams)
-    bool eob; // Initialize end-of-beam to false. Callback sets this to true.
-    // Sort beams by probability. In case of ties, prefer beams at eob.
-    bool operator<(const llama_beam & rhs) const {
-        return std::make_pair(p, eob) < std::make_pair(rhs.p, rhs.eob);
-    }
-    // Shift off first n tokens and discard them.
-    void shift_tokens(const size_t n) {
-        if (n) {
-            std::copy(tokens.begin() + n, tokens.end(), tokens.begin());
-            tokens.resize(tokens.size() - n);
-        }
-    }
-    llama_beam_view view() const { return {tokens.data(), tokens.size(), p, eob}; }
-};
-
-// A struct for calculating logit-related info.
-struct llama_logit_info {
-    const float * const logits;
-    const int n_vocab;
-    const float max_l;
-    const float normalizer;
-    struct sum_exp {
-        float max_l;
-        float operator()(float sum, float l) const { return sum + std::exp(l - max_l); }
-    };
-    llama_logit_info(llama_context * ctx)
-      : logits(llama_get_logits(ctx))
-      , n_vocab(llama_n_vocab(llama_get_model(ctx)))
-      , max_l(*std::max_element(logits, logits + n_vocab))
-      , normalizer(1.0f / std::accumulate(logits, logits + n_vocab, 0.0f, sum_exp{max_l}))
-      { }
-    llama_token_data get_token_data(const llama_token token_id) const {
-        constexpr auto p = std::numeric_limits<float>::quiet_NaN();  // never used
-        return {token_id, logits[token_id], p};
-    }
-    // Return top k token_data by logit.
-    std::vector<llama_token_data> top_k(size_t k) {
-        std::vector<llama_token_data> min_heap;  // min-heap by logit
-        const llama_token k_min = std::min(static_cast<llama_token>(k), n_vocab);
-        min_heap.reserve(k_min);
-        for (llama_token token_id = 0 ; token_id < k_min ; ++token_id) {
-            min_heap.push_back(get_token_data(token_id));
-        }
-        auto comp = [](const llama_token_data & a, const llama_token_data & b) { return a.logit > b.logit; };
-        std::make_heap(min_heap.begin(), min_heap.end(), comp);
-        for (llama_token token_id = k_min ; token_id < n_vocab ; ++token_id) {
-            if (min_heap.front().logit < logits[token_id]) {
-                std::pop_heap(min_heap.begin(), min_heap.end(), comp);
-                min_heap.back().id = token_id;
-                min_heap.back().logit = logits[token_id];
-                std::push_heap(min_heap.begin(), min_heap.end(), comp);
-            }
-        }
-        return min_heap;
-    }
-    float probability_from_logit(float logit) const {
-        return normalizer * std::exp(logit - max_l);
-    }
-};
-
-struct llama_beam_search_data {
-    llama_context * ctx;
-    size_t n_beams;
-    int n_past;
-    int n_predict;
-    std::vector<llama_beam> beams;
-    std::vector<llama_beam> next_beams;
-
-    // Re-calculated on each loop iteration
-    size_t common_prefix_length;
-
-    // Used to communicate to/from callback on beams state.
-    std::vector<llama_beam_view> beam_views;
-
-    llama_beam_search_data(llama_context * ctx, size_t n_beams, int n_past, int n_predict)
-      : ctx(ctx)
-      , n_beams(n_beams)
-      , n_past(n_past)
-      , n_predict(n_predict)
-      , beam_views(n_beams) {
-        beams.reserve(n_beams);
-        next_beams.reserve(n_beams);
-    }
-
-    // Collapse beams to a single beam given by index.
-    void collapse_beams(const size_t beam_idx) {
-        if (0u < beam_idx) {
-            std::swap(beams[0], beams[beam_idx]);
-        }
-        beams.resize(1);
-    }
-
-    // Min-heaps are used to efficiently collect the top-k elements (k=n_beams).
-    // The repetitive patterns below reflect the 2 stages of heaps:
-    //  * Gather elements until the vector is full, then call std::make_heap() on it.
-    //  * If the heap is full and a new element is found that should be included, pop the
-    //    least element to the back(), replace it with the new, then push it into the heap.
-    void fill_next_beams_by_top_probabilities(llama_beam & beam) {
-        // Min-heaps use a greater-than comparator.
-        const auto comp = [](const llama_beam & a, const llama_beam & b) { return a.p > b.p; };
-        if (beam.eob) {
-            // beam is at end-of-sentence, so just copy it to next_beams if its probability is high enough.
-            if (next_beams.size() < n_beams) {
