cui-llama.rn 1.0.9 → 1.0.11

package/cpp/common.cpp

@@ -116,8 +116,34 @@ int32_t cpu_get_num_physical_cores() {
     if (result == 0) {
         return num_physical_cores;
     }
-#elif defined(_WIN32)
-    //TODO: Implement
+#elif defined(_WIN32) && (_WIN32_WINNT >= 0x0601) && !defined(__MINGW64__) // windows 7 and later
+    // TODO: windows + arm64 + mingw64
+    unsigned int n_threads_win = std::thread::hardware_concurrency();
+    unsigned int default_threads = n_threads_win > 0 ? (n_threads_win <= 4 ? n_threads_win : n_threads_win / 2) : 4;
+
+    DWORD buffer_size = 0;
+    if (!GetLogicalProcessorInformationEx(RelationProcessorCore, nullptr, &buffer_size)) {
+        if (GetLastError() != ERROR_INSUFFICIENT_BUFFER) {
+            return default_threads;
+        }
+    }
+
+    std::vector<char> buffer(buffer_size);
+    if (!GetLogicalProcessorInformationEx(RelationProcessorCore, reinterpret_cast<PSYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX>(buffer.data()), &buffer_size)) {
+        return default_threads;
+    }
+
+    int32_t num_physical_cores = 0;
+    PSYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX info = reinterpret_cast<PSYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX>(buffer.data());
+    while (buffer_size > 0) {
+        if (info->Relationship == RelationProcessorCore) {
+            num_physical_cores += info->Processor.GroupCount;
+        }
+        buffer_size -= info->Size;
+        info = reinterpret_cast<PSYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX>(reinterpret_cast<char*>(info) + info->Size);
+    }
+
+    return num_physical_cores > 0 ? num_physical_cores : default_threads;
 #endif
     unsigned int n_threads = std::thread::hardware_concurrency();
     return n_threads > 0 ? (n_threads <= 4 ? n_threads : n_threads / 2) : 4;
@@ -1733,7 +1759,13 @@ std::string gpt_params_get_system_info(const gpt_params & params) {
     if (params.n_threads_batch != -1) {
         os << " (n_threads_batch = " << params.n_threads_batch << ")";
     }
+#if defined(_WIN32) && (_WIN32_WINNT >= 0x0601) && !defined(__MINGW64__) // windows 7 and later
+    // TODO: windows + arm64 + mingw64
+    DWORD logicalProcessorCount = GetActiveProcessorCount(ALL_PROCESSOR_GROUPS);
+    os << " / " << logicalProcessorCount << " | " << llama_print_system_info();
+#else
     os << " / " << std::thread::hardware_concurrency() << " | " << llama_print_system_info();
+#endif
 
     return os.str();
 }
@@ -2709,12 +2741,6 @@ std::string llama_detokenize(llama_context * ctx, const std::vector<llama_token>
     return text;
 }
 
-bool llama_should_add_bos_token(const llama_model * model) {
-    const int add_bos = llama_add_bos_token(model);
-
-    return add_bos != -1 ? bool(add_bos) : (llama_vocab_type(model) == LLAMA_VOCAB_TYPE_SPM);
-}
-
 //
 // Chat template utils
 //

package/cpp/common.h

@@ -392,10 +392,6 @@ std::string llama_detokenize(
         const std::vector<llama_token> & tokens,
                                   bool   special = true);
 
-// Uses the value from the model metadata if possible, otherwise
-// defaults to true when model type is SPM, otherwise false.
-bool llama_should_add_bos_token(const llama_model * model);
-
 //
 // Chat template utils
 //

package/cpp/ggml-backend.c

@@ -1018,10 +1018,6 @@ static bool lm_ggml_is_view_op(enum lm_ggml_op op) {
 #define LM_GGML_SCHED_MAX_BACKENDS 16
 #endif
 
-#ifndef LM_GGML_SCHED_MAX_SPLITS
-#define LM_GGML_SCHED_MAX_SPLITS 2048
-#endif
-
 #ifndef LM_GGML_SCHED_MAX_SPLIT_INPUTS
 #define LM_GGML_SCHED_MAX_SPLIT_INPUTS LM_GGML_MAX_SRC
 #endif
@@ -1125,7 +1121,8 @@ static int lm_ggml_backend_sched_backend_from_buffer(lm_ggml_backend_sched_t sch
 }
 
