llama_cpp 0.12.2 → 0.12.4

data/vendor/tmp/llama.cpp/llama.cpp

@@ -11,6 +11,12 @@
 #  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)
+#  include "ggml-sycl.h"
+#elif defined(GGML_USE_KOMPUTE)
+#   include "ggml-kompute.h"
 #endif
 
 #ifdef GGML_USE_METAL
@@ -52,6 +58,7 @@
 #include <algorithm>
 #include <array>
 #include <cassert>
+#include <cfloat>
 #include <cinttypes>
 #include <climits>
 #include <cmath>
@@ -192,8 +199,12 @@ enum llm_arch {
     LLM_ARCH_BLOOM,
     LLM_ARCH_STABLELM,
     LLM_ARCH_QWEN,
+    LLM_ARCH_QWEN2,
     LLM_ARCH_PHI2,
     LLM_ARCH_PLAMO,
+    LLM_ARCH_CODESHELL,
+    LLM_ARCH_ORION,
+    LLM_ARCH_INTERNLM2,
     LLM_ARCH_UNKNOWN,
 };
 
@@ -211,8 +222,12 @@ static std::map<llm_arch, std::string> LLM_ARCH_NAMES = {
     { LLM_ARCH_BLOOM,           "bloom"     },
     { LLM_ARCH_STABLELM,        "stablelm"  },
     { LLM_ARCH_QWEN,            "qwen"      },
+    { LLM_ARCH_QWEN2,           "qwen2"     },
     { LLM_ARCH_PHI2,            "phi2"      },
     { LLM_ARCH_PLAMO,           "plamo"     },
+    { LLM_ARCH_CODESHELL,       "codeshell" },
+    { LLM_ARCH_ORION,           "orion"     },
+    { LLM_ARCH_INTERNLM2,       "internlm2" },
 };
 
 enum llm_kv {
@@ -265,6 +280,7 @@ enum llm_kv {
     LLM_KV_TOKENIZER_PAD_ID,
     LLM_KV_TOKENIZER_ADD_BOS,
     LLM_KV_TOKENIZER_ADD_EOS,
+    LLM_KV_TOKENIZER_ADD_PREFIX,
     LLM_KV_TOKENIZER_HF_JSON,
     LLM_KV_TOKENIZER_RWKV,
 };
@@ -319,6 +335,7 @@ static std::map<llm_kv, std::string> LLM_KV_NAMES = {
     { LLM_KV_TOKENIZER_PAD_ID,              "tokenizer.ggml.padding_token_id"   },
     { LLM_KV_TOKENIZER_ADD_BOS,             "tokenizer.ggml.add_bos_token"      },
     { LLM_KV_TOKENIZER_ADD_EOS,             "tokenizer.ggml.add_eos_token"      },
+    { LLM_KV_TOKENIZER_ADD_PREFIX,          "tokenizer.ggml.add_space_prefix"   },
     { LLM_KV_TOKENIZER_HF_JSON,             "tokenizer.huggingface.json"        },
     { LLM_KV_TOKENIZER_RWKV,                "tokenizer.rwkv.world"              },
 };
@@ -566,6 +583,23 @@ static std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NAMES =
             { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
         },
     },
+    {
+        LLM_ARCH_QWEN2,
+        {
+            { 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_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_ARCH_PHI2,
         {
@@ -600,7 +634,62 @@ static std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NAMES =
             { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
         },
     },
-
+    {
+        LLM_ARCH_CODESHELL,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_OUTPUT,          "output" },
+            { LLM_TENSOR_ROPE_FREQS,      "rope_freqs" },
+            { 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_QKV,        "blk.%d.attn_qkv" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_ATTN_ROT_EMBD,   "blk.%d.attn_rot_embd" },
+            { 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_ARCH_ORION,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_OUTPUT,          "output" },
+            { LLM_TENSOR_ROPE_FREQS,      "rope_freqs" },
+            { 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_ATTN_ROT_EMBD,   "blk.%d.attn_rot_embd" },
+            { 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_ARCH_INTERNLM2,
+        {
+            { 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_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_ARCH_UNKNOWN,
         {
@@ -1091,10 +1180,10 @@ struct llama_mlock {
     #ifdef __APPLE__
         #define MLOCK_SUGGESTION \
             "Try increasing the sysctl values 'vm.user_wire_limit' and 'vm.global_user_wire_limit' and/or " \
-            "decreasing 'vm.global_no_user_wire_amount'.  Also try increasing RLIMIT_MLOCK (ulimit -l).\n"
+            "decreasing 'vm.global_no_user_wire_amount'.  Also try increasing RLIMIT_MEMLOCK (ulimit -l).\n"
     #else
         #define MLOCK_SUGGESTION \
-            "Try increasing RLIMIT_MLOCK ('ulimit -l' as root).\n"
+            "Try increasing RLIMIT_MEMLOCK ('ulimit -l' as root).\n"
     #endif
 
     bool raw_lock(const void * addr, size_t size) const {
@@ -1215,8 +1304,14 @@ static ggml_backend_buffer_type_t llama_default_buffer_type_cpu(bool host_buffer
     if (host_buffer) {
         buft = ggml_backend_cuda_host_buffer_type();
     }
+#elif defined(GGML_USE_SYCL)
+    buft = ggml_backend_sycl_host_buffer_type();
 #elif defined(GGML_USE_CPU_HBM)
     buft = ggml_backend_cpu_hbm_buffer_type();
+#elif defined(GGML_USE_VULKAN)
+    if (host_buffer) {
+        buft = ggml_backend_vk_host_buffer_type();
+    }
 #endif
 
     if (buft == nullptr) {
@@ -1234,8 +1329,17 @@ static ggml_backend_buffer_type_t llama_default_buffer_type_offload(int gpu) {
     buft = ggml_backend_metal_buffer_type();
 #elif defined(GGML_USE_CUBLAS)
     buft = ggml_backend_cuda_buffer_type(gpu);
+#elif defined(GGML_USE_VULKAN)
+    buft = ggml_backend_vk_buffer_type();
+#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) {
@@ -1284,12 +1388,16 @@ static llama_state g_state;
 // available llama models
 enum e_model {
     MODEL_UNKNOWN,
+    MODEL_0_5B,
     MODEL_1B,
     MODEL_3B,
+    MODEL_4B,
     MODEL_7B,
     MODEL_8B,
     MODEL_13B,
+    MODEL_14B,
     MODEL_15B,
+    MODEL_20B,
     MODEL_30B,
     MODEL_34B,
     MODEL_40B,
@@ -1393,6 +1501,9 @@ struct llama_cparams {
 
     bool mul_mat_q;
     bool offload_kqv;
+
+    ggml_backend_sched_eval_callback cb_eval;
+    void * cb_eval_user_data;
 };
 
 struct llama_layer {
@@ -1528,6 +1639,8 @@ struct llama_vocab {
     id special_suffix_id = 32008;
     id special_eot_id    = 32010;
 
+    bool add_space_prefix = true;
+
     int find_bpe_rank(const std::string & token_left, const std::string & token_right) const {
         GGML_ASSERT(token_left.find(' ') == std::string::npos);
         GGML_ASSERT(token_left.find('\n') == std::string::npos);
@@ -1596,7 +1709,7 @@ struct llama_model {
     std::unique_ptr<llama_mmap> mapping;
 
     // objects representing data potentially being locked in memory
-    llama_mlock mlock_buf;
+    std::vector<std::unique_ptr<llama_mlock>> mlock_bufs;
     llama_mlock mlock_mmap;
 
     // for quantize-stats only
@@ -1623,6 +1736,9 @@ struct llama_context {
         for (ggml_backend_t backend : backends) {
             ggml_backend_free(backend);
         }
+
+        ggml_backend_buffer_free(buf_input);
+        ggml_free(ctx_input);
     }
 
     llama_cparams cparams;
@@ -1669,8 +1785,14 @@ struct llama_context {
     // allocator for the input tensors
     ggml_tallocr * alloc = nullptr;
 
-    // temporary buffer for copying data to/from the backend
-    std::vector<no_init<uint8_t>> buf_copy;
+    // input tensors
+    ggml_backend_buffer_t buf_input = nullptr;
+    ggml_context * ctx_input = nullptr;
+    struct ggml_tensor * inp_tokens;    // I32 [n_batch]
+    struct ggml_tensor * inp_embd;      // F32 [n_embd, n_batch]
+    struct ggml_tensor * inp_pos;       // I32 [n_batch]
+    struct ggml_tensor * inp_KQ_mask;   // F32 [n_ctx, n_batch]
+    struct ggml_tensor * inp_K_shift;   // I32 [n_ctx]
 
 #ifdef GGML_USE_MPI
     ggml_mpi_context * ctx_mpi = NULL;
@@ -2254,20 +2376,21 @@ struct llama_model_loader {
             }
 
             switch (type_max) {
-                case GGML_TYPE_F32:  ftype = LLAMA_FTYPE_ALL_F32;       break;
-                case GGML_TYPE_F16:  ftype = LLAMA_FTYPE_MOSTLY_F16;    break;
-                case GGML_TYPE_Q4_0: ftype = LLAMA_FTYPE_MOSTLY_Q4_0;   break;
-                case GGML_TYPE_Q4_1: ftype = LLAMA_FTYPE_MOSTLY_Q4_1;   break;
-                case GGML_TYPE_Q5_0: ftype = LLAMA_FTYPE_MOSTLY_Q5_0;   break;
-                case GGML_TYPE_Q5_1: ftype = LLAMA_FTYPE_MOSTLY_Q5_1;   break;
-                case GGML_TYPE_Q8_0: ftype = LLAMA_FTYPE_MOSTLY_Q8_0;   break;
-                case GGML_TYPE_Q2_K: ftype = LLAMA_FTYPE_MOSTLY_Q2_K;   break;
-                case GGML_TYPE_Q3_K: ftype = LLAMA_FTYPE_MOSTLY_Q3_K_M; break;
-                case GGML_TYPE_Q4_K: ftype = LLAMA_FTYPE_MOSTLY_Q4_K_M; break;
-                case GGML_TYPE_Q5_K: ftype = LLAMA_FTYPE_MOSTLY_Q5_K_M; break;
-                case GGML_TYPE_Q6_K: ftype = LLAMA_FTYPE_MOSTLY_Q6_K;   break;
+                case GGML_TYPE_F32:     ftype = LLAMA_FTYPE_ALL_F32;        break;
+                case GGML_TYPE_F16:     ftype = LLAMA_FTYPE_MOSTLY_F16;     break;
+                case GGML_TYPE_Q4_0:    ftype = LLAMA_FTYPE_MOSTLY_Q4_0;    break;
+                case GGML_TYPE_Q4_1:    ftype = LLAMA_FTYPE_MOSTLY_Q4_1;    break;
+                case GGML_TYPE_Q5_0:    ftype = LLAMA_FTYPE_MOSTLY_Q5_0;    break;
+                case GGML_TYPE_Q5_1:    ftype = LLAMA_FTYPE_MOSTLY_Q5_1;    break;
+                case GGML_TYPE_Q8_0:    ftype = LLAMA_FTYPE_MOSTLY_Q8_0;    break;
+                case GGML_TYPE_Q2_K:    ftype = LLAMA_FTYPE_MOSTLY_Q2_K;    break;
+                case GGML_TYPE_Q3_K:    ftype = LLAMA_FTYPE_MOSTLY_Q3_K_M;  break;
+                case GGML_TYPE_Q4_K:    ftype = LLAMA_FTYPE_MOSTLY_Q4_K_M;  break;
+                case GGML_TYPE_Q5_K:    ftype = LLAMA_FTYPE_MOSTLY_Q5_K_M;  break;
+                case GGML_TYPE_Q6_K:    ftype = LLAMA_FTYPE_MOSTLY_Q6_K;    break;
                 case GGML_TYPE_IQ2_XXS: ftype = LLAMA_FTYPE_MOSTLY_IQ2_XXS; break;
                 case GGML_TYPE_IQ2_XS:  ftype = LLAMA_FTYPE_MOSTLY_IQ2_XS;  break;
+                case GGML_TYPE_IQ3_XXS: ftype = LLAMA_FTYPE_MOSTLY_IQ3_XXS; break;
                 default:
                     {
                         LLAMA_LOG_WARN("%s: unknown type %s\n", __func__, ggml_type_name(type_max));
@@ -2613,8 +2736,10 @@ static std::string llama_model_ftype_name(llama_ftype ftype) {
         case LLAMA_FTYPE_MOSTLY_Q5_K_S: return "Q5_K - Small";
         case LLAMA_FTYPE_MOSTLY_Q5_K_M: return "Q5_K - Medium";
         case LLAMA_FTYPE_MOSTLY_Q6_K:   return "Q6_K";
-        case LLAMA_FTYPE_MOSTLY_IQ2_XXS:return "IQ2_XSS - 2.0625 bpw";
+        case LLAMA_FTYPE_MOSTLY_IQ2_XXS:return "IQ2_XXS - 2.0625 bpw";
         case LLAMA_FTYPE_MOSTLY_IQ2_XS: return "IQ2_XS - 2.3125 bpw";
+        case LLAMA_FTYPE_MOSTLY_Q3_K_XS:return "Q3_K - Extra small";
+        case LLAMA_FTYPE_MOSTLY_IQ3_XXS:return "IQ3_XXS - 3.0625 bpw";
 