-                next_beams.push_back(std::move(beam));
-                if (next_beams.size() == n_beams) {
-                    std::make_heap(next_beams.begin(), next_beams.end(), comp);
-                }
-            } else if (next_beams.front().p < beam.p) {
-                std::pop_heap(next_beams.begin(), next_beams.end(), comp);
-                next_beams.back() = std::move(beam);
-                std::push_heap(next_beams.begin(), next_beams.end(), comp);
-            }
-        } else {
-            // beam is not at end-of-sentence, so branch with next top_k tokens.
-            if (!beam.tokens.empty()) {
-                llama_decode(ctx, llama_batch_get_one(beam.tokens.data(), beam.tokens.size(), n_past, 0));
-            }
-            llama_logit_info logit_info(ctx);
-            std::vector<llama_token_data> next_tokens = logit_info.top_k(n_beams);
-
-            // Clear the kv slot so that other beams may try different tokens at this position. The llama_decode()
-            // call in loop() will conclusively fill in the kv slot once the beams converge at this position.
-            llama_kv_cache_seq_rm(ctx, 0, n_past, -1);
-
-            size_t i=0;
-            if (next_beams.size() < n_beams) {
-                for (; next_beams.size() < n_beams ; ++i) {
-                    llama_beam next_beam = beam;
-                    next_beam.tokens.push_back(next_tokens[i].id);
-                    next_beam.p *= logit_info.probability_from_logit(next_tokens[i].logit);
-                    next_beams.push_back(std::move(next_beam));
-                }
-                std::make_heap(next_beams.begin(), next_beams.end(), comp);
-            } else {
-                for (; next_beams.front().p == 0.0f ; ++i) {
-                    std::pop_heap(next_beams.begin(), next_beams.end(), comp);
-                    next_beams.back() = beam;
-                    next_beams.back().tokens.push_back(next_tokens[i].id);
-                    next_beams.back().p *= logit_info.probability_from_logit(next_tokens[i].logit);
-                    std::push_heap(next_beams.begin(), next_beams.end(), comp);
-                }
-            }
-            for (; i < n_beams ; ++i) {
-                const float next_p = beam.p * logit_info.probability_from_logit(next_tokens[i].logit);
-                if (next_beams.front().p < next_p) {
-                    std::pop_heap(next_beams.begin(), next_beams.end(), comp);
-                    next_beams.back() = beam;
-                    next_beams.back().tokens.push_back(next_tokens[i].id);
-                    next_beams.back().p = next_p;
-                    std::push_heap(next_beams.begin(), next_beams.end(), comp);
-                }
-            }
-        }
-    }
-
-    // Find common_prefix_length based on beams.
-    // Requires beams is not empty.
-    size_t find_common_prefix_length() {
-        size_t common_prefix_length = beams[0].tokens.size();
-        for (size_t i = 1 ; i < beams.size() ; ++i) {
-            common_prefix_length = std::min(common_prefix_length, beams[i].tokens.size());
-            for (size_t j = 0 ; j < common_prefix_length ; ++j) {
-                if (beams[0].tokens[j] != beams[i].tokens[j]) {
-                    common_prefix_length = j;
-                    break;
-                }
-            }
-        }
-        return common_prefix_length;
-    }
-
-    // Construct beams_state to send back to caller via the callback function.
-    // Side effect: set common_prefix_length = find_common_prefix_length();
-    llama_beams_state get_beams_state(const bool last_call) {
-        for (size_t i = 0 ; i < beams.size() ; ++i) {
-            beam_views[i] = beams[i].view();
-        }
-        common_prefix_length = find_common_prefix_length();
-        return {beam_views.data(), beams.size(), common_prefix_length, last_call};
-    }
-
-    // Loop:
-    //  * while i < n_predict, AND
-    //  * any of the beams have not yet reached end-of-beam (eob), AND
-    //  * the highest probability beam(s) (plural in case of ties) are not at end-of-sentence
-    //    (since all other beam probabilities can only decrease)
-    void loop(const llama_beam_search_callback_fn_t callback, void * const callback_data) {
-        beams.push_back({{}, 1.0f, false});  // Start with one empty beam w/ probability = 1.0 and !eob.