 #if 0
-static char causes[LM_GGML_DEFAULT_GRAPH_SIZE*16 + LM_GGML_SCHED_MAX_SPLITS*LM_GGML_SCHED_MAX_SPLIT_INPUTS][128]; // debug only
+#define LM_GGML_SCHED_MAX_SPLITS_DEBUG 4096
+static char causes[LM_GGML_DEFAULT_GRAPH_SIZE*16 + LM_GGML_SCHED_MAX_SPLITS_DEBUG*LM_GGML_SCHED_MAX_SPLIT_INPUTS][128]; // debug only
 #define SET_CAUSE(node, ...) sprintf(causes[hash_id(node)], __VA_ARGS__)
 #define GET_CAUSE(node) causes[hash_id(node)]
 #else
@@ -1549,7 +1546,6 @@ static void lm_ggml_backend_sched_split_graph(lm_ggml_backend_sched_t sched, str
                     sched->splits = realloc(sched->splits, sched->splits_capacity * sizeof(struct lm_ggml_backend_sched_split));
                     LM_GGML_ASSERT(sched->splits != NULL);
                 }
-                LM_GGML_ASSERT(i_split < LM_GGML_SCHED_MAX_SPLITS);
                 split = &sched->splits[i_split];
                 split->backend_id = node_backend_id;
                 split->i_start = i;
@@ -1865,13 +1861,14 @@ lm_ggml_backend_sched_t lm_ggml_backend_sched_new(
     sched->hv_tensor_backend_ids = malloc(sched->hash_set.size * sizeof(sched->hv_tensor_backend_ids[0]));
     sched->hv_tensor_copies      = malloc(sched->hash_set.size * sched->n_backends * sched->n_copies * sizeof(struct lm_ggml_tensor *));
 
-    const size_t nodes_size = graph_size + LM_GGML_SCHED_MAX_SPLITS*LM_GGML_SCHED_MAX_SPLIT_INPUTS*2;
+    const size_t lm_ggml_sched_max_splits = graph_size; // at most there is one split for each node in the graph
+    const size_t nodes_size = graph_size + lm_ggml_sched_max_splits*LM_GGML_SCHED_MAX_SPLIT_INPUTS*2;
     sched->node_backend_ids = calloc(nodes_size, sizeof(sched->node_backend_ids[0]));
     sched->leaf_backend_ids = calloc(nodes_size, sizeof(sched->leaf_backend_ids[0]));
     sched->prev_node_backend_ids = calloc(nodes_size, sizeof(sched->prev_node_backend_ids[0]));
     sched->prev_leaf_backend_ids = calloc(nodes_size, sizeof(sched->prev_leaf_backend_ids[0]));
 
-    sched->context_buffer_size = LM_GGML_SCHED_MAX_SPLITS*LM_GGML_SCHED_MAX_SPLIT_INPUTS*2*sizeof(struct lm_ggml_tensor) + lm_ggml_graph_overhead_custom(graph_size, false);
+    sched->context_buffer_size = lm_ggml_sched_max_splits*LM_GGML_SCHED_MAX_SPLIT_INPUTS*2*sizeof(struct lm_ggml_tensor) + lm_ggml_graph_overhead_custom(graph_size, false);
     sched->context_buffer = malloc(sched->context_buffer_size);
 
     const int initial_splits_capacity = 16;

package/cpp/ggml-metal.m

@@ -310,7 +310,7 @@ static struct lm_ggml_backend_metal_context * lm_ggml_metal_init(int n_cb) {
     LM_GGML_METAL_LOG_INFO("%s: picking default device: %s\n", __func__, [[device name] UTF8String]);
 
     // Configure context
-    struct lm_ggml_backend_metal_context * ctx = malloc(sizeof(struct lm_ggml_backend_metal_context));
+    struct lm_ggml_backend_metal_context * ctx = calloc(1, sizeof(struct lm_ggml_backend_metal_context));
     ctx->device = device;
     ctx->n_cb   = MIN(n_cb, LM_GGML_METAL_MAX_BUFFERS);
     ctx->queue  = [ctx->device newCommandQueue];
@@ -2313,7 +2313,7 @@ static enum lm_ggml_status lm_ggml_metal_graph_compute(
                         memcpy(&beta_fast,   (int32_t *) dst->op_params +  9, sizeof(float));
                         memcpy(&beta_slow,   (int32_t *) dst->op_params + 10, sizeof(float));
 
-                        const bool is_neox = mode & 2;
+                        const bool is_neox = mode & LM_GGML_ROPE_TYPE_NEOX;
 
                         id<MTLComputePipelineState> pipeline = nil;
 

package/cpp/ggml.c

@@ -14094,7 +14094,7 @@ static void lm_ggml_compute_forward_rope_f32(
     float corr_dims[2];
     lm_ggml_rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow, corr_dims);
 
