cui-llama.rn 1.5.0 → 1.6.0

package/ios/rnllama.xcframework/ios-arm64/rnllama.framework/Headers/llama-model-loader.h

@@ -0,0 +1,169 @@
+#pragma once
+
+#include "llama.h"
+
+#include "llama-impl.h"
+#include "llama-arch.h"
+#include "llama-mmap.h"
+
+#include "ggml-cpp.h"
+
+#include <cstddef>
+#include <map>
+#include <stdexcept>
+#include <unordered_map>
+
+using llama_buf_map = std::unordered_map<uint32_t, lm_ggml_backend_buffer_t>;
+
+enum llama_fver {
+    LM_GGUF_FILE_VERSION_V1 = 1,
+    LM_GGUF_FILE_VERSION_V2 = 2,
+    LM_GGUF_FILE_VERSION_V3 = 3,
+};
+
+const char * llama_file_version_name(llama_fver version);
+
+struct llama_model_loader {
+    // Holds information on a model weight
+    struct llama_tensor_weight {
+        uint16_t  idx; // source file index
+        size_t   offs; // tensor data offset in the original file
+
+        lm_ggml_tensor * tensor;
+
+        llama_tensor_weight(const llama_file * file, uint16_t idx, const struct lm_gguf_context * lm_gguf_ctx, lm_ggml_tensor * tensor) : idx(idx), tensor(tensor) {
+            const int tensor_idx = lm_gguf_find_tensor(lm_gguf_ctx,  lm_ggml_get_name(tensor));
+            if (tensor_idx < 0) {
+                throw std::runtime_error(format("tensor '%s' not found in the model", lm_ggml_get_name(tensor)));
+            }
+
+            offs = lm_gguf_get_data_offset(lm_gguf_ctx) + lm_gguf_get_tensor_offset(lm_gguf_ctx, tensor_idx);
+            if (offs + lm_ggml_nbytes(tensor) < offs || offs + lm_ggml_nbytes(tensor) > file->size()) {
+                throw std::runtime_error(format("tensor '%s' data is not within the file bounds, model is corrupted or incomplete", lm_ggml_get_name(tensor)));
+            }
+        }
+    };
+
+    // custom comparator to sort weights more nicely by layer
+    struct weight_name_comparer {
+        bool operator()(const std::string & a, const std::string & b) const {
+            int a_layer = -1;
+            int b_layer = -1;
+            sscanf(a.c_str(), "blk.%d.", &a_layer);
+            sscanf(b.c_str(), "blk.%d.", &b_layer);
+            if (a_layer != b_layer) {
+                return a_layer < b_layer;
+            }
+            return a < b;
+        }
+    };
+
+    static const int TENSOR_NOT_REQUIRED = 1;
+    static const int TENSOR_DUPLICATED   = 2;
+
+    int n_kv      = 0;
+    int n_tensors = 0;
+    int n_created = 0;
+
+    uint64_t n_elements = 0;
+    size_t   n_bytes    = 0;
+
+    bool use_mmap = false;
+    bool check_tensors;
+
+    llama_files files;
+    llama_ftype ftype;
+    llama_fver  fver;
+
+    llama_mmaps mappings;
+
+    std::map<std::string, llama_tensor_weight, weight_name_comparer> weights_map;
+    std::unordered_map<std::string, llama_model_kv_override> kv_overrides;
+    const llama_model_tensor_buft_override * tensor_buft_overrides;
+
+    lm_gguf_context_ptr meta;
+    std::vector<lm_ggml_context_ptr> contexts;
+
+    std::string arch_name;
+    LLM_KV      llm_kv    = LLM_KV(LLM_ARCH_UNKNOWN);
+
+    size_t size_done = 0;
+    size_t size_data = 0;
+    std::vector<std::pair<size_t, size_t>> mmaps_used;
+
+    llama_model_loader(
+        const std::string & fname,
+        std::vector<std::string> & splits, // optional, only need if the split does not follow naming scheme
+        bool use_mmap,
+        bool check_tensors,
+        const llama_model_kv_override * param_overrides_p,
+        const llama_model_tensor_buft_override * param_tensor_buft_overrides_p);