-        const auto not_eob = [](const llama_beam & beam) { return !beam.eob; };
-        for (int i = 0 ; i < n_predict && std::any_of(beams.begin(),beams.end(),not_eob) &&
-                       !beams[top_beam_index()].eob ; ++i) {
-            callback(callback_data, get_beams_state(false));  // Sets common_prefix_length
-            update_beams_from_beam_views();   // Update values (p,eob) that callback may have changed.
-            if (common_prefix_length) {
-                llama_decode(ctx, llama_batch_get_one(beams[0].tokens.data(), common_prefix_length, n_past, 0));
-                n_past += common_prefix_length;
-            }
-            // Zero-out next_beam probabilities to place them last in following min-heap.
-            std::for_each(next_beams.begin(), next_beams.end(), [](llama_beam & beam) { beam.p = 0.0f; });
-            for (llama_beam & beam : beams) {
-                beam.shift_tokens(common_prefix_length);
-                fill_next_beams_by_top_probabilities(beam);
-            }
-            // next_beams become the beams of next/final iteration. Swap them to re-use memory.
-            beams.swap(next_beams);
-            renormalize_beam_probabilities(beams);
-        }
-        collapse_beams(top_beam_index());
-        callback(callback_data, get_beams_state(true));
-    }
-
-    // As beams grow, the cumulative probabilities decrease.
-    // Renormalize them to avoid floating point underflow.
-    static void renormalize_beam_probabilities(std::vector<llama_beam> & beams) {
-        const auto sum_p = [](float sum, llama_beam & beam) { return sum + beam.p; };
-        const float inv_sum = 1.0f / std::accumulate(beams.begin(), beams.end(), 0.0f, sum_p);
-        std::for_each(beams.begin(), beams.end(), [=](llama_beam & beam) { beam.p *= inv_sum; });
-    }
-
-    // Assumes beams is non-empty.  Uses llama_beam::operator<() for ordering.
-    size_t top_beam_index() {
-        return std::max_element(beams.begin(), beams.end()) - beams.begin();
-    }
-
-    // Copy (p,eob) for each beam which may have been changed by the callback.
-    void update_beams_from_beam_views() {
-        for (size_t i = 0 ; i < beams.size() ; ++i) {
-            beams[i].p = beam_views[i].p;
-            beams[i].eob = beam_views[i].eob;
-        }
-    }
-};
-
-void llama_beam_search(llama_context * ctx,
-                       llama_beam_search_callback_fn_t callback, void * callback_data,
-                       size_t n_beams, int n_past, int n_predict) {
-    assert(ctx);
-    const int64_t t_start_sample_us = ggml_time_us();
-
-    llama_beam_search_data beam_search_data(ctx, n_beams, n_past, n_predict);
-
-    beam_search_data.loop(callback, callback_data);
-
-    ctx->t_sample_us += ggml_time_us() - t_start_sample_us;
-    ctx->n_sample++;
-}
-
 //
 // quantization
 //
@@ -14412,8 +14946,6 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
         else if ((ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M) &&
                 use_more_bits(qs.i_attention_wv, qs.n_attention_wv)) new_type = GGML_TYPE_Q6_K;
         else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S && qs.i_attention_wv < 4) new_type = GGML_TYPE_Q5_K;
-        else if (QK_K == 64 && (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_S) &&
-                (qs.i_attention_wv < qs.n_attention_wv/8 || qs.i_attention_wv >= 7*qs.n_attention_wv/8)) new_type = GGML_TYPE_Q6_K;
         if (qs.model.type == MODEL_70B) {
             // In the 70B model we have 8 heads sharing the same attn_v weights. As a result, the attn_v.weight tensor is
             // 8x smaller compared to attn_q.weight. Hence, we can get a nice boost in quantization accuracy with
@@ -14716,6 +15248,14 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
         if (imatrix_data) {
             LLAMA_LOG_INFO("================================ Have weights data with %d entries\n",int(imatrix_data->size()));
             qs.has_imatrix = true;
+            // check imatrix for nans or infs
+            for (const auto & kv : *imatrix_data) {
+                for (float f : kv.second) {
+                    if (!std::isfinite(f)) {
+                        throw std::runtime_error(format("imatrix contains non-finite value %f\n", f));
+                    }
+                }
+            }
         }
     }
 