         default: return "unknown, may not work";
     }
@@ -2627,7 +2752,9 @@ static const char * llama_model_type_name(e_model type) {
         case MODEL_7B:     return "7B";
         case MODEL_8B:     return "8B";
         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_40B:    return "40B";
@@ -2640,6 +2767,14 @@ static const char * llama_model_type_name(e_model type) {
         default:           return "?B";
     }
 }
+static const char * llama_model_vocab_type_name(enum llama_vocab_type type){
+    switch (type) {
+        case LLAMA_VOCAB_TYPE_SPM:         return "SPM";
+        case LLAMA_VOCAB_TYPE_BPE:         return "BPE";
+        default:                           return "unknown";
+    }
+}
+
 
 static void llm_load_arch(llama_model_loader & ml, llama_model & model) {
     model.arch = ml.get_arch();
@@ -2830,6 +2965,7 @@ static void llm_load_hparams(
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps);
 
                 switch (hparams.n_layer) {
+                    case 24: model.type = e_model::MODEL_1B; break;
                     case 32: model.type = e_model::MODEL_3B; break;
                     default: model.type = e_model::MODEL_UNKNOWN;
                }
@@ -2844,6 +2980,17 @@ static void llm_load_hparams(
                     default: model.type = e_model::MODEL_UNKNOWN;
                 }
             } break;
+        case LLM_ARCH_QWEN2:
+            {
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+                switch (hparams.n_layer) {
+                    case 24: model.type = hparams.n_embd == 1024 ? e_model::MODEL_0_5B : e_model::MODEL_1B; break;
+                    case 32: model.type = e_model::MODEL_7B; break;
+                    case 40: model.type = hparams.n_head == 20 ? e_model::MODEL_4B : e_model::MODEL_13B; break;
+                    case 80: model.type = e_model::MODEL_70B; break;
+                    default: model.type = e_model::MODEL_UNKNOWN;
+                }
+            } break;
         case LLM_ARCH_PHI2:
             {
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps);
@@ -2874,7 +3021,32 @@ static void llm_load_hparams(
                     default: model.type = e_model::MODEL_UNKNOWN;
                 }
             } break;
+        case LLM_ARCH_CODESHELL:
+            {
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps);
+                switch (hparams.n_layer) {
+                    case 42: model.type = e_model::MODEL_SMALL; break;
+                    default: model.type = e_model::MODEL_UNKNOWN;
+                }
+            } break;
+        case LLM_ARCH_ORION:
+            {
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps);
 
+                switch (hparams.n_layer) {
+                    case 40: model.type = e_model::MODEL_14B; break;
+                    default: model.type = e_model::MODEL_UNKNOWN;
+                }
+            } break;
+        case LLM_ARCH_INTERNLM2:
+            {
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+                switch (hparams.n_layer) {
+                    case 32: model.type = e_model::MODEL_7B; break;
+                    case 48: model.type = e_model::MODEL_20B; break;
+                    default: model.type = e_model::MODEL_UNKNOWN;
+                }
+            } break;
         default: (void)0;
     }
 
@@ -2926,6 +3098,11 @@ static void llm_load_vocab(
             vocab.special_unk_id = 0;
             vocab.special_sep_id = -1;
             vocab.special_pad_id = -1;
+
+            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);
+            } // The default value of add_space_prefix is true.
         } else if (tokenizer_name == "gpt2") {
             vocab.type = LLAMA_VOCAB_TYPE_BPE;
 
@@ -3138,7 +3315,7 @@ static void llm_load_print_meta(llama_model_loader & ml, llama_model & model) {
     // hparams
     LLAMA_LOG_INFO("%s: format           = %s\n",     __func__, llama_file_version_name(ml.fver));
     LLAMA_LOG_INFO("%s: arch             = %s\n",     __func__, LLM_ARCH_NAMES.at(model.arch).c_str());
-    LLAMA_LOG_INFO("%s: vocab type       = %s\n",     __func__, vocab.type == LLAMA_VOCAB_TYPE_SPM ? "SPM" : "BPE"); // TODO: fix
+    LLAMA_LOG_INFO("%s: vocab type       = %s\n",     __func__, llama_model_vocab_type_name(vocab.type));
     LLAMA_LOG_INFO("%s: n_vocab          = %u\n",     __func__, hparams.n_vocab);
     LLAMA_LOG_INFO("%s: n_merges         = %u\n",     __func__, (int) vocab.bpe_ranks.size());
     LLAMA_LOG_INFO("%s: n_ctx_train      = %u\n",     __func__, hparams.n_ctx_train);
@@ -3435,7 +3612,12 @@ 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});
+                        if (gguf_find_tensor(ml.ctx_gguf, tn(LLM_TENSOR_OUTPUT, "weight").c_str()) >= 0) {
+                            model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT,     "weight"), {n_embd, n_vocab});
+                        } else {
+                            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
+                        }
                     }
 
                     for (int i = 0; i < n_layer; ++i) {
@@ -3629,6 +3811,11 @@ static bool llm_load_tensors(
                         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});
 
+                        // 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.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});
 
@@ -3666,6 +3853,41 @@ static bool llm_load_tensors(
                         layer.ffn_up   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_ff/2});
                     }
                 } break;
+            case LLM_ARCH_QWEN2:
+                {
+                    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});
+
+                        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});
+
+                        // optional bias tensors
+                        layer.bq = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q,   "bias", i), {n_embd});
+                        layer.bk = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K,   "bias", i), {n_embd_gqa});
+                        layer.bv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_V,   "bias", i), {n_embd_gqa});
+
+                        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_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});
+                    }
+                } break;
             case LLM_ARCH_PHI2:
                 {
                     model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
@@ -3776,6 +3998,101 @@ 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_CODESHELL:
+                {
+                    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});
+                    }
+
+                    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_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.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_ORION:
+                {
+                    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.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_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "bias", i),   {n_embd});
+
+                        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});
+                    }
+                } break;
+            case LLM_ARCH_INTERNLM2:
+                {
+                    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});
+                        // layer.wqkv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa});
+                        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_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});
+                    }
+                } break;
             default:
                 throw std::runtime_error("unknown architecture");
         }
@@ -3812,8 +4129,10 @@ static bool llm_load_tensors(
         else {
             buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx, buft);
             if (buf != nullptr && use_mlock && ggml_backend_buffer_is_host(buf)) {
-                model.mlock_buf.init   (ggml_backend_buffer_get_base(buf));
-                model.mlock_buf.grow_to(ggml_backend_buffer_get_size(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));
             }
         }
         if (buf == nullptr) {
@@ -3870,7 +4189,7 @@ static bool llm_load_tensors(
 }
 
 // Returns 0 on success, -1 on error, and -2 on cancellation via llama_progress_callback
-static int llama_model_load(const std::string & fname, llama_model & model, const llama_model_params & params) {
+static int llama_model_load(const std::string & fname, llama_model & model, llama_model_params & params) {
     try {
         llama_model_loader ml(fname, params.use_mmap, params.kv_overrides);
 
@@ -3891,6 +4210,22 @@ static int llama_model_load(const std::string & fname, llama_model & model, cons
             return 0;
         }
 
+#ifdef GGML_USE_KOMPUTE
+        if (params.n_gpu_layers > 0 && (
+            !(model.arch == LLM_ARCH_LLAMA || model.arch == LLM_ARCH_FALCON)
+            || !(
+                model.ftype == LLAMA_FTYPE_ALL_F32 ||
+                model.ftype == LLAMA_FTYPE_MOSTLY_F16 ||
+                model.ftype == LLAMA_FTYPE_MOSTLY_Q4_0 ||
+                model.ftype == LLAMA_FTYPE_MOSTLY_Q4_1
+            )
+        )) {
+            // TODO(cebtenzzre): propagate this error outside of llama_load_model_from_file
+            LLAMA_LOG_WARN("%s: disabling Kompute due to unsupported model arch or quantization\n", __func__);
+            params.n_gpu_layers = 0;
+        }
+#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
@@ -3939,22 +4274,24 @@ static struct ggml_tensor * llm_build_inp_embd(
         const llama_hparams & hparams,
           const llama_batch & batch,
          struct ggml_tensor * tok_embd,
+         struct ggml_tensor * inp_tokens,
+         struct ggml_tensor * inp_embd,
          const llm_build_cb & cb) {
     const int64_t n_embd = hparams.n_embd;
 
     struct ggml_tensor * inpL;
 
     if (batch.token) {
-        struct ggml_tensor * inp_tokens = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, batch.n_tokens);
+        struct ggml_tensor * inp_tokens_v = ggml_view_1d(ctx, inp_tokens, batch.n_tokens, 0);
         cb(inp_tokens, "inp_tokens", -1);
 
-        inpL = ggml_get_rows(ctx, tok_embd, inp_tokens);
+        inpL = ggml_get_rows(ctx, tok_embd, inp_tokens_v);
     } else {
 #ifdef GGML_USE_MPI
         GGML_ASSERT(false && "not implemented");
 #endif
 
-        inpL = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, batch.n_tokens);
+        inpL = ggml_view_2d(ctx, inp_embd, n_embd, batch.n_tokens, inp_embd->nb[1], 0);
     }
 
     return inpL;
@@ -3968,6 +4305,7 @@ static void llm_build_k_shift(
       const llama_cparams & cparams,
      const llama_kv_cache & kv,
        struct ggml_cgraph * graph,
+       struct ggml_tensor * K_shift,
             llm_rope_type   type,
                   int64_t   n_ctx,
                   float     freq_base,
@@ -3984,9 +4322,6 @@ static void llm_build_k_shift(
     const float   beta_fast     = cparams.yarn_beta_fast;
     const float   beta_slow     = cparams.yarn_beta_slow;
 
-    struct ggml_tensor * K_shift = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, n_ctx);
-    cb(K_shift, "K_shift", -1);
-
     int rope_type = 0;
 
     switch (type) {
@@ -4174,6 +4509,7 @@ static struct ggml_tensor * llm_build_kqv(
           const llama_model & model,
         const llama_hparams & hparams,
        const llama_kv_cache & kv,
+         struct ggml_cgraph * graph,
          struct ggml_tensor * wo,
          struct ggml_tensor * wo_b,
          struct ggml_tensor * q_cur,
@@ -4252,6 +4588,8 @@ static struct ggml_tensor * llm_build_kqv(
     struct ggml_tensor * cur = ggml_cont_2d(ctx, kqv_merged, n_embd_head_k*n_head, n_tokens);
     cb(cur, "kqv_merged_cont", il);
 
+    ggml_build_forward_expand(graph, cur);
+
     cur = ggml_mul_mat(ctx, wo, cur);
     if (wo_b) {
         cb(cur, "kqv_wo", il);
@@ -4264,17 +4602,56 @@ static struct ggml_tensor * llm_build_kqv(
     return cur;
 }
 