+
+    template<typename T>
+    typename std::enable_if<std::is_integral<T>::value, bool>::type
+    get_arr_n(const std::string & key, T & result, bool required = true);
+
+    template<typename T>
+    typename std::enable_if<std::is_integral<T>::value, bool>::type
+    get_arr_n(enum llm_kv kid, T & result, bool required = true);
+
+    template<typename T>
+    bool get_arr(const std::string & key, std::vector<T> & result, bool required = true);
+
+    template<typename T, size_t N_MAX>
+    bool get_arr(const std::string & key, std::array<T, N_MAX> & result, bool required = true);
+
+    template<typename T>
+    bool get_arr(enum llm_kv kid, T & result, bool required = true);
+
+    template<typename T>
+    bool get_key(const std::string & key, T & result, bool required = true);
+
+    template<typename T>
+    bool get_key(enum llm_kv kid, T & result, bool required = true);
+
+    template<typename T, size_t N_MAX>
+    bool get_key_or_arr(const std::string & key, std::array<T, N_MAX> & result, uint32_t n, bool required = true);
+
+    template<typename T>
+    bool get_key_or_arr(enum llm_kv kid, T & result, uint32_t n, bool required = true);
+
+    std::string get_arch_name() const;
+
+    enum llm_arch get_arch() const;
+
+    const llama_tensor_weight * get_weight(const char * name) const;
+
+    const llama_tensor_weight & require_weight(const char * name) const;
+
+    struct lm_ggml_tensor * get_tensor_meta(const char * name) const;
+
+    struct lm_ggml_tensor * require_tensor_meta(const std::string & name) const;
+
+    const struct lm_ggml_tensor * check_tensor_dims(const std::string & name, const std::vector<int64_t> & ne, bool required) const;
+
+    struct lm_ggml_tensor * create_tensor(struct lm_ggml_context * ctx, const std::string & name, const std::initializer_list<int64_t> & ne, int flags = 0);
+
+    struct lm_ggml_tensor * create_tensor_as_view(struct lm_ggml_context * ctx, struct lm_ggml_tensor * base, const std::string & name, const std::initializer_list<int64_t> & ne, size_t offset, bool required = true);
+
+    void done_getting_tensors() const;
+
+    void init_mappings(bool prefetch = true, llama_mlocks * mlock_mmaps = nullptr);
+
+    void get_mapping_range(size_t * first, size_t * last, void ** addr, int idx, lm_ggml_context * ctx) const;
+
+    // for backwards compatibility, does not support ggml-backend
+    void load_data_for(struct lm_ggml_tensor * cur) const;
+
+    // Returns false if cancelled by progress_callback
+    bool load_all_data(
+            struct lm_ggml_context * ctx,
+            llama_buf_map & bufs,
+            llama_mlocks * lmlocks,
+            llama_progress_callback progress_callback,
+            void * progress_callback_user_data);
+
+    std::string ftype_name() const;
+
+    void print_info() const;
+};

package/ios/rnllama.xcframework/ios-arm64/rnllama.framework/Headers/llama-model.h

@@ -0,0 +1,409 @@
+#pragma once
+
+#include "llama.h"
+#include "llama-arch.h"
+#include "llama-graph.h"
+#include "llama-hparams.h"
+#include "llama-memory.h"
+#include "llama-vocab.h"
+
+#include <memory>
+#include <string>
+#include <unordered_map>
+#include <vector>
+
+struct llama_cparams;
+struct llama_ubatch;
+struct llama_model_loader;
+
+// available models
+enum llm_type {
+    LLM_TYPE_UNKNOWN,
+    LLM_TYPE_14M,
+    LLM_TYPE_17M,
+    LLM_TYPE_22M,
+    LLM_TYPE_33M,
+    LLM_TYPE_60M,