@@ -15314,6 +15854,7 @@ struct llama_model_params llama_model_default_params() {
         /*.split_mode                  =*/ LLAMA_SPLIT_MODE_LAYER,
         /*.main_gpu                    =*/ 0,
         /*.tensor_split                =*/ nullptr,
+        /*.rpc_servers                 =*/ nullptr,
         /*.progress_callback           =*/ nullptr,
         /*.progress_callback_user_data =*/ nullptr,
         /*.kv_overrides                =*/ nullptr,
@@ -15384,7 +15925,9 @@ struct llama_model_quantize_params llama_model_quantize_default_params() {
 }
 
 size_t llama_max_devices(void) {
-#if defined(GGML_USE_METAL)
+#if defined(GGML_USE_RPC)
+    return GGML_RPC_MAX_SERVERS;
+#elif defined(GGML_USE_METAL)
     return 1;
 #elif defined(GGML_USE_CUDA)
     return GGML_CUDA_MAX_DEVICES;
@@ -15406,8 +15949,8 @@ bool llama_supports_mlock(void) {
 }
 
 bool llama_supports_gpu_offload(void) {
-#if defined(GGML_USE_CUDA) || defined(GGML_USE_CLBLAST) || defined(GGML_USE_METAL) || defined(GGML_USE_VULKAN) || \
-    defined(GGML_USE_SYCL) || defined(GGML_USE_KOMPUTE)
+#if defined(GGML_USE_CUDA) || defined(GGML_USE_METAL)   || defined(GGML_USE_VULKAN) || \
+    defined(GGML_USE_SYCL) || defined(GGML_USE_KOMPUTE) || defined(GGML_USE_RPC)
     // Defined when llama.cpp is compiled with support for offloading model layers to GPU.
     return true;
 #else
@@ -15424,10 +15967,6 @@ void llama_backend_init(void) {
         struct ggml_context * ctx = ggml_init(params);
         ggml_free(ctx);
     }
-
-#ifdef GGML_USE_MPI
-    ggml_mpi_backend_init();
-#endif
 }
 
 void llama_numa_init(enum ggml_numa_strategy numa) {
@@ -15437,9 +15976,6 @@ void llama_numa_init(enum ggml_numa_strategy numa) {
 }
 
 void llama_backend_free(void) {
-#ifdef GGML_USE_MPI
-    ggml_mpi_backend_free();
-#endif
     ggml_quantize_free();
 }
 
@@ -15470,7 +16006,17 @@ struct llama_model * llama_load_model_from_file(
             return true;
         };
     }
-
+    if (params.rpc_servers != nullptr && params.rpc_servers[0] != '\0') {
+        // split the servers set them into model->rpc_servers
+        std::string servers(params.rpc_servers);
+        size_t pos = 0;
+        while ((pos = servers.find(",")) != std::string::npos) {
+            std::string server = servers.substr(0, pos);
+            model->rpc_servers.push_back(server);
+            servers.erase(0, pos + 1);
+        }
+        model->rpc_servers.push_back(servers);
+    }
     int status = llama_model_load(path_model, *model, params);
     GGML_ASSERT(status <= 0);
     if (status < 0) {
@@ -15509,6 +16055,16 @@ struct llama_context * llama_new_context_with_model(
         return nullptr;
     }
 
+    if (params.flash_attn && model->arch == LLM_ARCH_GROK) {
+        LLAMA_LOG_WARN("%s: flash_attn is not compatible with Grok - forcing off\n", __func__);
+        params.flash_attn = false;
+    }
+
+    if (params.type_v != GGML_TYPE_F16 && !params.flash_attn) {
+        LLAMA_LOG_ERROR("%s: V cache quantization requires flash_attn\n", __func__);
+        return nullptr;
+    }
+
     llama_context * ctx = new llama_context(*model);
 
     const auto & hparams = model->hparams;
@@ -15532,7 +16088,7 @@ struct llama_context * llama_new_context_with_model(
     cparams.rope_freq_scale  = params.rope_freq_scale == 0.0f ? hparams.rope_freq_scale_train : params.rope_freq_scale;
 
     // this is necessary due to kv_self.n being padded later during inference
-    cparams.n_ctx            = GGML_PAD(cparams.n_ctx, 256);
+    cparams.n_ctx            = GGML_PAD(cparams.n_ctx, llama_kv_cache_get_padding(cparams));
 
     // with causal attention, the batch size is limited by the context size
     cparams.n_batch          = hparams.causal_attn ? std::min(cparams.n_ctx, params.n_batch) : params.n_batch;
@@ -15547,8 +16103,8 @@ struct llama_context * llama_new_context_with_model(
 
     cparams.n_ubatch         = std::min(cparams.n_batch, params.n_ubatch == 0 ? params.n_batch : params.n_ubatch);
 
-    cparams.n_yarn_orig_ctx  = params.yarn_orig_ctx    != 0 ? params.yarn_orig_ctx    :
-                               hparams.n_yarn_orig_ctx != 0 ? hparams.n_yarn_orig_ctx :
+    cparams.n_ctx_orig_yarn  = params.yarn_orig_ctx    != 0 ? params.yarn_orig_ctx    :
+                               hparams.n_ctx_orig_yarn != 0 ? hparams.n_ctx_orig_yarn :
                                                               hparams.n_ctx_train;
 
     cparams.cb_eval           = params.cb_eval;
@@ -15567,6 +16123,7 @@ struct llama_context * llama_new_context_with_model(
         cparams.yarn_ext_factor = rope_scaling_type == LLAMA_ROPE_SCALING_TYPE_YARN ? 1.0f : 0.0f;
     }
 
+    cparams.yarn_attn_factor *= hparams.rope_attn_factor;
     cparams.causal_attn = hparams.causal_attn;
 
     if (cparams.pooling_type == LLAMA_POOLING_TYPE_UNSPECIFIED) {
@@ -15577,11 +16134,6 @@ struct llama_context * llama_new_context_with_model(
         }
     }
 