-struct llm_build_context {
-    const llama_model    & model;
-    const llama_hparams  & hparams;
-    const llama_cparams  & cparams;
-    const llama_batch    & batch;
-    const llama_kv_cache & kv_self;
-
-    const int64_t n_embd;
-    const int64_t n_layer;
-    const int64_t n_ctx;       // user-specified context size (can be different from n_ctx_train)
-    const int64_t n_head;
+static struct ggml_tensor * llm_build_kv(
+        struct ggml_context * ctx,
+          const llama_model & model,
+        const llama_hparams & hparams,
+       const llama_kv_cache & kv,
+         struct ggml_cgraph * graph,
+         struct ggml_tensor * wo,
+         struct ggml_tensor * wo_b,
+         struct ggml_tensor * k_cur,
+         struct ggml_tensor * v_cur,
+         struct ggml_tensor * q_cur,
+         struct ggml_tensor * kq_mask,
+                    int64_t   n_ctx,
+                    int32_t   n_tokens,
+                    int32_t   kv_head,
+                    int32_t   n_kv,
+                    float     max_alibi_bias,
+                    float     kq_scale,
+         const llm_build_cb & cb,
+                    int       il) {
+
+    // these nodes are added to the graph together so that they are not reordered
+    // by doing so, the number of splits in the graph is reduced
+    ggml_build_forward_expand(graph, q_cur);
+    ggml_build_forward_expand(graph, k_cur);
+    ggml_build_forward_expand(graph, v_cur);
+
+    llm_build_kv_store(ctx, hparams, kv, graph, k_cur, v_cur, n_ctx, n_tokens, kv_head, cb, il);
+
+    struct ggml_tensor * cur;
+    cur  = llm_build_kqv(ctx, model, hparams, kv, graph,
+            wo, wo_b,
+            q_cur, kq_mask, n_ctx, n_tokens, n_kv, max_alibi_bias, kq_scale, cb, il);
+    cb(cur, "kqv_out", il);
+
+    return cur;
+}
+
+struct llm_build_context {
+    const llama_model    & model;
+    const llama_context  & lctx;
+    const llama_hparams  & hparams;
+    const llama_cparams  & cparams;
+    const llama_batch    & batch;
+    const llama_kv_cache & kv_self;
+
+    const int64_t n_embd;
+    const int64_t n_layer;
+    const int64_t n_ctx;       // user-specified context size (can be different from n_ctx_train)
+    const int64_t n_head;
     const int64_t n_head_kv;
     const int64_t n_embd_head_k;
     const int64_t n_embd_k_gqa;
@@ -4312,6 +4689,7 @@ struct llm_build_context {
     const llm_build_cb & cb,
                   bool   worst_case) :
         model            (lctx.model),
+        lctx             (lctx),
         hparams          (model.hparams),
         cparams          (lctx.cparams),
         batch            (batch),
@@ -4372,20 +4750,20 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, lctx.inp_tokens, lctx.inp_embd, cb);
         cb(inpL, "inp_embd", -1);
 
         // inp_pos - contains the positions
-        struct ggml_tensor * inp_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens);
+        struct ggml_tensor * inp_pos = ggml_view_1d(ctx0, lctx.inp_pos, n_tokens, 0);
         cb(inp_pos, "inp_pos", -1);
 
         // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
-        struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1);
+        struct ggml_tensor * KQ_mask = ggml_view_2d(ctx0, lctx.inp_KQ_mask, n_kv, n_tokens, n_kv*ggml_type_size(lctx.inp_KQ_mask->type), 0);
         cb(KQ_mask, "KQ_mask", -1);
 
         // shift the entire K-cache if needed
         if (do_rope_shift) {
-            llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE, n_ctx, freq_base, freq_scale, cb);
+            llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, lctx.inp_K_shift, LLM_ROPE, n_ctx, freq_base, freq_scale, cb);
         }
 
         for (int il = 0; il < n_layer; ++il) {
@@ -4421,12 +4799,6 @@ struct llm_build_context {
                     cb(Vcur, "Vcur", il);
                 }
 
-                // these nodes are added to the graph together so that they are not reordered
-                // by doing so, the number of splits in the graph is reduced
-                ggml_build_forward_expand(gf, Qcur);
-                ggml_build_forward_expand(gf, Kcur);
-                ggml_build_forward_expand(gf, Vcur);
-
                 Qcur = ggml_rope_custom(
                     ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens), inp_pos,
                     hparams.n_rot, 0, 0, n_orig_ctx, freq_base, freq_scale,
@@ -4441,11 +4813,9 @@ struct llm_build_context {
                 );
                 cb(Kcur, "Kcur", il);
 
-                llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il);
-
-                cur = llm_build_kqv(ctx0, model, hparams, kv_self,
+                cur = llm_build_kv(ctx0, model, hparams, kv_self, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+                        Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
                 cb(cur, "kqv_out", il);
             }
 
@@ -4564,20 +4934,20 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, lctx.inp_tokens, lctx.inp_embd, cb);
         cb(inpL, "inp_embd", -1);
 
         // inp_pos - contains the positions
-        struct ggml_tensor * inp_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens);
+        struct ggml_tensor * inp_pos = ggml_view_1d(ctx0, lctx.inp_pos, n_tokens, 0);
         cb(inp_pos, "inp_pos", -1);
 
         // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
-        struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1);
+        struct ggml_tensor * KQ_mask = ggml_view_2d(ctx0, lctx.inp_KQ_mask, n_kv, n_tokens, n_kv*ggml_type_size(lctx.inp_KQ_mask->type), 0);
         cb(KQ_mask, "KQ_mask", -1);
 
         // shift the entire K-cache if needed
         if (do_rope_shift) {
-            llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE, n_ctx, freq_base, freq_scale, cb);
+            llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, lctx.inp_K_shift, LLM_ROPE, n_ctx, freq_base, freq_scale, cb);
         }
 
         for (int il = 0; il < n_layer; ++il) {
@@ -4622,14 +4992,13 @@ struct llm_build_context {
                 cb(Qcur, "Qcur", il);
                 cb(Kcur, "Kcur", il);
 
-                llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il);
 
                 // apply ALiBi for 13B model
                 const float max_alibi_bias = model.type == MODEL_13B ? 8.0f : -1.0f;
 
-                cur = llm_build_kqv(ctx0, model, hparams, kv_self,
+                cur = llm_build_kv(ctx0, model, hparams, kv_self, gf,
                         model.layers[il].wo, NULL,
-                        Qcur, KQ_mask, n_ctx, n_tokens, n_kv, max_alibi_bias, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+                        Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, max_alibi_bias, 1.0f/sqrtf(float(n_embd_head)), cb, il);
                 cb(cur, "kqv_out", il);
             }
 
@@ -4686,20 +5055,20 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, lctx.inp_tokens, lctx.inp_embd, cb);
         cb(inpL, "inp_embd", -1);
 
         // inp_pos - contains the positions
-        struct ggml_tensor * inp_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens);
+        struct ggml_tensor * inp_pos = ggml_view_1d(ctx0, lctx.inp_pos, n_tokens, 0);
         cb(inp_pos, "inp_pos", -1);
 
         // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
-        struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1);
+        struct ggml_tensor * KQ_mask = ggml_view_2d(ctx0, lctx.inp_KQ_mask, n_kv, n_tokens, n_kv*ggml_type_size(lctx.inp_KQ_mask->type), 0);
         cb(KQ_mask, "KQ_mask", -1);
 
         // shift the entire K-cache if needed
         if (do_rope_shift) {
-            llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE_NEOX, n_ctx, freq_base, freq_scale, cb);
+            llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, lctx.inp_K_shift, LLM_ROPE_NEOX, n_ctx, freq_base, freq_scale, cb);
         }
 
         for (int il = 0; il < n_layer; ++il) {
@@ -4751,11 +5120,9 @@ struct llm_build_context {
                 );
                 cb(Kcur, "Kcur", il);
 
-                llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il);
-
-                cur = llm_build_kqv(ctx0, model, hparams, kv_self,
+                cur = llm_build_kv(ctx0, model, hparams, kv_self, gf,
                         model.layers[il].wo, NULL,
-                        Qcur, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+                        Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
                 cb(cur, "kqv_out", il);
             }
 
@@ -4810,15 +5177,15 @@ struct llm_build_context {
         struct ggml_tensor * pos;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, lctx.inp_tokens, lctx.inp_embd, cb);
         cb(inpL, "inp_embd", -1);
 
         // inp_pos - contains the positions
-        struct ggml_tensor * inp_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens);
+        struct ggml_tensor * inp_pos = ggml_view_1d(ctx0, lctx.inp_pos, n_tokens, 0);
         cb(inp_pos, "inp_pos", -1);
 
         // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
-        struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1);
+        struct ggml_tensor * KQ_mask = ggml_view_2d(ctx0, lctx.inp_KQ_mask, n_kv, n_tokens, n_kv*ggml_type_size(lctx.inp_KQ_mask->type), 0);
         cb(KQ_mask, "KQ_mask", -1);
 
         pos = ggml_get_rows(ctx0, model.pos_embd, inp_pos);
@@ -4852,11 +5219,9 @@ struct llm_build_context {
 
                 Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
 
-                llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il);
-
-                cur = llm_build_kqv(ctx0, model, hparams, kv_self,
+                cur = llm_build_kv(ctx0, model, hparams, kv_self, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+                        Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
                 cb(cur, "kqv_out", il);
             }
 
@@ -4909,19 +5274,19 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, lctx.inp_tokens, lctx.inp_embd, cb);
         cb(inpL, "inp_embd", -1);
 
         // inp_pos - contains the positions
-        struct ggml_tensor * inp_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens);
+        struct ggml_tensor * inp_pos = ggml_view_1d(ctx0, lctx.inp_pos, n_tokens, 0);
         cb(inp_pos, "inp_pos", -1);
 
         // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
-        struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1);
+        struct ggml_tensor * KQ_mask = ggml_view_2d(ctx0, lctx.inp_KQ_mask, n_kv, n_tokens, n_kv*ggml_type_size(lctx.inp_KQ_mask->type), 0);
         cb(KQ_mask, "KQ_mask", -1);
 
         if (do_rope_shift) {
-            llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE_NEOX, n_ctx, freq_base, freq_scale, cb);
+            llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, lctx.inp_K_shift, LLM_ROPE_NEOX, n_ctx, freq_base, freq_scale, cb);
         }
 
         for (int il = 0; il < n_layer; ++il) {
@@ -5059,12 +5424,9 @@ struct llm_build_context {
                         );
                 cb(Vcur, "Vcur", il);
 
-                llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il);
-
-                // TODO: not tested, could be broken
-                cur = llm_build_kqv(ctx0, model, hparams, kv_self,
+                cur = llm_build_kv(ctx0, model, hparams, kv_self, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Q, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+                        Kcur, Vcur, Q, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
                 cb(cur, "kqv_out", il);
             }
 
@@ -5119,11 +5481,11 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, lctx.inp_tokens, lctx.inp_embd, cb);
         cb(inpL, "inp_embd", -1);
 
         // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
-        struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1);
+        struct ggml_tensor * KQ_mask = ggml_view_2d(ctx0, lctx.inp_KQ_mask, n_kv, n_tokens, n_kv*ggml_type_size(lctx.inp_KQ_mask->type), 0);
         cb(KQ_mask, "KQ_mask", -1);
 
         for (int il = 0; il < n_layer; ++il) {
@@ -5151,11 +5513,9 @@ struct llm_build_context {
                 Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
                 cb(Qcur, "Qcur", il);
 
-                llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il);
-
-                cur = llm_build_kqv(ctx0, model, hparams, kv_self,
+                cur = llm_build_kv(ctx0, model, hparams, kv_self, gf,
                         model.layers[il].wo, NULL,
-                        Qcur, KQ_mask, n_ctx, n_tokens, n_kv, 8.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+                        Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, 8.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
                 cb(cur, "kqv_out", il);
             }
 
@@ -5211,11 +5571,11 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, lctx.inp_tokens, lctx.inp_embd, cb);
         cb(inpL, "inp_embd", -1);
 
         // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
-        struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1);
+        struct ggml_tensor * KQ_mask = ggml_view_2d(ctx0, lctx.inp_KQ_mask, n_kv, n_tokens, n_kv*ggml_type_size(lctx.inp_KQ_mask->type), 0);
         cb(KQ_mask, "KQ_mask", -1);
 
         inpL = llm_build_norm(ctx0, inpL, hparams,
@@ -5249,11 +5609,9 @@ struct llm_build_context {
 
                 Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
 
-                llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il);
-
-                cur = llm_build_kqv(ctx0, model, hparams, kv_self,
+                cur = llm_build_kv(ctx0, model, hparams, kv_self, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, KQ_mask, n_ctx, n_tokens, n_kv, 8.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+                        Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, 8.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
                 cb(cur, "kqv_out", il);
             }
 
@@ -5306,11 +5664,11 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, lctx.inp_tokens, lctx.inp_embd, cb);
         cb(inpL, "inp_embd", -1);
 
         // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
-        struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1);
+        struct ggml_tensor * KQ_mask = ggml_view_2d(ctx0, lctx.inp_KQ_mask, n_kv, n_tokens, n_kv*ggml_type_size(lctx.inp_KQ_mask->type), 0);
         cb(KQ_mask, "KQ_mask", -1);
 
         for (int il = 0; il < n_layer; ++il) {
@@ -5344,11 +5702,9 @@ struct llm_build_context {
 
                 Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
 
-                llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il);
-
-                cur = llm_build_kqv(ctx0, model, hparams, kv_self,
+                cur = llm_build_kv(ctx0, model, hparams, kv_self, gf,
                         model.layers[il].wo, NULL,
-                        Qcur, KQ_mask, n_ctx, n_tokens, n_kv, hparams.f_max_alibi_bias, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+                        Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, hparams.f_max_alibi_bias, 1.0f/sqrtf(float(n_embd_head)), cb, il);
                 cb(cur, "kqv_out", il);
             }
 
@@ -5404,20 +5760,20 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, lctx.inp_tokens, lctx.inp_embd, cb);
         cb(inpL, "inp_embd", -1);
 
         // inp_pos - contains the positions
-        struct ggml_tensor * inp_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens);
+        struct ggml_tensor * inp_pos = ggml_view_1d(ctx0, lctx.inp_pos, n_tokens, 0);
         cb(inp_pos, "inp_pos", -1);
 
         // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
-        struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1);
+        struct ggml_tensor * KQ_mask = ggml_view_2d(ctx0, lctx.inp_KQ_mask, n_kv, n_tokens, n_kv*ggml_type_size(lctx.inp_KQ_mask->type), 0);
         cb(KQ_mask, "KQ_mask", -1);
 
         // shift the entire K-cache if needed
         if (do_rope_shift) {
-            llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE_NEOX, n_ctx, freq_base, freq_scale, cb);
+            llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, lctx.inp_K_shift, LLM_ROPE_NEOX, n_ctx, freq_base, freq_scale, cb);
         }
 
         for (int il = 0; il < n_layer; ++il) {
@@ -5435,12 +5791,24 @@ struct llm_build_context {
                 // compute Q and K and RoPE them
                 struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].wq, cur);
                 cb(Qcur, "Qcur", il);
+                if (model.layers[il].bq) {
+                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                    cb(Qcur, "Qcur", il);
+                }
 
                 struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, model.layers[il].wk, cur);
                 cb(Kcur, "Kcur", il);
+                if (model.layers[il].bk) {
+                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                    cb(Kcur, "Kcur", il);
+                }
 
                 struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv, cur);
                 cb(Vcur, "Vcur", il);
+                if (model.layers[il].bv) {
+                    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,
@@ -5456,11 +5824,9 @@ struct llm_build_context {
                 );
                 cb(Kcur, "Kcur", il);
 
-                llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il);
-
-                cur = llm_build_kqv(ctx0, model, hparams, kv_self,
+                cur = llm_build_kv(ctx0, model, hparams, kv_self, gf,
                         model.layers[il].wo, NULL,
-                        Qcur, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+                        Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
                 cb(cur, "kqv_out", il);
             }
 
@@ -5517,20 +5883,20 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, lctx.inp_tokens, lctx.inp_embd, cb);
         cb(inpL, "inp_embd", -1);
 
         // inp_pos - contains the positions
-        struct ggml_tensor * inp_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens);
+        struct ggml_tensor * inp_pos = ggml_view_1d(ctx0, lctx.inp_pos, n_tokens, 0);
         cb(inp_pos, "inp_pos", -1);
 
         // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
-        struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1);
+        struct ggml_tensor * KQ_mask = ggml_view_2d(ctx0, lctx.inp_KQ_mask, n_kv, n_tokens, n_kv*ggml_type_size(lctx.inp_KQ_mask->type), 0);
         cb(KQ_mask, "KQ_mask", -1);
 
         // shift the entire K-cache if needed
         if (do_rope_shift) {
-            llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE_NEOX, n_ctx, freq_base, freq_scale, cb);
+            llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, lctx.inp_K_shift, LLM_ROPE_NEOX, n_ctx, freq_base, freq_scale, cb);
         }
 
         for (int il = 0; il < n_layer; ++il) {
@@ -5573,11 +5939,9 @@ struct llm_build_context {
                 );
                 cb(Kcur, "Kcur", il);
 
-                llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il);
-
-                cur = llm_build_kqv(ctx0, model, hparams, kv_self,
+                cur = llm_build_kv(ctx0, model, hparams, kv_self, gf,
                         model.layers[il].wo, NULL,
-                        Qcur, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+                        Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
                 cb(cur, "kqv_out", il);
             }
 
@@ -5622,6 +5986,126 @@ struct llm_build_context {
 
         return gf;
     }
+
+    struct ggml_cgraph * build_qwen2() {
+        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 == hparams.n_rot);
+
+        struct ggml_tensor * cur;
+        struct ggml_tensor * inpL;
+
+        inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, lctx.inp_tokens, lctx.inp_embd, cb);
+        cb(inpL, "inp_embd", -1);
+
+        // inp_pos - contains the positions
+        struct ggml_tensor * inp_pos = ggml_view_1d(ctx0, lctx.inp_pos, n_tokens, 0);
+        cb(inp_pos, "inp_pos", -1);
+
+        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+        struct ggml_tensor * KQ_mask = ggml_view_2d(ctx0, lctx.inp_KQ_mask, n_kv, n_tokens, n_kv*ggml_type_size(lctx.inp_KQ_mask->type), 0);
+        cb(KQ_mask, "KQ_mask", -1);
+
+        // shift the entire K-cache if needed
+        if (do_rope_shift) {
+            llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, lctx.inp_K_shift, LLM_ROPE_NEOX, n_ctx, freq_base, freq_scale, cb);
+        }
+
+        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);
+                Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                cb(Qcur, "Qcur", il);
+
+                struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, model.layers[il].wk, cur);
+                cb(Kcur, "Kcur", il);
+                Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                cb(Kcur, "Kcur", il);
+
+                struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv, cur);
+                cb(Vcur, "Vcur", il);
+                Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                cb(Vcur, "Vcur", il);
+
+                // these nodes are added to the graph together so that they are not reordered
+                // by doing so, the number of splits in the graph is reduced
+                ggml_build_forward_expand(gf, Qcur);
+                ggml_build_forward_expand(gf, Kcur);
+                ggml_build_forward_expand(gf, Vcur);
+
+                Qcur = ggml_rope_custom(
+                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens), inp_pos,
+                    hparams.n_rot, 2, 0, n_orig_ctx, 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,
+                    hparams.n_rot, 2, 0, n_orig_ctx, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Kcur, "Kcur", il);
+
+                cur = llm_build_kv(ctx0, model, hparams, kv_self, gf,
+                        model.layers[il].wo, model.layers[il].bo,
+                        Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+                cb(cur, "kqv_out", il);
+            }
+
+            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
+        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_phi2() {
         struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
 
@@ -5634,20 +6118,20 @@ struct llm_build_context {
         struct ggml_tensor * ffn_output;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, lctx.inp_tokens, lctx.inp_embd, cb);
         cb(inpL, "inp_embd", -1);
 
         // inp_pos - contains the positions
-        struct ggml_tensor * inp_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens);
+        struct ggml_tensor * inp_pos = ggml_view_1d(ctx0, lctx.inp_pos, n_tokens, 0);
         cb(inp_pos, "inp_pos", -1);
 
         // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
-        struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1);
+        struct ggml_tensor * KQ_mask = ggml_view_2d(ctx0, lctx.inp_KQ_mask, n_kv, n_tokens, n_kv*ggml_type_size(lctx.inp_KQ_mask->type), 0);
         cb(KQ_mask, "KQ_mask", -1);
 
         // shift the entire K-cache if needed
         if (do_rope_shift) {
-            llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE_NEOX, n_ctx, freq_base, freq_scale, cb);
+            llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, lctx.inp_K_shift, LLM_ROPE_NEOX, n_ctx, freq_base, freq_scale, cb);
         }
 
         for (int il = 0; il < n_layer; ++il) {
@@ -5703,11 +6187,9 @@ struct llm_build_context {
                 );
                 cb(Kcur, "Kcur", il);
 
-                llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il);
-
-                cur = llm_build_kqv(ctx0, model, hparams, kv_self,
+                cur = llm_build_kv(ctx0, model, hparams, kv_self, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, 1.0f, cb, il);
+                        Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, -1.0f, 1.0f, cb, il);
                 cb(cur, "kqv_out", il);
             }
 
@@ -5758,20 +6240,20 @@ struct llm_build_context {
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, lctx.inp_tokens, lctx.inp_embd, cb);
         cb(inpL, "inp_embd", -1);
 
         // inp_pos - contains the positions
-        struct ggml_tensor * inp_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens);
+        struct ggml_tensor * inp_pos = ggml_view_1d(ctx0, lctx.inp_pos, n_tokens, 0);
         cb(inp_pos, "inp_pos", -1);
 
         // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
-        struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1);
+        struct ggml_tensor * KQ_mask = ggml_view_2d(ctx0, lctx.inp_KQ_mask, n_kv, n_tokens, n_kv*ggml_type_size(lctx.inp_KQ_mask->type), 0);
         cb(KQ_mask, "KQ_mask", -1);
 
         // shift the entire K-cache if needed
         if (do_rope_shift) {
-            llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE, n_ctx, freq_base, freq_scale, cb);
+            llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, lctx.inp_K_shift, LLM_ROPE, n_ctx, freq_base, freq_scale, cb);
         }
 
         for (int il = 0; il < n_layer; ++il) {
@@ -5808,11 +6290,9 @@ struct llm_build_context {
                         ext_factor, attn_factor, beta_fast, beta_slow);
                 cb(Kcur, "Kcur", il);
 
-                llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il);
-
-                cur = llm_build_kqv(ctx0, model, hparams, kv_self,
+                cur = llm_build_kv(ctx0, model, hparams, kv_self, gf,
                         model.layers[il].wo, NULL,
-                        Qcur, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+                        Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
                 cb(cur, "kqv_out", il);
             }
             struct ggml_tensor * sa_out = cur;
@@ -5867,15 +6347,15 @@ struct llm_build_context {
         struct ggml_tensor * pos;
         struct ggml_tensor * inpL;
 
-        inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, cb);
+        inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, lctx.inp_tokens, lctx.inp_embd, cb);
         cb(inpL, "inp_embd", -1);
 
         // inp_pos - contains the positions
-        struct ggml_tensor * inp_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens);
+        struct ggml_tensor * inp_pos = ggml_view_1d(ctx0, lctx.inp_pos, n_tokens, 0);
         cb(inp_pos, "inp_pos", -1);
 
         // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
-        struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1);
+        struct ggml_tensor * KQ_mask = ggml_view_2d(ctx0, lctx.inp_KQ_mask, n_kv, n_tokens, n_kv*ggml_type_size(lctx.inp_KQ_mask->type), 0);
         cb(KQ_mask, "KQ_mask", -1);
 
         pos = ggml_get_rows(ctx0, model.pos_embd, inp_pos);
@@ -5903,51 +6383,396 @@ struct llm_build_context {
                 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(Qcur, "Qcur", il);
+                cb(Kcur, "Kcur", il);
+                cb(Vcur, "Vcur", il);
+
+                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+
+                cur = llm_build_kv(ctx0, model, hparams, kv_self, gf,
+                        model.layers[il].wo, model.layers[il].bo,
+                        Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+                cb(cur, "kqv_out", il);
+            }
+
+            // add the input
+            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
+            cb(ffn_inp, "ffn_inp", il);
+
+            // FF
+            {
+                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_codeshell() {
+        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);
+        GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+        struct ggml_tensor * cur;
+        struct ggml_tensor * inpL;
+
+        inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, lctx.inp_tokens, lctx.inp_embd, cb);
+        cb(inpL, "inp_embd", -1);
+
+        // inp_pos - contains the positions
+        struct ggml_tensor * inp_pos = ggml_view_1d(ctx0, lctx.inp_pos, n_tokens, 0);
+        cb(inp_pos, "inp_pos", -1);
+
+        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+        struct ggml_tensor * KQ_mask = ggml_view_2d(ctx0, lctx.inp_KQ_mask, n_kv, n_tokens, n_kv*ggml_type_size(lctx.inp_KQ_mask->type), 0);
+        cb(KQ_mask, "KQ_mask", -1);
+
+        // shift the entire K-cache if needed
+        if (do_rope_shift) {
+            llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, lctx.inp_K_shift, LLM_ROPE, n_ctx, freq_base, freq_scale, cb);
+        }
+
+        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 * tmpq = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd,     n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd)));
+                struct ggml_tensor * tmpk = 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(tmpq, "tmpq", il);
+                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,
+                    hparams.n_rot, 2, 0, n_orig_ctx, 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,
+                    hparams.n_rot, 2, 0, n_orig_ctx, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Kcur, "Kcur", il);
+
+                cur = llm_build_kv(ctx0, model, hparams, kv_self, gf,
+                        model.layers[il].wo, model.layers[il].bo,
+                        Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+                cb(cur, "kqv_out", il);
+            }
+
+            // add the input
+            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
+            cb(ffn_inp, "ffn_inp", il);
+
+            // FF
+            {
+                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_orion() {
+        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 == hparams.n_rot);
+
+        struct ggml_tensor * cur;
+        struct ggml_tensor * inpL;
+
+        inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, lctx.inp_tokens, lctx.inp_embd, cb);
+        cb(inpL, "inp_embd", -1);
+
+        // inp_pos - contains the positions
+        struct ggml_tensor * inp_pos = ggml_view_1d(ctx0, lctx.inp_pos, n_tokens, 0);
+        cb(inp_pos, "inp_pos", -1);
+
+        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+        struct ggml_tensor * KQ_mask = ggml_view_2d(ctx0, lctx.inp_KQ_mask, n_kv, n_tokens, n_kv*ggml_type_size(lctx.inp_KQ_mask->type), 0);
+        cb(KQ_mask, "KQ_mask", -1);
+
+        // shift the entire K-cache if needed
+        if (do_rope_shift) {
+            llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, lctx.inp_K_shift, LLM_ROPE, n_ctx, freq_base, freq_scale, cb);
+        }
+
+        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, model.layers[il].attn_norm_b,
+                    LLM_NORM, 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);
+                // if (model.layers[il].bq) {
+                //     Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                //     cb(Qcur, "Qcur", il);
+                // }
+
+                struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, model.layers[il].wk, cur);
+                cb(Kcur, "Kcur", il);
+                // if (model.layers[il].bk) {
+                //     Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                //     cb(Kcur, "Kcur", il);
+                // }
+
+                struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv, cur);
+                cb(Vcur, "Vcur", il);
+                // if (model.layers[il].bv) {
+                //     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,
+                    hparams.n_rot, 2, 0, n_orig_ctx, 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,
+                    hparams.n_rot, 2, 0, n_orig_ctx, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Kcur, "Kcur", il);
+
+                cur = llm_build_kv(ctx0, model, hparams, kv_self, gf,
+                        model.layers[il].wo, NULL,
+                        Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+                cb(cur, "kqv_out", il);
+            }
+
+            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);
+            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, model.output_norm_b,
+                LLM_NORM, 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_internlm2() {
+        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 == hparams.n_rot);
+
+        struct ggml_tensor * cur;
+        struct ggml_tensor * inpL;
+
+        inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, lctx.inp_tokens, lctx.inp_embd, cb);
+        cb(inpL, "inp_embd", -1);
+
+        // inp_pos - contains the positions
+        struct ggml_tensor * inp_pos = ggml_view_1d(ctx0, lctx.inp_pos, n_tokens, 0);
+        cb(inp_pos, "inp_pos", -1);
+
+        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+        struct ggml_tensor * KQ_mask = ggml_view_2d(ctx0, lctx.inp_KQ_mask, n_kv, n_tokens, n_kv*ggml_type_size(lctx.inp_KQ_mask->type), 0);
+        cb(KQ_mask, "KQ_mask", -1);
+
+        // shift the entire K-cache if needed
+        if (do_rope_shift) {
+            llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, lctx.inp_K_shift, LLM_ROPE, n_ctx, freq_base, freq_scale, cb);
+        }
+
+        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);
+                if (model.layers[il].bq) {
+                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                    cb(Qcur, "Qcur", il);
+                }
+
+                struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, model.layers[il].wk, cur);
                 cb(Kcur, "Kcur", il);
+                if (model.layers[il].bk) {
+                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                    cb(Kcur, "Kcur", il);
+                }
+
+                struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv, cur);
                 cb(Vcur, "Vcur", il);
+                if (model.layers[il].bv) {
+                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                    cb(Vcur, "Vcur", il);
+                }
 
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+                Qcur = ggml_rope_custom(
+                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens), inp_pos,
+                    hparams.n_rot, 0, 0, n_orig_ctx, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Qcur, "Qcur", il);
 
-                llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il);
+                Kcur = ggml_rope_custom(
+                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos,
+                    hparams.n_rot, 0, 0, n_orig_ctx, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Kcur, "Kcur", il);
 
-                cur = llm_build_kqv(ctx0, model, hparams, kv_self,
+                cur = llm_build_kv(ctx0, model, hparams, kv_self, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+                        Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
                 cb(cur, "kqv_out", il);
             }
 
-            // add the input
-            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
+            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
             cb(ffn_inp, "ffn_inp", il);
 
-            // FF
-            {
-                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);
+            // 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,   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 = 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);
 
-            inpL = ggml_add(ctx0, cur, ffn_inp);
-            cb(inpL, "l_out", il);
+            cur = ggml_add(ctx0, cur, ffn_inp);
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
         }
 
-        cur = llm_build_norm(ctx0, inpL, hparams,
-                model.output_norm,
-                model.output_norm_b,
-                LLM_NORM, cb, -1);
+        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);
 
@@ -5955,6 +6780,7 @@ struct llm_build_context {
 
         return gf;
     }
+
 };
 
 static struct ggml_cgraph * llama_build_graph(
@@ -5965,15 +6791,7 @@ static struct ggml_cgraph * llama_build_graph(
     // check if we should build the worst-case graph (for memory measurement)
     const bool worst_case = ggml_tallocr_is_measure(lctx.alloc);
 
-    // keep track of the input that has already been allocated
-    bool alloc_inp_tokens   = false;
-    bool alloc_inp_embd     = false;
-    bool alloc_inp_pos      = false;
-    bool alloc_inp_KQ_mask  = false;
-    bool alloc_inp_K_shift  = false;
-
     // this callback allows us to apply custom logic to each tensor (e.g. ggml-alloc, offloading, etc.)
-    // TODO: improve handling of input and output tensors, then replace this with ggml_set_name
     llm_build_cb cb = [&](struct ggml_tensor * cur, const char * name, int il) {
         if (il >= 0) {
             ggml_format_name(cur, "%s-%d", name, il);
@@ -5981,118 +6799,78 @@ static struct ggml_cgraph * llama_build_graph(
             ggml_set_name(cur, name);
         }
 
-        //
-        // allocate input tensors and set input data
-        //
-
-        if (!alloc_inp_tokens && strcmp(name, "inp_tokens") == 0) {
-            ggml_tallocr_alloc(lctx.alloc, cur);
-
-            if (!ggml_tallocr_is_measure(lctx.alloc) && batch.token) {
-                const int64_t n_tokens = cur->ne[0];
-
-                ggml_backend_tensor_set(cur, batch.token, 0, n_tokens*ggml_element_size(cur));
+        if (!lctx.cparams.offload_kqv) {
+            if (strcmp(name, "kqv_merged_cont") == 0) {
+                // all nodes between the KV store and the attention output are run on the CPU
+                ggml_backend_sched_set_node_backend(lctx.sched, cur, lctx.backend_cpu);
             }
-
-            alloc_inp_tokens = true;
         }
+    };
 
-        if (!alloc_inp_embd && strcmp(name, "inp_embd") == 0 && batch.embd) {
-            ggml_tallocr_alloc(lctx.alloc, cur);
+    struct ggml_cgraph * result = NULL;
 
-            if (!ggml_tallocr_is_measure(lctx.alloc) && batch.embd) {
-                const int64_t n_embd   = cur->ne[0];
-                const int64_t n_tokens = cur->ne[1];
+    struct llm_build_context llm(lctx, batch, cb, worst_case);
 
-                ggml_backend_tensor_set(cur, batch.embd, 0, n_tokens*n_embd*ggml_element_size(cur));
-            }
+    //
+    // set input data
+    //
+
+    if (!ggml_tallocr_is_measure(lctx.alloc)) {
+        if (batch.token) {
+            const int64_t n_tokens = batch.n_tokens;
 
-            alloc_inp_embd = true;
+            ggml_backend_tensor_set(lctx.inp_tokens, batch.token, 0, n_tokens*ggml_element_size(lctx.inp_tokens));
         }
 
-        if (!alloc_inp_pos && strcmp(name, "inp_pos") == 0) {
-            ggml_tallocr_alloc(lctx.alloc, cur);
+        if (batch.embd) {
+            const int64_t n_embd   = llm.n_embd;
+            const int64_t n_tokens = batch.n_tokens;
 
-            if (!ggml_tallocr_is_measure(lctx.alloc) && batch.pos) {
-                const int64_t n_tokens = cur->ne[0];
+            ggml_backend_tensor_set(lctx.inp_embd, batch.embd, 0, n_tokens*n_embd*ggml_element_size(lctx.inp_embd));
+        }
 
-                static_assert(std::is_same<llama_pos, int32_t>::value, "llama_pos must be int32_t");
-                ggml_backend_tensor_set(cur, batch.pos, 0, n_tokens*ggml_element_size(cur));
-            }
+        if (batch.pos) {
+            const int64_t n_tokens = batch.n_tokens;
 
-            alloc_inp_pos = true;
+            ggml_backend_tensor_set(lctx.inp_pos, batch.pos, 0, n_tokens*ggml_element_size(lctx.inp_pos));
         }
 
-        if (!alloc_inp_KQ_mask && strcmp(name, "KQ_mask") == 0) {
-            ggml_tallocr_alloc(lctx.alloc, cur);
+        {
+            const int64_t n_kv     = llm.n_kv;
+            const int64_t n_tokens = batch.n_tokens;
 
-            if (!ggml_tallocr_is_measure(lctx.alloc)) {
-                const int64_t n_kv     = cur->ne[0];
-                const int64_t n_tokens = cur->ne[1];
+            GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_KQ_mask->buffer));
+            float * data = (float *) lctx.inp_KQ_mask->data;
 
-                float * data;
-                if (ggml_backend_buffer_is_host(cur->buffer)) {
-                    data = (float *) cur->data;
-                } else {
-                    lctx.buf_copy.resize(ggml_nbytes(cur));
-                    data = (float *) lctx.buf_copy.data();
-                }
+            for (int h = 0; h < 1; ++h) {
+                for (int j = 0; j < n_tokens; ++j) {
+                    const llama_pos    pos    = batch.pos[j];
+                    const llama_seq_id seq_id = batch.seq_id[j][0];
 
-                for (int h = 0; h < 1; ++h) {
-                    for (int j = 0; j < n_tokens; ++j) {
-                        const llama_pos    pos    = batch.pos[j];
-                        const llama_seq_id seq_id = batch.seq_id[j][0];
-
-                        for (int i = 0; i < n_kv; ++i) {
-                            float f;
-                            if (!lctx.kv_self.cells[i].has_seq_id(seq_id) || lctx.kv_self.cells[i].pos > pos) {
-                                f = -INFINITY;
-                            } else {
-                                f = 0;
-                            }
-                            data[h*(n_kv*n_tokens) + j*n_kv + i] = f;
+                    for (int i = 0; i < n_kv; ++i) {
+                        float f;
+                        if (!lctx.kv_self.cells[i].has_seq_id(seq_id) || lctx.kv_self.cells[i].pos > pos) {
+                            f = -INFINITY;
+                        } else {
+                            f = 0;
                         }
+                        data[h*(n_kv*n_tokens) + j*n_kv + i] = f;
                     }
                 }
-
-                if (data != cur->data) {
-                    ggml_backend_tensor_set(cur, data, 0, ggml_nbytes(cur));
-                }
             }
-
-            alloc_inp_KQ_mask = true;
         }
 
-        if (!alloc_inp_K_shift && strcmp(name, "K_shift") == 0) {
-            ggml_tallocr_alloc(lctx.alloc, cur);
-
-            if (!ggml_tallocr_is_measure(lctx.alloc)) {
-                const int64_t n_ctx = cur->ne[0];
-
-                int32_t * data;
-                if (ggml_backend_buffer_is_host(cur->buffer)) {
-                    data = (int32_t *) cur->data;
-                } else {
-                    lctx.buf_copy.resize(ggml_nbytes(cur));
-                    data = (int32_t *) lctx.buf_copy.data();
-                }
+        if (llm.do_rope_shift) {
+            const int64_t n_ctx = llm.n_ctx;
 