+    LLM_TYPE_70M,
+    LLM_TYPE_80M,
+    LLM_TYPE_109M,
+    LLM_TYPE_137M,
+    LLM_TYPE_160M,
+    LLM_TYPE_190M,
+    LLM_TYPE_220M,
+    LLM_TYPE_250M,
+    LLM_TYPE_270M,
+    LLM_TYPE_335M,
+    LLM_TYPE_410M,
+    LLM_TYPE_450M,
+    LLM_TYPE_770M,
+    LLM_TYPE_780M,
+    LLM_TYPE_0_5B,
+    LLM_TYPE_1B,
+    LLM_TYPE_1_3B,
+    LLM_TYPE_1_4B,
+    LLM_TYPE_1_5B,
+    LLM_TYPE_1_6B,
+    LLM_TYPE_1_8B,
+    LLM_TYPE_2B,
+    LLM_TYPE_2_8B,
+    LLM_TYPE_2_9B,
+    LLM_TYPE_3B,
+    LLM_TYPE_4B,
+    LLM_TYPE_6B,
+    LLM_TYPE_6_9B,
+    LLM_TYPE_7B,
+    LLM_TYPE_8B,
+    LLM_TYPE_9B,
+    LLM_TYPE_11B,
+    LLM_TYPE_12B,
+    LLM_TYPE_13B,
+    LLM_TYPE_14B,
+    LLM_TYPE_15B,
+    LLM_TYPE_16B,
+    LLM_TYPE_20B,
+    LLM_TYPE_30B,
+    LLM_TYPE_32B,
+    LLM_TYPE_34B,
+    LLM_TYPE_35B,
+    LLM_TYPE_40B,
+    LLM_TYPE_65B,
+    LLM_TYPE_70B,
+    LLM_TYPE_236B,
+    LLM_TYPE_314B,
+    LLM_TYPE_671B,
+    LLM_TYPE_SMALL,
+    LLM_TYPE_MEDIUM,
+    LLM_TYPE_LARGE,
+    LLM_TYPE_XL,
+    LLM_TYPE_A1_7B,
+    LLM_TYPE_A2_7B,
+    LLM_TYPE_8x7B,
+    LLM_TYPE_8x22B,
+    LLM_TYPE_16x12B,
+    LLM_TYPE_16x3_8B,
+    LLM_TYPE_10B_128x3_66B,
+    LLM_TYPE_57B_A14B,
+    LLM_TYPE_27B,
+    LLM_TYPE_290B,
+    LLM_TYPE_17B_16E, // llama4 Scout
+    LLM_TYPE_17B_128E, // llama4 Maverick
+};
+
+struct llama_layer_posnet {
+    // resnet
+    struct lm_ggml_tensor * norm1   = nullptr;
+    struct lm_ggml_tensor * norm1_b = nullptr;
+
+    struct lm_ggml_tensor * conv1   = nullptr;
+    struct lm_ggml_tensor * conv1_b = nullptr;
+
+    struct lm_ggml_tensor * norm2   = nullptr;
+    struct lm_ggml_tensor * norm2_b = nullptr;
+
+    struct lm_ggml_tensor * conv2   = nullptr;
+    struct lm_ggml_tensor * conv2_b = nullptr;
+
+    // attention
+    struct lm_ggml_tensor * attn_norm   = nullptr;
+    struct lm_ggml_tensor * attn_norm_b = nullptr;
+
+    struct lm_ggml_tensor * attn_q   = nullptr;
+    struct lm_ggml_tensor * attn_q_b = nullptr;
+
+    struct lm_ggml_tensor * attn_k   = nullptr;
+    struct lm_ggml_tensor * attn_k_b = nullptr;
+
+    struct lm_ggml_tensor * attn_v   = nullptr;
+    struct lm_ggml_tensor * attn_v_b = nullptr;
+
+    struct lm_ggml_tensor * attn_o   = nullptr;
+    struct lm_ggml_tensor * attn_o_b = nullptr;
+
+    // normalize
+    struct lm_ggml_tensor * norm   = nullptr;
+    struct lm_ggml_tensor * norm_b = nullptr;
+};
+
+struct llama_layer_convnext {
+    struct lm_ggml_tensor * dw   = nullptr;
+    struct lm_ggml_tensor * dw_b = nullptr;
+
+    struct lm_ggml_tensor * norm   = nullptr;
+    struct lm_ggml_tensor * norm_b = nullptr;
+
+    struct lm_ggml_tensor * pw1   = nullptr;
+    struct lm_ggml_tensor * pw1_b = nullptr;
+
+    struct lm_ggml_tensor * pw2   = nullptr;
+    struct lm_ggml_tensor * pw2_b = nullptr;
+
+    struct lm_ggml_tensor * gamma = nullptr;
+};
+
+struct llama_layer {
+    // normalization
+    struct lm_ggml_tensor * attn_norm       = nullptr;
+    struct lm_ggml_tensor * attn_norm_b     = nullptr;
+    struct lm_ggml_tensor * attn_norm_2     = nullptr;
+    struct lm_ggml_tensor * attn_norm_2_b   = nullptr;
+    struct lm_ggml_tensor * attn_q_norm     = nullptr;
+    struct lm_ggml_tensor * attn_q_norm_b   = nullptr;
+    struct lm_ggml_tensor * attn_k_norm     = nullptr;
+    struct lm_ggml_tensor * attn_k_norm_b   = nullptr;
+    struct lm_ggml_tensor * attn_out_norm   = nullptr;
+    struct lm_ggml_tensor * attn_out_norm_b = nullptr;