-    const bool is_neox = mode & 2;
+    const bool is_neox = mode & LM_GGML_ROPE_TYPE_NEOX;
 
     const float * freq_factors = NULL;
     if (src2 != NULL) {
@@ -14219,7 +14219,7 @@ static void lm_ggml_compute_forward_rope_f16(
     float corr_dims[2];
     lm_ggml_rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow, corr_dims);
 
-    const bool is_neox = mode & 2;
+    const bool is_neox = mode & LM_GGML_ROPE_TYPE_NEOX;
 
     const float * freq_factors = NULL;
     if (src2 != NULL) {
@@ -21129,7 +21129,7 @@ struct lm_gguf_context * lm_gguf_init_from_file(const char * fname, struct lm_gg
                 (int64_t) info->ne[2] *
                 (int64_t) info->ne[3];
 
-            if (ne % lm_ggml_blck_size(info->type) != 0) {
+            if (lm_ggml_blck_size(info->type) == 0 || ne % lm_ggml_blck_size(info->type) != 0) {
                 fprintf(stderr, "%s: tensor '%s' of type %d (%s) number of elements (%" PRId64 ") is not a multiple of block size (%" PRId64 ")\n",
                         __func__, info->name.data, (int) info->type, lm_ggml_type_name(info->type), ne, lm_ggml_blck_size(info->type));
                 fclose(file);

package/cpp/ggml.h

@@ -244,6 +244,8 @@
 #define LM_GGML_EXIT_SUCCESS 0
 #define LM_GGML_EXIT_ABORTED 1
 
+#define LM_GGML_ROPE_TYPE_NEOX 2
+
 #define LM_GGUF_MAGIC "GGUF"
 
 #define LM_GGUF_VERSION 3
@@ -1453,8 +1455,8 @@ extern "C" {
             struct lm_ggml_tensor  * b);
 
     // rotary position embedding
-    // if mode & 1 == 1, skip n_past elements (NOT SUPPORTED)
-    // if mode & 2 == 1, GPT-NeoX style
+    // if (mode & 1) - skip n_past elements (NOT SUPPORTED)
+    // if (mode & LM_GGML_ROPE_TYPE_NEOX) - GPT-NeoX style
     //
     // b is an int32 vector with size a->ne[2], it contains the positions
     LM_GGML_API struct lm_ggml_tensor * lm_ggml_rope(

package/cpp/grammar-parser.cpp

@@ -369,6 +369,9 @@ namespace grammar_parser {
             }
             // Validate the state to ensure that all rules are defined
             for (const auto & rule : state.rules) {
+                if (rule.empty()) {
+                    throw std::runtime_error("Undefined rule");
+                }
                 for (const auto & elem : rule) {
                     if (elem.type == LLAMA_GRETYPE_RULE_REF) {
                         // Ensure that the rule at that location exists

package/cpp/llama-sampling.cpp

@@ -85,14 +85,14 @@ void llama_sample_top_k_impl(struct llama_sampling * smpl, llama_token_data_arra
             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;
+            constexpr float bucket_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);
+                int ib = int(bucket_scale * val + bucket_inter); //nbuckets * (val - bucket_low) / (bucket_high - bucket_low);
                 ib = std::max(0, std::min(nbuckets-1, ib));
                 bucket_idx[i] = ib;
                 ++histo[ib];

package/cpp/llama-vocab.cpp

@@ -388,6 +388,7 @@ struct llm_tokenizer_bpe {
             case LLAMA_VOCAB_PRE_TYPE_COMMAND_R:
             case LLAMA_VOCAB_PRE_TYPE_SMOLLM:
             case LLAMA_VOCAB_PRE_TYPE_CODESHELL:
+            case LLAMA_VOCAB_PRE_TYPE_EXAONE:
                 regex_exprs = {
                     "\\p{N}",
                     "'s|'t|'re|'ve|'m|'ll|'d| ?\\p{L}+| ?\\p{N}+| ?[^\\s\\p{L}\\p{N}]+|\\s+(?!\\S)",
@@ -410,6 +411,8 @@ struct llm_tokenizer_bpe {
                 };
                 break;
             case LLAMA_VOCAB_PRE_TYPE_PORO:
+            case LLAMA_VOCAB_PRE_TYPE_BLOOM:
+            case LLAMA_VOCAB_PRE_TYPE_GPT3_FINNISH:
                 regex_exprs = {
                     " ?[^(\\s|.,!?…。，、।۔،)]+",
                 };
@@ -1466,11 +1469,11 @@ llama_token llama_token_pad_impl(const struct llama_vocab & vocab) {
     return vocab.special_pad_id;
 }
 