-    if (cparams.flash_attn && model->arch == LLM_ARCH_GROK) {
-        LLAMA_LOG_WARN("%s: flash_attn is not compatible with Grok - forcing off\n", __func__);
-        cparams.flash_attn = false;
-    }
-
     if (params.seed == LLAMA_DEFAULT_SEED) {
         params.seed = time(NULL);
     }
@@ -15617,7 +16169,7 @@ struct llama_context * llama_new_context_with_model(
 
     if (!hparams.vocab_only) {
         // initialize backends
-#ifdef GGML_USE_METAL
+#if defined(GGML_USE_METAL)
         if (model->n_gpu_layers > 0) {
             ctx->backend_metal = ggml_backend_metal_init();
             if (ctx->backend_metal == nullptr) {
@@ -15656,7 +16208,7 @@ struct llama_context * llama_new_context_with_model(
             return nullptr;
         }
         if (model->split_mode == LLAMA_SPLIT_MODE_NONE) {
-            ggml_backend_t backend = ggml_backend_vk_init(0);
+            ggml_backend_t backend = ggml_backend_vk_init(model->main_gpu);
             if (backend == nullptr) {
                 LLAMA_LOG_ERROR("%s: failed to initialize Vulkan backend\n", __func__);
                 llama_free(ctx);
@@ -15679,8 +16231,7 @@ struct llama_context * llama_new_context_with_model(
         if (model->split_mode == LLAMA_SPLIT_MODE_NONE || model->split_mode == LLAMA_SPLIT_MODE_ROW) {
             ggml_backend_t backend = ggml_backend_sycl_init(model->main_gpu);
             if (backend == nullptr) {
-                int main_gpu_id = ggml_backend_sycl_get_device_id(model->main_gpu);
-                LLAMA_LOG_ERROR("%s: failed to initialize SYCL%d (index %d) backend\n", __func__, main_gpu_id, model->main_gpu);
+                LLAMA_LOG_ERROR("%s: failed to initialize SYCL%d backend\n", __func__, model->main_gpu);
                 llama_free(ctx);
                 return nullptr;
             }
@@ -15710,6 +16261,29 @@ struct llama_context * llama_new_context_with_model(
             ctx->backends.push_back(backend);
         }
 #endif
+
+#ifdef GGML_USE_BLAS
+        ctx->backend_blas = ggml_backend_blas_init();
+        if (ctx->backend_blas == nullptr) {
+            LLAMA_LOG_WARN("%s: failed to initialize BLAS backend\n", __func__);
+        } else {
+            ctx->backends.push_back(ctx->backend_blas);
+        }
+#endif
+
+#if defined(GGML_USE_RPC)
+        if (model->n_gpu_layers > 0) {
+            for (const auto & endpoint : model->rpc_servers) {
+                ggml_backend_t backend = ggml_backend_rpc_init(endpoint.c_str());
+                if (backend == nullptr) {
+                    LLAMA_LOG_ERROR("%s: failed to initialize RPC to '%s'\n", __func__, endpoint.c_str());
+                    llama_free(ctx);
+                    return nullptr;
+                }
+                ctx->backends.push_back(backend);
+            }
+        }
+#endif
         ctx->backend_cpu = ggml_backend_cpu_init();
         if (ctx->backend_cpu == nullptr) {
             LLAMA_LOG_ERROR("%s: failed to initialize CPU backend\n", __func__);
@@ -15773,7 +16347,11 @@ struct llama_context * llama_new_context_with_model(
             ctx->buf_compute_meta.resize(ggml_tensor_overhead()*LLAMA_MAX_NODES + ggml_graph_overhead_custom(LLAMA_MAX_NODES, false));
 
             // enabling pipeline parallelism in the scheduler increases memory usage, so it is only done when necessary
-            bool pipeline_parallel = llama_get_device_count() > 1 && model->n_gpu_layers > (int)model->hparams.n_layer && model->split_mode == LLAMA_SPLIT_MODE_LAYER;
+            bool pipeline_parallel =
+                llama_get_device_count(*model) > 1 &&
+                model->n_gpu_layers > (int)model->hparams.n_layer &&
+                model->split_mode == LLAMA_SPLIT_MODE_LAYER &&
+                params.offload_kqv;
 #ifndef GGML_USE_CUDA
             // pipeline parallelism requires support for async compute and events
             // currently this is only implemented in the CUDA backend
@@ -15816,20 +16394,6 @@ struct llama_context * llama_new_context_with_model(
         }
     }
 
-#ifdef GGML_USE_MPI
-    ctx->ctx_mpi = ggml_mpi_init();
-
-    if (ggml_mpi_rank(ctx->ctx_mpi) > 0) {
-        // Enter a blocking eval loop with dummy input, letting rank=0 drive the process
-        // TODO: needs fix after #3228
-        GGML_ASSERT(false && "not implemented");
-        //const std::vector<llama_token> tmp(ctx->model.hparams.n_ctx, llama_token_bos(ctx));
-        //while (!llama_eval(ctx, tmp.data(), tmp.size(), 0, 0)) {};
-        llama_backend_free();
-        exit(1);
-    }
-#endif
-
     return ctx;
 }
 