-                for (int i = 0; i < n_ctx; ++i) {
-                    data[i] = lctx.kv_self.cells[i].delta;
-                }
+            GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_K_shift->buffer));
+            int32_t * data = (int32_t *) lctx.inp_K_shift->data;
 
-                if (data != cur->data) {
-                    ggml_backend_tensor_set(cur, data, 0, ggml_nbytes(cur));
-                }
+            for (int i = 0; i < n_ctx; ++i) {
+                data[i] = lctx.kv_self.cells[i].delta;
             }
-
-            alloc_inp_K_shift = true;
         }
-    };
-
-    struct ggml_cgraph * result = NULL;
-
-    struct llm_build_context llm(lctx, batch, cb, worst_case);
+    }
 
     llm.init();
 
@@ -6137,6 +6915,10 @@ static struct ggml_cgraph * llama_build_graph(
             {
                 result = llm.build_qwen();
             } break;
+        case LLM_ARCH_QWEN2:
+            {
+                result = llm.build_qwen2();
+            } break;
         case LLM_ARCH_PHI2:
             {
                 result = llm.build_phi2();
@@ -6149,6 +6931,18 @@ static struct ggml_cgraph * llama_build_graph(
             {
                 result = llm.build_gpt2();
             } break;
+        case LLM_ARCH_CODESHELL:
+            {
+                result = llm.build_codeshell();
+            } break;
+        case LLM_ARCH_ORION:
+            {
+                result = llm.build_orion();
+            } break;
+        case LLM_ARCH_INTERNLM2:
+            {
+                result = llm.build_internlm2();
+            } break;
         default:
             GGML_ASSERT(false);
     }
@@ -6254,6 +7048,7 @@ static int llama_decode_internal(
     //printf("kv_self.n = %5d, kv_self.used = %5d, kv_self.head = %5d\n", kv_self.n, kv_self.used, kv_self.head);
 
     ggml_backend_sched_reset(lctx.sched);
+    ggml_backend_sched_set_eval_callback(lctx.sched, lctx.cparams.cb_eval, lctx.cparams.cb_eval_user_data);
 
     ggml_cgraph * gf = llama_build_graph(lctx, batch);
 
@@ -6279,11 +7074,6 @@ static int llama_decode_internal(
         n_threads = std::min(4, n_threads);
     }
 
-    const bool fully_offloaded = model.n_gpu_layers >= (int) hparams.n_layer + 1;
-    if (ggml_cpu_has_cublas() && fully_offloaded) {
-        n_threads = 1;
-    }
-
 #ifdef GGML_USE_MPI
     const int64_t n_layer = hparams.n_layer;
     ggml_mpi_graph_compute_pre(lctx.ctx_mpi, gf, n_layer);
@@ -7095,7 +7885,9 @@ static std::vector<llama_vocab::id> llama_tokenize_internal(const llama_vocab &
                         //
                         auto raw_text = fragment.raw_text.substr(fragment.offset, fragment.length);
                         if (&fragment == &fragment_buffer.front()) {
-                            raw_text = " " + raw_text; // prefix with space if the first token is not special
+                            if (vocab.add_space_prefix) {
+                                raw_text = " " + raw_text; // prefix with space if the first token is not special
+                            }
                         }
 
 #ifdef PRETOKENIZERDEBUG
@@ -7574,6 +8366,11 @@ void llama_sample_softmax(struct llama_context * ctx, llama_token_data_array * c
 }
 
 void llama_sample_top_k(struct llama_context * ctx, llama_token_data_array * candidates, int32_t k, size_t min_keep) {
+    // TODO: move bucket sort to separate function so that top_p/tail_free/typical/softmax first is equally fast
+    // if (k >= (int32_t)candidates->size) {
+    //     return;
+    // }
+
     const int64_t t_start_sample_us = ggml_time_us();
 
     k = std::max(k, (int) min_keep);
@@ -7584,10 +8381,57 @@ void llama_sample_top_k(struct llama_context * ctx, llama_token_data_array * can
         auto comp = [](const llama_token_data & a, const llama_token_data & b) {
             return a.logit > b.logit;
         };
-        if (k == (int) candidates->size) {
-            std::sort(candidates->data, candidates->data + candidates->size, comp);
-        } else {
+        if (k <= 128) {
             std::partial_sort(candidates->data, candidates->data + k, candidates->data + candidates->size, comp);
+        } else {
+            constexpr int   nbuckets     = 128;
+            constexpr float bucket_low   = -10.0f;
+            constexpr float bucket_high  =  10.0f;
+            constexpr float bucket_scale = nbuckets/(bucket_high - bucket_low);
+            constexpr float bucker_inter = -bucket_low * bucket_scale;
+
+            std::vector<int> bucket_idx(candidates->size);
+            std::vector<int> histo(nbuckets, 0);
+
+            for (int i = 0; i < (int)candidates->size; ++i) {
+                const float val = candidates->data[i].logit;
+                int ib = int(bucket_scale * val + bucker_inter); //nbuckets * (val - bucket_low) / (bucket_high - bucket_low);
+                ib = std::max(0, std::min(nbuckets-1, ib));
+                bucket_idx[i] = ib;
+                ++histo[ib];
+            }
+            int nhave = 0;
+            int ib = nbuckets - 1;
+            for ( ; ib >= 0; --ib) {
+                nhave += histo[ib];
+                if (nhave >= k) break;
+            }
+            std::vector<llama_token_data> tmp_tokens(nhave);
+            auto ptr = tmp_tokens.data();
+            std::vector<llama_token_data*> bucket_ptrs;
+            bucket_ptrs.reserve(nbuckets - ib);
+            for (int j = nbuckets - 1; j >= ib; --j) {
+                bucket_ptrs.push_back(ptr);
+                ptr += histo[j];
+            }
+            for (int i = 0; i < (int)candidates->size; ++i) {
+                int j = bucket_idx[i];
+                if (j >= ib) {
+                    *bucket_ptrs[nbuckets-1-j]++ = candidates->data[i];
+                }
+            }
+
+            ptr = tmp_tokens.data();
+            int ndone = 0;
+            for (int j = nbuckets-1; j > ib; --j) {
+                std::sort(ptr, ptr + histo[j], comp);
+                ptr += histo[j];
+                ndone += histo[j];
+            }
+            std::partial_sort(ptr, ptr + k - ndone, ptr + histo[ib], comp);
+
+            std::memcpy(candidates->data, tmp_tokens.data(), k*sizeof(llama_token_data));
+
         }
         candidates->sorted = true;
     }
@@ -7635,21 +8479,56 @@ void llama_sample_min_p(struct llama_context * ctx, llama_token_data_array * can
         return;
     }
 
-    llama_sample_softmax(ctx, candidates);
-
     const int64_t t_start_sample_us = ggml_time_us();
 
-    float scale = candidates->data[0].p; // scale by max prob
-    size_t i = 1; // first token always matches
+    bool min_p_applied = false;
+
+    // if the candidates aren't sorted, try the unsorted implementation first
+    if (!candidates->sorted) {
+        std::vector<llama_token_data> filtered_tokens;
+
+        float max_logit = -FLT_MAX;
+        for (size_t i = 0; i < candidates->size; ++i) {
+            max_logit = std::max(max_logit, candidates->data[i].logit);
+        }
+        const float min_logit = max_logit + logf(p); // min logit for p_i >= p * p_max
+
+        for (size_t i = 0; i < candidates->size; ++i) {
+            if (candidates->data[i].logit >= min_logit) {
+                filtered_tokens.push_back(candidates->data[i]);
+            }
+        }
 
-    for (; i < candidates->size; ++i) {
-        if (candidates->data[i].p < p * scale && i >= min_keep) {
-            break; // prob too small
+        // if we have enough values the operation was a success
+        if (filtered_tokens.size() >= min_keep) {
+            memcpy(candidates->data, filtered_tokens.data(), filtered_tokens.size()*sizeof(llama_token_data));
+            candidates->size = filtered_tokens.size();
+            min_p_applied = true;
         }
     }
 
-    // Resize the output vector to keep only the matching tokens
-    candidates->size = i;
+    // if the candidates are sorted or the unsorted implementation failed, use this implementation
+    if (!min_p_applied) {
+        // Sort the logits in descending order
+        if (!candidates->sorted) {
+            std::sort(candidates->data, candidates->data + candidates->size, [](const llama_token_data & a, const llama_token_data & b) {
+                return a.logit > b.logit;
+            });
+            candidates->sorted = true;
+        }
+
+        const float min_logit = candidates->data[0].logit + logf(p); // min logit for p_i >= p * p_max
+        size_t i = 1; // first token always matches
+
+        for (; i < candidates->size; ++i) {
+            if (candidates->data[i].logit < min_logit && i >= min_keep) {
+                break; // prob too small
+            }
+        }
+
+        // Resize the output vector to keep only the matching tokens
+        candidates->size = i;
+    }
 
     if (ctx) {
         ctx->t_sample_us += ggml_time_us() - t_start_sample_us;
@@ -7779,6 +8658,73 @@ void llama_sample_typical(struct llama_context * ctx, llama_token_data_array * c
     }
 }
 
+void llama_sample_entropy(struct llama_context * ctx, llama_token_data_array * candidates_p, float min_temp, float max_temp, float exponent_val) {
+    const int64_t t_start_sample_us = ggml_time_us();
+
+    // no need to do anything if there is only one (or zero) candidates
+    if(candidates_p->size <= 1) {
+        return;
+    }
+
+    // Calculate maximum possible entropy
+    float max_entropy = -logf(1.0f / candidates_p->size);
+
+    llama_sample_softmax(nullptr, candidates_p);
+
+    // Calculate entropy of the softmax probabilities
+    float entropy = 0.0f;
+    for (size_t i = 0; i < candidates_p->size; ++i) {
+        float prob = candidates_p->data[i].p;
+        if (prob > 0.0f) { // Ensure no log(0)
+            entropy -= prob * logf(prob);
+        }
+    }
+
+    // Normalize the entropy (max_entropy cannot be 0 here because we checked candidates_p->size != 1 above)
+    float normalized_entropy = entropy / max_entropy;
+
+    // Map the normalized entropy to the desired temperature range using the power function
+    float dyn_temp = min_temp + (max_temp - min_temp) * powf(normalized_entropy, exponent_val);
+
+#ifdef DEBUG
+    LLAMA_LOG_INFO("Your text maxtemp value is: %f\n", max_temp);
+    LLAMA_LOG_INFO("Entropy: %f\n", entropy);
+    LLAMA_LOG_INFO("Max Possible Entropy: %f\n", max_entropy);
+    LLAMA_LOG_INFO("Normalized Entropy: %f\n", normalized_entropy);
+    LLAMA_LOG_INFO("Exponent: %f\n", exponent_val);
+    LLAMA_LOG_INFO("Dynamic Temperature (dyn_temp): %f\n", dyn_temp);
+#endif
+
+    // Apply the dynamically calculated temperature scaling
+    for (size_t i = 0; i < candidates_p->size; ++i) {
+        candidates_p->data[i].logit /= dyn_temp;
+    }
+
+    // Re-compute softmax probabilities after scaling logits with dynamic temperature
+    double max_l_double = candidates_p->data[0].logit;
+    double cum_sum_double = 0.0;
+    for (size_t i = 0; i < candidates_p->size; ++i) {
+        double p = exp(candidates_p->data[i].logit - max_l_double);
+        candidates_p->data[i].p = p; // Store the scaled probability
+        cum_sum_double += p;
+    }
+    for (size_t i = 0; i < candidates_p->size; ++i) {
+        candidates_p->data[i].p /= cum_sum_double; // Re-normalize the probabilities
+    }
+
+#ifdef DEBUG
+    // Print the updated top 25 probabilities after temperature scaling
+    LLAMA_LOG_INFO("\nUpdated Top 25 Probabilities After Dynamic Temperature Scaling (in percentages):\n");
+    for (size_t i = 0; i < 25 && i < candidates_p->size; ++i) {
+        LLAMA_LOG_INFO("Token %zu: %f%%\n", i + 1, candidates_p->data[i].p * 100.0f);
+    }
+#endif
+
+    if (ctx) {
+        ctx->t_sample_us += ggml_time_us() - t_start_sample_us;
+    }
+}
+
 void llama_sample_temp(struct llama_context * ctx, llama_token_data_array * candidates_p, float temp) {
     const int64_t t_start_sample_us = ggml_time_us();
 