-int32_t llama_add_bos_token_impl(const struct llama_vocab & vocab) {
+bool llama_add_bos_token_impl(const struct llama_vocab & vocab) {
     return vocab.tokenizer_add_bos;
 }
 
-int32_t llama_add_eos_token_impl(const struct llama_vocab & vocab) {
+bool llama_add_eos_token_impl(const struct llama_vocab & vocab) {
     return vocab.tokenizer_add_eos;
 }
 

package/cpp/llama-vocab.h

@@ -95,8 +95,8 @@ llama_token llama_token_sep_impl(const struct llama_vocab & vocab);
 llama_token llama_token_nl_impl (const struct llama_vocab & vocab);
 llama_token llama_token_pad_impl(const struct llama_vocab & vocab);
 
-int32_t llama_add_bos_token_impl(const struct llama_vocab & vocab);
-int32_t llama_add_eos_token_impl(const struct llama_vocab & vocab);
+bool llama_add_bos_token_impl(const struct llama_vocab & vocab);
+bool llama_add_eos_token_impl(const struct llama_vocab & vocab);
 
 llama_token llama_token_prefix_impl(const struct llama_vocab & vocab);
 llama_token llama_token_middle_impl(const struct llama_vocab & vocab);

package/cpp/llama.cpp

@@ -221,6 +221,8 @@ enum llm_arch {
     LLM_ARCH_T5,
     LLM_ARCH_T5ENCODER,
     LLM_ARCH_JAIS,
+    LLM_ARCH_NEMOTRON,
+    LLM_ARCH_EXAONE,
     LLM_ARCH_UNKNOWN,
 };
 
@@ -266,6 +268,8 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_T5,              "t5"           },
     { LLM_ARCH_T5ENCODER,       "t5encoder"    },
     { LLM_ARCH_JAIS,            "jais"         },
+    { LLM_ARCH_NEMOTRON,        "nemotron"     },
+    { LLM_ARCH_EXAONE,          "exaone"       },
     { LLM_ARCH_UNKNOWN,         "(unknown)"    },
 };
 
@@ -1307,6 +1311,43 @@ static const std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NA
             { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
         },
     },
+    {
+        LLM_ARCH_NEMOTRON,
+        {
+            { 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_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+        },
+    },
+    {
+        LLM_ARCH_EXAONE,
+        {
+            { 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_UNKNOWN,
         {
@@ -3586,13 +3627,8 @@ namespace GGUFMeta {
 
 using llama_buf_map = std::unordered_map<uint32_t, lm_ggml_backend_buffer_t>;
 
-// TODO: update when needed or think of some clever automatic way to do this
-static size_t llama_model_max_nodes(const llama_model & /*model*/) {
-    //if (model.arch == LLM_ARCH_LLAMA && model.hparams.n_layer > ??) { // llama-3 405B
-    //    return 32768;
-    //}
-
-    return 8192;
+static size_t llama_model_max_nodes(const llama_model & model) {
+    return std::max<size_t>(8192, model.tensors_by_name.size()*5);
 }
 
 struct llama_model_loader {
@@ -4912,7 +4948,6 @@ static void llm_load_hparams(
             } break;
         case LLM_ARCH_PHI3:
             {
-                ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa);
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
 
                 switch (hparams.n_layer) {
@@ -4921,6 +4956,22 @@ static void llm_load_hparams(
                     case 40: model.type = e_model::MODEL_14B; break;
                     default: model.type = e_model::MODEL_UNKNOWN;
                 }
+
+                // for backward compatibility ; see: https://github.com/ggerganov/llama.cpp/pull/8931
+                if ((hparams.n_layer == 32 || hparams.n_layer == 40) && hparams.n_ctx_train == 4096) {
+                    // default value for Phi-3-mini-4k-instruct and Phi-3-medium-4k-instruct
+                    hparams.n_swa = 2047;
+                } else if (hparams.n_layer == 32 && hparams.n_head_kv(0) == 32 && hparams.n_ctx_train == 131072) {
+                    // default value for Phi-3-mini-128k-instruct
+                    hparams.n_swa = 262144;
+                } else if (hparams.n_layer == 40 && hparams.n_ctx_train == 131072) {
+                    // default value for Phi-3-medium-128k-instruct
+                    hparams.n_swa = 131072;
+                }
+                bool found_swa = ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa, false);
+                if (!found_swa && hparams.n_swa == 0) {
+                    throw std::runtime_error("invalid value for sliding_window");
+                }
             } break;
         case LLM_ARCH_PLAMO:
             {
@@ -5236,6 +5287,23 @@ static void llm_load_hparams(
                     default: model.type = e_model::MODEL_UNKNOWN;
                 }
             } break;
+        case LLM_ARCH_NEMOTRON:
+            {
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps);
+                switch (hparams.n_layer) {
+                    case 32: model.type = e_model::MODEL_4B; break;
+                    default: model.type = e_model::MODEL_UNKNOWN;
+                }
+            } break;
+        case LLM_ARCH_EXAONE:
+            {
+                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_8B; break;
+                    default: model.type = e_model::MODEL_UNKNOWN;
+                }
+            } break;
         default: (void)0;
     }
 