@@ -15866,7 +16430,6 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) {
         // these models do not use RoPE
         case LLM_ARCH_GPT2:
         case LLM_ARCH_GPTJ:
-        case LLM_ARCH_GPTNEOX:
         case LLM_ARCH_MPT:
         case LLM_ARCH_REFACT:
         case LLM_ARCH_BLOOM:
@@ -15886,13 +16449,14 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) {
         case LLM_ARCH_XVERSE:
         case LLM_ARCH_COMMAND_R:
         case LLM_ARCH_OLMO:
+        case LLM_ARCH_ARCTIC:
+        case LLM_ARCH_DEEPSEEK2:
             return LLAMA_ROPE_TYPE_NORM;
 
         // the pairs of head values are offset by n_rot/2
         case LLM_ARCH_FALCON:
         case LLM_ARCH_GROK:
         case LLM_ARCH_DBRX:
-        case LLM_ARCH_PERSIMMON:
         case LLM_ARCH_BERT:
         case LLM_ARCH_NOMIC_BERT:
         case LLM_ARCH_STABLELM:
@@ -15903,6 +16467,7 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) {
         case LLM_ARCH_PHI3:
         case LLM_ARCH_GEMMA:
         case LLM_ARCH_STARCODER2:
+        case LLM_ARCH_GPTNEOX:
             return LLAMA_ROPE_TYPE_NEOX;
 
         // all model arches should be listed explicitly here
@@ -16062,6 +16627,7 @@ static bool llama_control_vector_init(struct llama_control_vector & cvec, const
     }
 
     // make tensors
+    cvec.tensors.reserve(model.hparams.n_layer);
     cvec.tensors.push_back(nullptr); // there's never a tensor for layer 0
     for (size_t il = 1; il < model.hparams.n_layer; il++) {
         struct ggml_context * ctx = ctx_map.at(model.buft_layer[il].buft);
@@ -16070,6 +16636,8 @@ static bool llama_control_vector_init(struct llama_control_vector & cvec, const
     }
 
     // allocate tensors / buffers and zero
+    cvec.ctxs.reserve(ctx_map.size());
+    cvec.bufs.reserve(ctx_map.size());
     for (auto it : ctx_map) {
         ggml_backend_buffer_type_t buft = it.first;
         ggml_context * ctx = it.second;
@@ -16893,13 +17461,13 @@ static size_t llama_state_seq_get_data_internal(struct llama_context * ctx, llam
             }
             else {
                 if (cell_range_begin != kv_self.size) {
-                    cell_ranges.push_back({ cell_range_begin, i });
+                    cell_ranges.emplace_back(cell_range_begin, i);
                     cell_range_begin = kv_self.size;
                 }
             }
         }
         if (cell_range_begin != kv_self.size) {
-            cell_ranges.push_back({ cell_range_begin, kv_self.size });
+            cell_ranges.emplace_back(cell_range_begin, kv_self.size);
         }
 
         // DEBUG CHECK: Sum of cell counts in ranges should equal the total cell count
@@ -17278,6 +17846,14 @@ void llama_set_n_threads(struct llama_context * ctx, uint32_t n_threads, uint32_
     ctx->cparams.n_threads_batch = n_threads_batch;
 }
 
+uint32_t llama_n_threads(struct llama_context * ctx) {
+    return ctx->cparams.n_threads;
+}
+
+uint32_t llama_n_threads_batch(struct llama_context * ctx) {
+    return ctx->cparams.n_threads_batch;
+}
+
 void llama_set_abort_callback(struct llama_context * ctx, bool (*abort_callback)(void * data), void * abort_callback_data) {
     ctx->abort_callback      = abort_callback;
     ctx->abort_callback_data = abort_callback_data;
@@ -17489,9 +18065,9 @@ float llama_token_get_score(const struct llama_model * model, llama_token token)
     return model->vocab.id_to_token[token].score;
 }
 
-llama_token_type llama_token_get_type(const struct llama_model * model, llama_token token) {
+llama_token_attr llama_token_get_attr(const struct llama_model * model, llama_token token) {
     GGML_ASSERT(model->vocab.type != LLAMA_VOCAB_TYPE_NONE);
-    return model->vocab.id_to_token[token].type;
+    return model->vocab.id_to_token[token].attr;
 }
 
 bool llama_token_is_eog(const struct llama_model * model, llama_token token) {
@@ -17501,6 +18077,10 @@ bool llama_token_is_eog(const struct llama_model * model, llama_token token) {
     );
 }
 