@@ -8367,9 +9313,13 @@ struct quantize_state_internal {
     const llama_model_quantize_params * params;
 
     int n_attention_wv    = 0;
-    int n_feed_forward_w2 = 0;
+    int n_ffn_down        = 0;
+    int n_ffn_gate        = 0;
+    int n_ffn_up          = 0;
     int i_attention_wv    = 0;
-    int i_feed_forward_w2 = 0;
+    int i_ffn_down        = 0;
+    int i_ffn_gate        = 0;
+    int i_ffn_up          = 0;
 
     int n_k_quantized     = 0;
     int n_fallback        = 0;
@@ -8453,6 +9403,23 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty
     auto use_more_bits = [](int i_layer, int num_layers) -> bool {
         return i_layer < num_layers/8 || i_layer >= 7*num_layers/8 || (i_layer - num_layers/8)%3 == 2;
     };
+    const int n_expert = std::max(1, (int)qs.model.hparams.n_expert);
+    auto layer_info = [n_expert] (int i_layer, int n_layer, const char * name) {
+        if (n_expert > 1) {
+            // Believe it or not, "experts" in the FFN of Mixtral-8x7B are not consecutive, but iccasionally randomly
+            // sprinkled in the model. Hence, simply dividing i_ffn_down by n_expert does not work
+            // for getting the current layer as I initially thought, and we need to resort to parsing the
+            // tensor name.
+            n_layer /= n_expert;
+            if (sscanf(name, "blk.%d.", &i_layer) != 1) {
+                throw std::runtime_error(format("Failed to determine layer for tensor %s", name));
+            }
+            if (i_layer < 0 || i_layer >= n_layer) {
+                throw std::runtime_error(format("Bad layer %d for tensor %s. Must be in [0, %d)", i_layer, name, n_layer));
+            }
+        }
+        return std::make_pair(i_layer, n_layer);
+    };
 
     if (name == tn(LLM_TENSOR_OUTPUT, "weight")) {
         int nx = tensor->ne[0];
@@ -8465,6 +9432,13 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty
         else if (new_type != GGML_TYPE_Q8_0) {
             new_type = GGML_TYPE_Q6_K;
         }
+    } else if (name == "token_embd.weight") {
+        if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS) {
+            new_type = GGML_TYPE_Q2_K;
+        }
+        else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS) {
+            new_type = GGML_TYPE_Q4_K;
+        }
     } else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS) {
         if (name.find("attn_v.weight") != std::string::npos) {
             if (qs.model.hparams.n_gqa() >= 4 || qs.model.hparams.n_expert >= 4) new_type = GGML_TYPE_Q4_K;
@@ -8472,12 +9446,19 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty
             ++qs.i_attention_wv;
         }
         else if (name.find("ffn_down") != std::string::npos) {
-            if (qs.i_feed_forward_w2 < qs.n_feed_forward_w2/8) new_type = GGML_TYPE_Q2_K;
-            ++qs.i_feed_forward_w2;
+            if (qs.i_ffn_down < qs.n_ffn_down/8) new_type = GGML_TYPE_Q2_K;
+            ++qs.i_ffn_down;
         }
-        else if (name == "token_embd.weight") new_type = GGML_TYPE_Q2_K;
     } else if (name.find("attn_v.weight") != std::string::npos) {
-        if      (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q3_K;
+        if      (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) {
+            new_type = qs.model.hparams.n_gqa() >= 4 ? GGML_TYPE_Q4_K : GGML_TYPE_Q3_K;
+        }
+        else if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K_S && qs.model.hparams.n_gqa() >= 4) {
+            new_type = GGML_TYPE_Q4_K;
+        }
+        else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS && qs.model.hparams.n_gqa() >= 4) {
+            new_type = GGML_TYPE_Q4_K;
+        }
         else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M) {
             new_type = qs.i_attention_wv < 2 ? GGML_TYPE_Q5_K : GGML_TYPE_Q4_K;
         }
@@ -8505,29 +9486,19 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty
             // TODO: explore better strategies
             new_type = GGML_TYPE_Q8_0;
         }
-    } else if (name.find("ffn_down") != std::string::npos) {
-        const int n_expert = std::max(1, (int)qs.model.hparams.n_expert);
-        int i_layer, n_layer;
-        if (n_expert == 1) {
-            i_layer = qs.i_feed_forward_w2;
-            n_layer = qs.n_feed_forward_w2;
-        } else {
-            // Believe it or not, "experts" in the FFN of Mixtral-8x7B are not consecutive, but iccasionally randomly
-            // sprinkled in the model. Hence, simply dividing i_feed_forward_w2 by n_expert does not work
-            // for getting the current layer as I initially thought, and we need to resort to parsing the
-            // tensor name.
-            n_layer = qs.n_feed_forward_w2 / n_expert;
-            if (sscanf(name.c_str(), "blk.%d.ffn_down", &i_layer) != 1) {
-                throw std::runtime_error(format("Failed to determine layer for tensor %s", name.c_str()));
-            }
-            if (i_layer < 0 || i_layer >= n_layer) {
-                throw std::runtime_error(format("Bad layer %d for tensor %s. Must be in [0, %d)", i_layer, name.c_str(), n_layer));
-            }
+        else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_XS) {
+            new_type = GGML_TYPE_Q2_K;
         }
+    } else if (name.find("ffn_down") != std::string::npos) {
+        auto info = layer_info(qs.i_ffn_down, qs.n_ffn_down, name.c_str());
+        int i_layer = info.first, n_layer = info.second;
         if      (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q3_K;
-        else if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K_S) {
+        else if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K_S || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_XS) {
             if (i_layer < n_layer/8) new_type = GGML_TYPE_Q4_K;
         }
+        //else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS) {
+        //    if (i_layer < n_layer/8) new_type = GGML_TYPE_Q5_K;
+        //}
         else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M) {
             new_type = i_layer < n_layer/16 ? GGML_TYPE_Q5_K
                      : arch != LLM_ARCH_FALCON || use_more_bits(i_layer, n_layer) ? GGML_TYPE_Q4_K
@@ -8555,16 +9526,18 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty
             // same quantization as before imatrix stuff, and b) Q4_1/Q5_1 do go crazy on ffn_down without an imatrix.
             new_type = ftype == LLAMA_FTYPE_MOSTLY_Q4_0 ? GGML_TYPE_Q4_1 : GGML_TYPE_Q5_1;
         }
-        ++qs.i_feed_forward_w2;
+        ++qs.i_ffn_down;
     } else if (name.find("attn_output.weight") != std::string::npos) {
         if (arch != LLM_ARCH_FALCON) {
             if (qs.model.hparams.n_expert == 8) {
-                if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K   || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_S || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M ||
+                if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K   || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS ||
+                    ftype == LLAMA_FTYPE_MOSTLY_Q3_K_S || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M ||
                     ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S || ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M) {
                     new_type = GGML_TYPE_Q5_K;
                 }
             } else {
                 if      (ftype == LLAMA_FTYPE_MOSTLY_Q2_K  ) new_type = GGML_TYPE_Q3_K;
+                else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS) new_type = GGML_TYPE_Q3_K;
                 else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M) new_type = GGML_TYPE_Q4_K;
                 else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) new_type = GGML_TYPE_Q5_K;
             }
@@ -8577,6 +9550,24 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty
         else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M) new_type = GGML_TYPE_Q5_K;
         else if (ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M) new_type = GGML_TYPE_Q6_K;
     }
+    else if (name.find("ffn_gate") != std::string::npos) {
+        auto info = layer_info(qs.i_ffn_gate, qs.n_ffn_gate, name.c_str());
+        int i_layer = info.first, n_layer = info.second;
+        if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_XS && !use_more_bits(i_layer, n_layer)) {
+            new_type = GGML_TYPE_Q2_K;
+        }
+        ++qs.i_ffn_gate;
+    }
+    else if (name.find("ffn_up") != std::string::npos) {
+        auto info = layer_info(qs.i_ffn_up, qs.n_ffn_up, name.c_str());
+        int i_layer = info.first, n_layer = info.second;
+        if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_XS && !use_more_bits(i_layer, n_layer)) {
+            new_type = GGML_TYPE_Q2_K;
+        }
+        ++qs.i_ffn_up;
+    }
+    //    if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q3_K;
+    //}
     // IK: let's remove this, else Q2_K is almost the same as Q3_K_S
     //else if (name.find("ffn_gate") != std::string::npos || name.find("ffn_up") != std::string::npos) {
     //    if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q3_K;
@@ -8589,7 +9580,8 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty
     bool convert_incompatible_tensor = false;
     if (new_type == GGML_TYPE_Q2_K || new_type == GGML_TYPE_Q3_K || new_type == GGML_TYPE_Q4_K ||
         new_type == GGML_TYPE_Q5_K || new_type == GGML_TYPE_Q6_K ||
-        new_type == GGML_TYPE_IQ2_XS || new_type == GGML_TYPE_IQ2_XXS) {
+        new_type == GGML_TYPE_IQ2_XS || new_type == GGML_TYPE_IQ2_XXS ||
+        new_type == GGML_TYPE_IQ3_XXS) {
         int nx = tensor->ne[0];
         int ny = tensor->ne[1];
         if (nx % QK_K != 0) {
@@ -8603,6 +9595,7 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty
         switch (new_type) {
             case GGML_TYPE_IQ2_XXS:
             case GGML_TYPE_IQ2_XS:
+            case GGML_TYPE_IQ3_XXS:
             case GGML_TYPE_Q2_K: new_type = GGML_TYPE_Q4_0; break;
             case GGML_TYPE_Q3_K: new_type = GGML_TYPE_Q4_1; break;
             case GGML_TYPE_Q4_K: new_type = GGML_TYPE_Q5_0; break;
@@ -8631,8 +9624,9 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
         case LLAMA_FTYPE_ALL_F32:     quantized_type = GGML_TYPE_F32;  break;
 
         // K-quants
+        case LLAMA_FTYPE_MOSTLY_Q2_K_S:
         case LLAMA_FTYPE_MOSTLY_Q2_K:   quantized_type = GGML_TYPE_Q2_K; break;
-        case LLAMA_FTYPE_MOSTLY_Q2_K_S: quantized_type = GGML_TYPE_Q2_K; break;
+        case LLAMA_FTYPE_MOSTLY_Q3_K_XS:
         case LLAMA_FTYPE_MOSTLY_Q3_K_S:
         case LLAMA_FTYPE_MOSTLY_Q3_K_M:
         case LLAMA_FTYPE_MOSTLY_Q3_K_L: quantized_type = GGML_TYPE_Q3_K; break;
@@ -8643,6 +9637,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
         case LLAMA_FTYPE_MOSTLY_Q6_K:   quantized_type = GGML_TYPE_Q6_K; break;
         case LLAMA_FTYPE_MOSTLY_IQ2_XXS:quantized_type = GGML_TYPE_IQ2_XXS; break;
         case LLAMA_FTYPE_MOSTLY_IQ2_XS :quantized_type = GGML_TYPE_IQ2_XS;  break;
+        case LLAMA_FTYPE_MOSTLY_IQ3_XXS:quantized_type = GGML_TYPE_IQ3_XXS; break;
 
         default: throw std::runtime_error(format("invalid output file type %d\n", ftype));
     }
@@ -8700,12 +9695,18 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
             ++qs.n_attention_wv;
         }
         else if (name.find("ffn_down") != std::string::npos) {
-            ++qs.n_feed_forward_w2;
+            ++qs.n_ffn_down;
+        }
+        else if (name.find("ffn_gate") != std::string::npos) {
+            ++qs.n_ffn_gate;
+        }
+        else if (name.find("ffn_up") != std::string::npos) {
+            ++qs.n_ffn_up;
         }
     }
-    if (qs.n_attention_wv != qs.n_feed_forward_w2 || (uint32_t)qs.n_attention_wv != model.hparams.n_layer) {
-        LLAMA_LOG_WARN("%s ============ Strange model: n_attention_wv = %d, n_feed_forward_w2 = %d, hparams.n_layer = %d\n",
-                __func__, qs.n_attention_wv, qs.n_feed_forward_w2, model.hparams.n_layer);
+    if (qs.n_attention_wv != qs.n_ffn_down || (uint32_t)qs.n_attention_wv != model.hparams.n_layer) {
+        LLAMA_LOG_WARN("%s ============ Strange model: n_attention_wv = %d, n_ffn_down = %d, hparams.n_layer = %d\n",
+                __func__, qs.n_attention_wv, qs.n_ffn_down, model.hparams.n_layer);
     }
 
     size_t total_size_org = 0;
@@ -8738,8 +9739,6 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
     // placeholder for the meta data
     ::zeros(fout, meta_size);
 