@@ -5468,6 +5536,15 @@ static void llm_load_vocab(
             } else if (
                 tokenizer_pre == "codeshell") {
                 vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_CODESHELL;
+            } else if (
+                tokenizer_pre == "bloom") {
+                vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_BLOOM;
+            } else if (
+                tokenizer_pre == "gpt3-finnish") {
+                vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_GPT3_FINNISH;
+            } else if (
+                tokenizer_pre == "exaone") {
+                vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_EXAONE;
             } else {
                 throw std::runtime_error(format("unknown pre-tokenizer type: '%s'", tokenizer_pre.c_str()));
             }
@@ -7563,6 +7640,78 @@ static bool llm_load_tensors(
                         layer.ffn_down   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd});
                     }
                 } break;
+            case LLM_ARCH_NEMOTRON:
+                {
+                    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) {
+                        lm_ggml_context * ctx_layer = ctx_for_layer(i);
+                        lm_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});
+
+                        // optional bias tensors
+                        layer.bq = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q,   "bias", i), {n_embd},     llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.bk = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K,   "bias", i), {n_embd_gqa}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.bv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_V,   "bias", i), {n_embd_gqa}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.bo = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd},     llama_model_loader::TENSOR_NOT_REQUIRED);
+
+                        layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
+                        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_up   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff});
+
+                        // optional MLP bias
+                        layer.ffn_down_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.ffn_up_b   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_UP,   "bias", i), {n_ff}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                    }
+                } break;
+            case LLM_ARCH_EXAONE:
+                {
+                    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) {
+                        lm_ggml_context * ctx_layer = ctx_for_layer(i);
+                        lm_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_head_k * n_head});
+                        layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_k_gqa});
+                        layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_v_gqa});
+                        layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd});
+
+                        layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
+                        layer.rope_freqs = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ROPE_FREQS, "weight"), {n_embd/n_head/2}, llama_model_loader::TENSOR_NOT_REQUIRED | (i != 0 ? llama_model_loader::TENSOR_DUPLICATED : 0));
+                        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");
         }
@@ -8249,7 +8398,7 @@ static struct lm_ggml_tensor * llm_build_kqv(
         struct lm_ggml_tensor * kq = lm_ggml_mul_mat(ctx, k, q);
         cb(kq, "kq", il);
 
-        if (model.arch == LLM_ARCH_PHI2 || model.arch == LLM_ARCH_PHI3 || model.arch == LLM_ARCH_GPTNEOX || model.arch == LLM_ARCH_QWEN2) {
+        if (model.arch == LLM_ARCH_PHI2 || model.arch == LLM_ARCH_PHI3 || model.arch == LLM_ARCH_GPTNEOX || model.arch == LLM_ARCH_QWEN2 || model.arch == LLM_ARCH_NEMOTRON) {
             // for this arch, we need to perform the KQ multiplication with F32 precision, otherwise we get NaNs
             // ref: https://github.com/ggerganov/llama.cpp/pull/4490#issuecomment-1859055847
             lm_ggml_mul_mat_set_prec(kq, LM_GGML_PREC_F32);
@@ -13750,6 +13899,254 @@ struct llm_build_context {
 