+bool llama_token_is_control(const struct llama_model * model, llama_token token) {
+    return llama_is_control_token(model->vocab, token);
+}
+
 llama_token llama_token_bos(const struct llama_model * model) {
     return model->vocab.special_bos_id;
 }
@@ -17572,7 +18152,16 @@ static std::string llama_decode_text(const std::string & text) {
 
     const auto cpts = unicode_cpts_from_utf8(text);
     for (const auto cpt : cpts) {
-        decoded_text += unicode_utf8_to_byte(unicode_cpt_to_utf8(cpt));
+        const auto utf8 = unicode_cpt_to_utf8(cpt);
+        try {
+            decoded_text += unicode_utf8_to_byte(utf8);
+        } catch (const std::out_of_range & e) {
+            decoded_text += "[UNK_BYTE_0x";
+            for (const auto c : utf8) {
+                decoded_text += format("%02x", (uint8_t) c);
+            }
+            decoded_text += text + "]";
+        }
     }
 
     return decoded_text;
@@ -17580,69 +18169,88 @@ static std::string llama_decode_text(const std::string & text) {
 
 // does not write null-terminator to buf
 int32_t llama_token_to_piece(const struct llama_model * model, llama_token token, char * buf, int32_t length, bool special) {
+    // ref: https://github.com/ggerganov/llama.cpp/pull/7587#discussion_r1620983843
+    if (!special && llama_is_control_token(model->vocab, token)) {
+        return 0;
+    }
+
+    // if we have a cache - use it
+    {
+        const auto & cache = model->vocab.cache_token_to_piece;
+
+        if (!cache.empty()) {
+            const auto & res = cache.at(token);
+            if (length < (int) res.size()) {
+                return -(int) res.size();
+            }
+            memcpy(buf, res.c_str(), res.size());
+            return res.size();
+        }
+    }
+
     if (0 <= token && token < llama_n_vocab(model)) {
         switch (llama_vocab_get_type(model->vocab)) {
-        case LLAMA_VOCAB_TYPE_WPM:
-        case LLAMA_VOCAB_TYPE_SPM: {
-            // NOTE: we accept all unsupported token types,
-            // suppressing them like CONTROL tokens.
-            if (llama_is_normal_token(model->vocab, token)) {
-                std::string result = model->vocab.id_to_token[token].text;
-                llama_unescape_whitespace(result);
-                if (length < (int) result.length()) {
-                    return -(int) result.length();
-                }
-                memcpy(buf, result.c_str(), result.length());
-                return result.length();
-            } else if (
-                    (llama_is_user_defined_token(model->vocab, token)) ||
-                    (llama_is_control_token     (model->vocab, token) && special)) {
-                std::string result = model->vocab.id_to_token[token].text;
-                if (length < (int) result.length()) {
-                    return -(int) result.length();
-                }
-                memcpy(buf, result.c_str(), result.length());
-                return result.length();
-            } else if (llama_is_unknown_token(model->vocab, token)) { // NOLINT
-                if (length < 3) {
-                    return -3;
-                }
-                memcpy(buf, "\xe2\x96\x85", 3);
-                return 3;
-            } else if (llama_is_byte_token(model->vocab, token)) {
-                if (length < 1) {
-                    return -1;
+            case LLAMA_VOCAB_TYPE_WPM:
+            case LLAMA_VOCAB_TYPE_SPM: {
+                // NOTE: we accept all unsupported token types,
+                // suppressing them like CONTROL tokens.
+                if (llama_is_normal_token(model->vocab, token)) {
+                    std::string result = model->vocab.id_to_token[token].text;
+                    llama_unescape_whitespace(result);
+                    if (length < (int) result.length()) {
+                        return -(int) result.length();
+                    }
+                    memcpy(buf, result.c_str(), result.length());
+                    return result.length();
+                } else if (
+                        (llama_is_user_defined_token(model->vocab, token)) ||
+                        (llama_is_control_token     (model->vocab, token) && special)) {
+                    std::string result = model->vocab.id_to_token[token].text;
+                    if (length < (int) result.length()) {
+                        return -(int) result.length();
+                    }
+                    memcpy(buf, result.c_str(), result.length());
+                    return result.length();
+                } else if (llama_is_unknown_token(model->vocab, token)) { // NOLINT
+                    if (length < 3) {
+                        return -3;
+                    }
+                    memcpy(buf, "\xe2\x96\x85", 3);
+                    return 3;
+                } else if (llama_is_byte_token(model->vocab, token)) {
+                    if (length < 1) {
+                        return -1;
+                    }
+                    buf[0] = llama_token_to_byte(model->vocab, token);
+                    return 1;
                 }
-                buf[0] = llama_token_to_byte(model->vocab, token);