-    std::set<ggml_type> used_iq2;
-
     for (int i = 0; i < ml.n_tensors; ++i) {
         struct ggml_tensor * tensor = ml.get_tensor_meta(i);
 
@@ -8792,11 +9791,6 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
         } else {
             const size_t nelements = ggml_nelements(tensor);
 
-            if ((new_type == GGML_TYPE_IQ2_XXS || new_type == GGML_TYPE_IQ2_XS) && used_iq2.find(new_type) == used_iq2.end()) {
-                ggml_init_iq2_quantization(new_type);
-                used_iq2.insert(new_type);
-            }
-
             const float * imatrix = nullptr;
             if (imatrix_data) {
                 auto it = imatrix_data->find(tensor->name);
@@ -8922,10 +9916,6 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
 
     fout.close();
 
-    for (auto type : used_iq2) {
-        ggml_deinit_iq2_quantization(type);
-    }
-
     gguf_free(ctx_out);
 
     LLAMA_LOG_INFO("%s: model size  = %8.2f MB\n", __func__, total_size_org/1024.0/1024.0);
@@ -9271,6 +10261,8 @@ struct llama_context_params llama_context_default_params() {
         /*.yarn_beta_fast              =*/ 32.0f,
         /*.yarn_beta_slow              =*/ 1.0f,
         /*.yarn_orig_ctx               =*/ 0,
+        /*.cb_eval                     =*/ nullptr,
+        /*.cb_eval_user_data           =*/ nullptr,
         /*.type_k                      =*/ GGML_TYPE_F16,
         /*.type_v                      =*/ GGML_TYPE_F16,
         /*.mul_mat_q                   =*/ true,
@@ -9296,18 +10288,45 @@ struct llama_model_quantize_params llama_model_quantize_default_params() {
     return result;
 }
 
-int32_t llama_max_devices(void) {
-    return LLAMA_MAX_DEVICES;
+size_t llama_max_devices(void) {
+#if defined(GGML_USE_METAL)
+    return 1;
+#elif defined(GGML_USE_CUBLAS)
+    return GGML_CUDA_MAX_DEVICES;
+#elif defined(GGML_USE_SYCL)
+    return GGML_SYCL_MAX_DEVICES;
+#else
+    return 1;
+#endif
 }
 
-bool llama_mmap_supported(void) {
+bool llama_supports_mmap(void) {
     return llama_mmap::SUPPORTED;
 }
 
-bool llama_mlock_supported(void) {
+bool llama_supports_mlock(void) {
     return llama_mlock::SUPPORTED;
 }
 
+bool llama_supports_gpu_offload(void) {
+#if defined(GGML_USE_CUBLAS) || defined(GGML_USE_CLBLAST) || defined(GGML_USE_METAL) || defined(GGML_USE_VULKAN) || \
+    defined(GGML_USE_SYCL)   || defined(GGML_USE_KOMPUTE)
+    // Defined when llama.cpp is compiled with support for offloading model layers to GPU.
+    return true;
+#else
+    return false;
+#endif
+}
+
+// deprecated:
+bool llama_mmap_supported(void) {
+    return llama_supports_mmap();
+}
+
+bool llama_mlock_supported(void) {
+    return llama_supports_mlock();
+}
+
 void llama_backend_init(bool numa) {
     ggml_time_init();
 
@@ -9331,6 +10350,7 @@ void llama_backend_free(void) {
 #ifdef GGML_USE_MPI
     ggml_mpi_backend_free();
 #endif
+    ggml_quantize_free();
 }
 
 int64_t llama_time_us(void) {
@@ -9338,8 +10358,8 @@ int64_t llama_time_us(void) {
 }
 
 struct llama_model * llama_load_model_from_file(
-                             const char * path_model,
-              struct llama_model_params   params) {
+        const char * path_model,
+        struct llama_model_params   params) {
     ggml_time_init();
 
     llama_model * model = new llama_model;
@@ -9411,6 +10431,9 @@ struct llama_context * llama_new_context_with_model(
                                hparams.n_yarn_orig_ctx != 0 ? hparams.n_yarn_orig_ctx :
                                                               hparams.n_ctx_train;
 
+    cparams.cb_eval           = params.cb_eval;
+    cparams.cb_eval_user_data = params.cb_eval_user_data;
+
     auto rope_scaling_type = params.rope_scaling_type;
     if (rope_scaling_type == LLAMA_ROPE_SCALING_UNSPECIFIED) {
         rope_scaling_type = hparams.rope_scaling_type_train;
@@ -9477,6 +10500,36 @@ struct llama_context * llama_new_context_with_model(
                 }
             }
         }
+#elif defined(GGML_USE_VULKAN)
+        if (model->n_gpu_layers > 0) {
+            ggml_backend_t backend = ggml_backend_vk_init();
+            if (backend == nullptr) {
+                LLAMA_LOG_ERROR("%s: failed to initialize Vulkan backend\n", __func__);
+                llama_free(ctx);
+                return nullptr;
+            }
+            ctx->backends.push_back(backend);
+        }
+#elif defined(GGML_USE_SYCL)
+        if (model->n_gpu_layers > 0) {
+            ggml_backend_t backend = ggml_backend_sycl_init(model->main_gpu);
+            if (backend == nullptr) {
+                LLAMA_LOG_ERROR("%s: failed to initialize SYCL%d backend\n", __func__, model->main_gpu);
+                llama_free(ctx);
+                return nullptr;
+            }
+            ctx->backends.push_back(backend);
+        }
+#elif defined(GGML_USE_KOMPUTE)
+        if (model->n_gpu_layers > 0) {
+            auto * backend = ggml_backend_kompute_init(model->main_gpu);
+            if (backend == nullptr) {
+                LLAMA_LOG_ERROR("%s: failed to initialize Kompute backend\n", __func__);
+                llama_free(ctx);
+                return nullptr;
+            }
+            ctx->backends.push_back(backend);
+        }
 #endif
         ctx->backend_cpu = ggml_backend_cpu_init();
         if (ctx->backend_cpu == nullptr) {
@@ -9518,6 +10571,35 @@ struct llama_context * llama_new_context_with_model(
             ctx->embedding.resize(hparams.n_embd);
         }
 
+        // graph inputs
+        {
+            ggml_init_params init_params = {
+                /* .mem_size   */ ggml_tensor_overhead()*5,
+                /* .mem_buffer */ nullptr,
+                /* .no_alloc   */ true,
+            };
+            ctx->ctx_input = ggml_init(init_params);
+
+            ctx->inp_tokens  = ggml_new_tensor_1d(ctx->ctx_input, GGML_TYPE_I32, cparams.n_batch);
+            ctx->inp_embd    = ggml_new_tensor_2d(ctx->ctx_input, GGML_TYPE_F32, hparams.n_embd, cparams.n_batch);
+            ctx->inp_pos     = ggml_new_tensor_1d(ctx->ctx_input, GGML_TYPE_I32, cparams.n_batch);
+            ctx->inp_KQ_mask = ggml_new_tensor_2d(ctx->ctx_input, GGML_TYPE_F32, cparams.n_ctx, cparams.n_batch);
+            ctx->inp_K_shift = ggml_new_tensor_1d(ctx->ctx_input, GGML_TYPE_I32, cparams.n_ctx);
+
+            ggml_set_name(ctx->inp_tokens,  "inp_tokens");
+            ggml_set_name(ctx->inp_embd,    "inp_embd");
+            ggml_set_name(ctx->inp_pos,     "inp_pos");
+            ggml_set_name(ctx->inp_KQ_mask, "inp_KQ_mask");
+            ggml_set_name(ctx->inp_K_shift, "inp_K_shift");
+
+            ctx->buf_input = ggml_backend_alloc_ctx_tensors_from_buft(ctx->ctx_input, llama_default_buffer_type_cpu(true));
+
+            LLAMA_LOG_INFO("%s: %10s input buffer size   = %8.2f MiB\n", __func__,
+                    ggml_backend_buffer_name(ctx->buf_input),
+                    ggml_backend_buffer_get_size(ctx->buf_input) / 1024.0 / 1024.0);
+        }
+
+        // scheduler and compute buffers
         {
             // buffer types used for the compute buffer of each backend
             std::vector<ggml_backend_buffer_type_t> backend_buft;
@@ -9544,9 +10626,6 @@ struct llama_context * llama_new_context_with_model(
 
             // initialize scheduler with the worst-case graph
             ggml_backend_sched_init_measure(ctx->sched, gf);
-            // note: the number of splits during measure is higher than during inference due to the kv shift
-            int n_splits = ggml_backend_sched_get_n_splits(ctx->sched);
-            LLAMA_LOG_INFO("%s: graph splits (measure): %d\n", __func__, n_splits);
             ctx->alloc = ggml_backend_sched_get_tallocr(ctx->sched, ctx->backend_cpu);
 
             for (ggml_backend_t backend : ctx->backends) {
@@ -9555,6 +10634,10 @@ struct llama_context * llama_new_context_with_model(
                         ggml_backend_buffer_name(buf),
                         ggml_backend_buffer_get_size(buf) / 1024.0 / 1024.0);
             }
+
+            // note: the number of splits during measure is higher than during inference due to the kv shift
+            int n_splits = ggml_backend_sched_get_n_splits(ctx->sched);
+            LLAMA_LOG_INFO("%s: graph splits (measure): %d\n", __func__, n_splits);
         }
     }
 
@@ -10294,22 +11377,24 @@ struct llama_batch llama_batch_get_one(
     };
 }
 
-struct llama_batch llama_batch_init(int32_t n_tokens, int32_t embd, int32_t n_seq_max) {
+struct llama_batch llama_batch_init(int32_t n_tokens_alloc, int32_t embd, int32_t n_seq_max) {
     llama_batch batch = { 0, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, 0, 0, 0, };
 
     if (embd) {
-        batch.embd = (float *) malloc(sizeof(float) * n_tokens * embd);
+        batch.embd = (float *) malloc(sizeof(float) * n_tokens_alloc * embd);
     } else {
-        batch.token = (llama_token *) malloc(sizeof(llama_token) * n_tokens);
+        batch.token = (llama_token *) malloc(sizeof(llama_token) * n_tokens_alloc);
     }
 
-    batch.pos      = (llama_pos *)     malloc(sizeof(llama_pos)      * n_tokens);
-    batch.n_seq_id = (int32_t *)       malloc(sizeof(int32_t)        * n_tokens);
-    batch.seq_id   = (llama_seq_id **) malloc(sizeof(llama_seq_id *) * n_tokens);
-    for (int i = 0; i < n_tokens; ++i) {
+    batch.pos      = (llama_pos *)     malloc(sizeof(llama_pos)      * n_tokens_alloc);
+    batch.n_seq_id = (int32_t *)       malloc(sizeof(int32_t)        * n_tokens_alloc);
+    batch.seq_id   = (llama_seq_id **) malloc(sizeof(llama_seq_id *) * (n_tokens_alloc + 1));
+    for (int i = 0; i < n_tokens_alloc; ++i) {
         batch.seq_id[i] = (llama_seq_id *) malloc(sizeof(llama_seq_id) * n_seq_max);
     }
-    batch.logits   = (int8_t *)        malloc(sizeof(int8_t)         * n_tokens);
+    batch.seq_id[n_tokens_alloc] = nullptr;
+
+    batch.logits   = (int8_t *)        malloc(sizeof(int8_t)         * n_tokens_alloc);
 
     return batch;
 }
@@ -10320,7 +11405,7 @@ void llama_batch_free(struct llama_batch batch) {
     if (batch.pos)      free(batch.pos);
     if (batch.n_seq_id) free(batch.n_seq_id);
     if (batch.seq_id) {
-        for (int i = 0; i < batch.n_tokens; ++i) {
+        for (int i = 0; batch.seq_id[i] != nullptr; ++i) {
             free(batch.seq_id[i]);
         }
         free(batch.seq_id);