         return gf;
     }
+
+    struct lm_ggml_cgraph * build_nemotron() {
+        struct lm_ggml_cgraph * gf = lm_ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
+
+        const int64_t n_embd_head = hparams.n_embd_head_v;
+        LM_GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+        //LM_GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+        struct lm_ggml_tensor * cur;
+        struct lm_ggml_tensor * inpL;
+
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+
+        // inp_pos - contains the positions
+        struct lm_ggml_tensor * inp_pos = build_inp_pos();
+
+        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+        struct lm_ggml_tensor * KQ_mask = build_inp_KQ_mask();
+
+        for (int il = 0; il < n_layer; ++il) {
+            struct lm_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 lm_ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
+                cb(Qcur, "Qcur", il);
+                if (model.layers[il].bq) {
+                    Qcur = lm_ggml_add(ctx0, Qcur, model.layers[il].bq);
+                    cb(Qcur, "Qcur", il);
+                }
+
+                struct lm_ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur);
+                cb(Kcur, "Kcur", il);
+                if (model.layers[il].bk) {
+                    Kcur = lm_ggml_add(ctx0, Kcur, model.layers[il].bk);
+                    cb(Kcur, "Kcur", il);
+                }
+
+                struct lm_ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur);
+                cb(Vcur, "Vcur", il);
+                if (model.layers[il].bv) {
+                    Vcur = lm_ggml_add(ctx0, Vcur, model.layers[il].bv);
+                    cb(Vcur, "Vcur", il);
+                }
+
+                Qcur = lm_ggml_rope_ext(
+                    ctx0, lm_ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Qcur, "Qcur", il);
+
+                Kcur = lm_ggml_rope_ext(
+                    ctx0, lm_ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Kcur, "Kcur", il);
+
+                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
+                        model.layers[il].wo, model.layers[il].bo,
+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+            }
+
+            if (il == n_layer - 1) {
+                // skip computing output for unused tokens
+                struct lm_ggml_tensor * inp_out_ids = build_inp_out_ids();
+                cur   = lm_ggml_get_rows(ctx0,   cur, inp_out_ids);
+                inpSA = lm_ggml_get_rows(ctx0, inpSA, inp_out_ids);
+            }
+
+            struct lm_ggml_tensor * ffn_inp = lm_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, lctx, cur,
+                    model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
+                    NULL,                      NULL,                        NULL,
+                    model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+                    NULL,
+                    LLM_FFN_RELU_SQR, LLM_FFN_SEQ, cb, il);
+
+            cur = lm_ggml_add(ctx0, cur, ffn_inp);
+            cb(cur, "ffn_out", il);
+
+            cur = lctx.cvec.apply_to(ctx0, cur, il);
+            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 = llm_build_lora_mm(lctx, ctx0, model.output, cur);
+        cb(cur, "result_output", -1);
+
+        lm_ggml_build_forward_expand(gf, cur);
+
+        return gf;
+    }
+
+    struct lm_ggml_cgraph * build_exaone() {
+        struct lm_ggml_cgraph * gf = lm_ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
+
+        // mutable variable, needed during the last layer of the computation to skip unused tokens
+        int32_t n_tokens = this->n_tokens;
+
+        const int64_t n_embd_head = hparams.n_embd_head_v;
+        LM_GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+        LM_GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+        struct lm_ggml_tensor * cur;
+        struct lm_ggml_tensor * inpL;
+
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+
+        // inp_pos - contains the positions
+        struct lm_ggml_tensor * inp_pos = build_inp_pos();
+
+        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+        struct lm_ggml_tensor * KQ_mask = build_inp_KQ_mask();
+
+        for (int il = 0; il < n_layer; ++il) {
+            struct lm_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
+            {
+                // rope freq factors for llama3; may return nullptr for llama2 and other models
+                struct lm_ggml_tensor * rope_factors = build_rope_factors(il);
+
+                // compute Q and K and RoPE them
+                struct lm_ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
+                cb(Qcur, "Qcur", il);
+                if (model.layers[il].bq) {
+                    Qcur = lm_ggml_add(ctx0, Qcur, model.layers[il].bq);
+                    cb(Qcur, "Qcur", il);
+                }
+
+                struct lm_ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur);
+                cb(Kcur, "Kcur", il);
+                if (model.layers[il].bk) {
+                    Kcur = lm_ggml_add(ctx0, Kcur, model.layers[il].bk);
+                    cb(Kcur, "Kcur", il);
+                }
+
+                struct lm_ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur);
+                cb(Vcur, "Vcur", il);
+                if (model.layers[il].bv) {
+                    Vcur = lm_ggml_add(ctx0, Vcur, model.layers[il].bv);
+                    cb(Vcur, "Vcur", il);
+                }
+
+                Qcur = lm_ggml_rope_ext(
+                    ctx0, lm_ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, rope_factors,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Qcur, "Qcur", il);
+
+                Kcur = lm_ggml_rope_ext(
+                    ctx0, lm_ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, rope_factors,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Kcur, "Kcur", il);
+
+                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
+                        model.layers[il].wo, model.layers[il].bo,
+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+            }
+
+            if (il == n_layer - 1) {
+                // skip computing output for unused tokens
+                struct lm_ggml_tensor * inp_out_ids = build_inp_out_ids();
+                n_tokens = n_outputs;
+                cur   = lm_ggml_get_rows(ctx0,   cur, inp_out_ids);
+                inpSA = lm_ggml_get_rows(ctx0, inpSA, inp_out_ids);
+            }
+
+            struct lm_ggml_tensor * ffn_inp = lm_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, lctx, cur,
+                    model.layers[il].ffn_up,   NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL,
+                    LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
+            cb(cur, "ffn_out", il);
+
+            cur = lm_ggml_add(ctx0, cur, ffn_inp);
+            cb(cur, "ffn_out", il);
+
+            cur = lctx.cvec.apply_to(ctx0, cur, il);
+            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 = llm_build_lora_mm(lctx, ctx0, model.output, cur);
+        cb(cur, "result_output", -1);
+
+        lm_ggml_build_forward_expand(gf, cur);
+
+        return gf;
+    }
 };
 