-                return 1;
+                break;
             }
-            break;
-        }
-        case LLAMA_VOCAB_TYPE_BPE: {
-            // NOTE: we accept all unsupported token types,
-            // suppressing them like CONTROL tokens.
-            if (llama_is_normal_token(model->vocab, token)) {
-                std::string result = model->vocab.id_to_token[token].text;
-                result = llama_decode_text(result);
-                if (length < (int) result.length()) {
-                    return -(int) result.length();
-                }
-                memcpy(buf, result.c_str(), result.length());
-                return result.length();
-            } else if (
-                    (llama_is_user_defined_token(model->vocab, token)) ||
-                    (llama_is_control_token     (model->vocab, token) && special)) {
-                std::string result = model->vocab.id_to_token[token].text;
-                if (length < (int) result.length()) {
-                    return -(int) result.length();
+            case LLAMA_VOCAB_TYPE_BPE: {
+                // NOTE: we accept all unsupported token types,
+                // suppressing them like CONTROL tokens.
+                if (llama_is_normal_token(model->vocab, token)) {
+                    std::string result = model->vocab.id_to_token[token].text;
+                    result = llama_decode_text(result);
+                    if (length < (int) result.length()) {
+                        return -(int) result.length();
+                    }
+                    memcpy(buf, result.c_str(), result.length());
+                    return result.length();
+                } else if (
+                        (llama_is_user_defined_token(model->vocab, token)) ||
+                        (llama_is_control_token     (model->vocab, token) && special)) {
+                    std::string result = model->vocab.id_to_token[token].text;
+                    if (length < (int) result.length()) {
+                        return -(int) result.length();
+                    }
+                    memcpy(buf, result.c_str(), result.length());
+                    return result.length();
                 }
-                memcpy(buf, result.c_str(), result.length());
-                return result.length();
+                break;
             }
-            break;
-        }
-        default:
-            GGML_ASSERT(false);
+            default:
+                GGML_ASSERT(false);
         }
     }
     return 0;
@@ -17712,6 +18320,15 @@ static int32_t llama_chat_apply_template_internal(
             }
         }
         // llama2 templates seem to not care about "add_generation_prompt"
+    } else if (tmpl == "phi3" || (tmpl.find("<|assistant|>") != std::string::npos && tmpl.find("<|end|>") != std::string::npos)) {
+        // Phi 3
+        for (auto message : chat) {
+            std::string role(message->role);
+            ss << "<|" << role << "|>\n" << message->content << "<|end|>\n";
+        }
+        if (add_ass) {
+            ss << "<|assistant|>\n";
+        }
     } else if (tmpl == "zephyr" || tmpl.find("<|user|>") != std::string::npos) {
         // zephyr template
         for (auto message : chat) {
@@ -17844,15 +18461,6 @@ static int32_t llama_chat_apply_template_internal(
         if (add_ass) {
             ss << "<|start_header_id|>assistant<|end_header_id|>\n\n";
         }
-    } else if (tmpl == "phi3" || (tmpl.find("<|assistant|>") != std::string::npos && tmpl.find("<|end|>") != std::string::npos )) {
-        // Phi 3
-        for (auto message : chat) {
-            std::string role(message->role);
-            ss << "<|" << role << "|>\n" << trim(message->content) << "<|end|>\n";
-        }
-        if (add_ass) {
-            ss << "<|assistant|>\n";
-        }
     } else {
         // template not supported
         return -1;
@@ -17974,8 +18582,10 @@ const char * llama_print_system_info(void) {
     s += "AVX512 = "      + std::to_string(ggml_cpu_has_avx512())      + " | ";
     s += "AVX512_VBMI = " + std::to_string(ggml_cpu_has_avx512_vbmi()) + " | ";
     s += "AVX512_VNNI = " + std::to_string(ggml_cpu_has_avx512_vnni()) + " | ";
+    s += "AVX512_BF16 = " + std::to_string(ggml_cpu_has_avx512_bf16()) + " | ";
     s += "FMA = "         + std::to_string(ggml_cpu_has_fma())         + " | ";
     s += "NEON = "        + std::to_string(ggml_cpu_has_neon())        + " | ";
+    s += "SVE = "         + std::to_string(ggml_cpu_has_sve())         + " | ";
     s += "ARM_FMA = "     + std::to_string(ggml_cpu_has_arm_fma())     + " | ";
     s += "F16C = "        + std::to_string(ggml_cpu_has_f16c())        + " | ";
     s += "FP16_VA = "     + std::to_string(ggml_cpu_has_fp16_va())     + " | ";
@@ -18034,6 +18644,8 @@ void llama_log_set(ggml_log_callback log_callback, void * user_data) {
     g_state.log_callback_user_data = user_data;
 #ifdef GGML_USE_METAL
     ggml_backend_metal_log_set_callback(g_state.log_callback, g_state.log_callback_user_data);
+#elif defined(GGML_USE_CUDA)
+    ggml_backend_cuda_log_set_callback(g_state.log_callback, g_state.log_callback_user_data);
 #endif
 }