 static struct lm_ggml_cgraph * llama_build_graph_defrag(llama_context & lctx, const std::vector<uint32_t> & ids) {
@@ -14005,6 +14402,14 @@ static struct lm_ggml_cgraph * llama_build_graph(
             {
                 result = llm.build_jais();
             } break;
+        case LLM_ARCH_NEMOTRON:
+            {
+                result = llm.build_nemotron();
+            } break;
+        case LLM_ARCH_EXAONE:
+            {
+                result = llm.build_exaone();
+            } break;
         default:
             LM_GGML_ABORT("fatal error");
     }
@@ -14718,12 +15123,15 @@ static int llama_decode_internal(
             res  = nullptr;
             embd = nullptr;
         } else if (cparams.embeddings) {
-            res = nullptr; // do not extract logits for embedding case
-            embd = gf->nodes[gf->n_nodes - 1];
-            if (strcmp(embd->name, "result_embd_pooled") != 0) {
-                embd = gf->nodes[gf->n_nodes - 2];
+            res  = nullptr; // do not extract logits for embedding case
+            embd = nullptr;
+            for (int i = gf->n_nodes - 1; i >= 0; --i) {
+                if (strcmp(gf->nodes[i]->name, "result_embd_pooled") == 0) {
+                    embd = gf->nodes[i];
+                    break;
+                }
             }
-            LM_GGML_ASSERT(strcmp(embd->name, "result_embd_pooled") == 0 && "missing embeddings tensor");
+            LM_GGML_ASSERT(embd != nullptr && "missing embeddings tensor");
         } else {
             embd = nullptr; // do not extract embeddings when not needed
             LM_GGML_ASSERT(strcmp(res->name, "result_output") == 0 && "missing result_output tensor");
@@ -17072,6 +17480,8 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) {
         case LLM_ARCH_OPENELM:
         case LLM_ARCH_GPTNEOX:
         case LLM_ARCH_CODESHELL:
+        case LLM_ARCH_NEMOTRON:
+        case LLM_ARCH_EXAONE:
             return LLAMA_ROPE_TYPE_NEOX;
 
         // all model arches should be listed explicitly here
@@ -18697,11 +19107,11 @@ llama_token llama_token_pad(const struct llama_model * model) {
     return llama_token_pad_impl(model->vocab);
 }
 
-int32_t llama_add_bos_token(const struct llama_model * model) {
+bool llama_add_bos_token(const struct llama_model * model) {
     return llama_add_bos_token_impl(model->vocab);
 }
 
-int32_t llama_add_eos_token(const struct llama_model * model) {
+bool llama_add_eos_token(const struct llama_model * model) {
     return llama_add_eos_token_impl(model->vocab);
 }
 
@@ -19002,6 +19412,22 @@ static int32_t llama_chat_apply_template_internal(
         if (add_ass) {
             ss << "Assistant:";
         }
+    } else if (tmpl == "exaone3" || (tmpl_contains("[|system|]") && tmpl_contains("[|assistant|]") && tmpl_contains("[|endofturn|]"))) {
+        // ref: https://huggingface.co/LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct/discussions/8#66bae61b1893d14ee8ed85bb
+        // EXAONE-3.0-7.8B-Instruct
+        for (auto message : chat) {
+            std::string role(message->role);
+            if (role == "system") {
+                ss << "[|system|]" << trim(message->content) << "[|endofturn|]\n";
+            } else if (role == "user") {
+                ss << "[|user|]" << trim(message->content) << "\n";
+            } else if (role == "assistant") {
+                ss << "[|assistant|]" << trim(message->content) << "[|endofturn|]\n";
+            }
+        }
+        if (add_ass) {
+            ss << "[|assistant|]";
+        }
     } else {
         // template not supported
         return -1;

package/cpp/llama.h

@@ -93,15 +93,15 @@ extern "C" {
         LLAMA_VOCAB_PRE_TYPE_TEKKEN         = 20,
         LLAMA_VOCAB_PRE_TYPE_SMOLLM         = 21,
         LLAMA_VOCAB_PRE_TYPE_CODESHELL      = 22,
+        LLAMA_VOCAB_PRE_TYPE_BLOOM          = 23,
+        LLAMA_VOCAB_PRE_TYPE_GPT3_FINNISH   = 24,
+        LLAMA_VOCAB_PRE_TYPE_EXAONE         = 25,
     };
 
-    // note: these values should be synchronized with lm_ggml_rope
-    // TODO: maybe move this enum to ggml.h (lm_ggml_rope_type)
     enum llama_rope_type {
         LLAMA_ROPE_TYPE_NONE = -1,
-        LLAMA_ROPE_TYPE_NORM =  0,
-        LLAMA_ROPE_TYPE_NEOX =  2,
-        LLAMA_ROPE_TYPE_GLM  =  4,
+        LLAMA_ROPE_TYPE_NORM = 0,
+        LLAMA_ROPE_TYPE_NEOX = LM_GGML_ROPE_TYPE_NEOX,
     };
 
     enum llama_token_type { //TODO: remove, required until per token attributes are available from GGUF file
@@ -915,11 +915,8 @@ extern "C" {
     LLAMA_API llama_token llama_token_nl (const struct llama_model * model); // next-line
     LLAMA_API llama_token llama_token_pad(const struct llama_model * model); // padding
 
-    // Returns -1 if unknown, 1 for true or 0 for false.
-    LLAMA_API int32_t llama_add_bos_token(const struct llama_model * model);
-
-    // Returns -1 if unknown, 1 for true or 0 for false.
-    LLAMA_API int32_t llama_add_eos_token(const struct llama_model * model);
+    LLAMA_API bool llama_add_bos_token(const struct llama_model * model);
+    LLAMA_API bool llama_add_eos_token(const struct llama_model * model);
 
     // Codellama infill tokens
     LLAMA_API llama_token llama_token_prefix(const struct llama_model * model); // Beginning of infill prefix

package/cpp/rn-llama.hpp

@@ -297,7 +297,9 @@ struct llama_rn_context
         }
 
         // do Context Shift , may be buggy! TODO: Verify functionality
-        purge_missing_tokens(ctx, embd, prompt_tokens, params.n_predict, params.n_ctx);
+        if(!params.embedding){
+            purge_missing_tokens(ctx, embd, prompt_tokens, params.n_predict, params.n_ctx);
+        }
 
         // push the prompt into the sampling context (do not apply grammar)
         for (auto & token : prompt_tokens)
@@ -305,7 +307,7 @@ struct llama_rn_context
            llama_sampling_accept(ctx_sampling, ctx, token, false);
         }
         // compare the evaluated prompt with the new prompt
-        n_past = common_part(embd, prompt_tokens);
+        n_past = params.embedding? 0 :  common_part(embd, prompt_tokens);
         LLAMA_LOG_INFO("%s:        n_past: %zu", __func__,  n_past);
         LLAMA_LOG_INFO("%s:        embd size: %zu", __func__,  embd.size());
         LLAMA_LOG_INFO("%s:        prompt_tokens size: %zu", __func__,  prompt_tokens.size());
@@ -342,9 +344,9 @@ struct llama_rn_context
         completion_token_output result;
         result.tok = -1;
 
+        // this truncation should never trigger with good context shifting
         if (embd.size() >= (size_t)params.n_ctx)
         {
-            // Shift context
 
             const int n_left    = n_past - params.n_keep - 1;
             const int n_discard = n_left/2;
@@ -546,9 +548,21 @@ struct llama_rn_context
             LOG_WARNING("embedding disabled, embedding: %s", params.embedding);
             return std::vector<float>(n_embd, 0.0f);
         }
-        const float *data = llama_get_embeddings(ctx);
-        std::vector<float> embedding(data, data + n_embd);
-        return embedding;
+        float *data;
+        
+        if(params.pooling_type == 0){
+            data = llama_get_embeddings(ctx);
+        }
+        else {
+            data = llama_get_embeddings_seq(ctx, 0);
+        }
+        
+        if(!data) {
+            return std::vector<float>(n_embd, 0.0f);
+        }
+        std::vector<float> embedding(data, data + n_embd), out(data, data + n_embd);
+        llama_embd_normalize(embedding.data(), out.data(), n_embd, params.embd_normalize);
+        return out;
     }
 
     std::string bench(int pp, int tg, int pl, int nr)

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "cui-llama.rn",
-  "version": "1.0.9",
+  "version": "1.0.11",
   "description": "Fork of llama.rn for ChatterUI",
   "main": "lib/commonjs/index",
   "module": "lib/module/index",