+    struct lm_ggml_tensor * attn_q_a_norm   = nullptr;
+    struct lm_ggml_tensor * attn_kv_a_norm  = nullptr;
+    struct lm_ggml_tensor * attn_sub_norm   = nullptr;
+    struct lm_ggml_tensor * attn_post_norm  = nullptr;
+    struct lm_ggml_tensor * ffn_sub_norm    = nullptr;
+    struct lm_ggml_tensor * attn_norm_cross = nullptr;
+    struct lm_ggml_tensor * attn_norm_enc   = nullptr;
+
+    // attention
+    struct lm_ggml_tensor * wq        = nullptr;
+    struct lm_ggml_tensor * wk        = nullptr;
+    struct lm_ggml_tensor * wv        = nullptr;
+    struct lm_ggml_tensor * wo        = nullptr;
+    struct lm_ggml_tensor * wqkv      = nullptr;
+    struct lm_ggml_tensor * wq_a      = nullptr;
+    struct lm_ggml_tensor * wq_b      = nullptr;
+    struct lm_ggml_tensor * wkv_a_mqa = nullptr;
+    struct lm_ggml_tensor * wkv_b     = nullptr;
+    struct lm_ggml_tensor * wq_cross  = nullptr;
+    struct lm_ggml_tensor * wk_cross  = nullptr;
+    struct lm_ggml_tensor * wv_cross  = nullptr;
+    struct lm_ggml_tensor * wo_cross  = nullptr;
+    struct lm_ggml_tensor * wq_enc    = nullptr;
+    struct lm_ggml_tensor * wk_enc    = nullptr;
+    struct lm_ggml_tensor * wv_enc    = nullptr;
+    struct lm_ggml_tensor * wo_enc    = nullptr;
+
+    // attention bias
+    struct lm_ggml_tensor * bq   = nullptr;
+    struct lm_ggml_tensor * bk   = nullptr;
+    struct lm_ggml_tensor * bv   = nullptr;
+    struct lm_ggml_tensor * bo   = nullptr;
+    struct lm_ggml_tensor * bqkv = nullptr;
+
+    // relative position bias
+    struct lm_ggml_tensor * attn_rel_b       = nullptr;
+    struct lm_ggml_tensor * attn_rel_b_enc   = nullptr;
+    struct lm_ggml_tensor * attn_rel_b_cross = nullptr;
+
+    // normalization
+    struct lm_ggml_tensor * ffn_norm         = nullptr;
+    struct lm_ggml_tensor * ffn_norm_b       = nullptr;
+    struct lm_ggml_tensor * ffn_post_norm    = nullptr;
+    struct lm_ggml_tensor * layer_out_norm   = nullptr;
+    struct lm_ggml_tensor * layer_out_norm_b = nullptr;
+    struct lm_ggml_tensor * ffn_norm_exps    = nullptr;
+    struct lm_ggml_tensor * ffn_norm_enc     = nullptr;
+
+    // ff
+    struct lm_ggml_tensor * ffn_gate     = nullptr; // w1
+    struct lm_ggml_tensor * ffn_down     = nullptr; // w2
+    struct lm_ggml_tensor * ffn_up       = nullptr; // w3
+    struct lm_ggml_tensor * ffn_gate_enc = nullptr;
+    struct lm_ggml_tensor * ffn_down_enc = nullptr;
+    struct lm_ggml_tensor * ffn_up_enc   = nullptr;
+
+    // ff MoE
+    struct lm_ggml_tensor * ffn_gate_inp  = nullptr;
+    struct lm_ggml_tensor * ffn_gate_exps = nullptr;
+    struct lm_ggml_tensor * ffn_down_exps = nullptr;
+    struct lm_ggml_tensor * ffn_up_exps   = nullptr;
+
+    // ff shared expert (shexp)
+    struct lm_ggml_tensor * ffn_gate_inp_shexp = nullptr;
+    struct lm_ggml_tensor * ffn_gate_shexp     = nullptr;
+    struct lm_ggml_tensor * ffn_down_shexp     = nullptr;
+    struct lm_ggml_tensor * ffn_up_shexp       = nullptr;
+
+    // ff bias
+    struct lm_ggml_tensor * ffn_gate_b = nullptr;
+    struct lm_ggml_tensor * ffn_down_b = nullptr; // b2
+    struct lm_ggml_tensor * ffn_up_b   = nullptr; // b3
+    struct lm_ggml_tensor * ffn_act    = nullptr;
+    struct lm_ggml_tensor * ffn_exp_probs_b = nullptr;
+
+    // mamba proj
+    struct lm_ggml_tensor * ssm_in  = nullptr;
+    struct lm_ggml_tensor * ssm_x   = nullptr;
+    struct lm_ggml_tensor * ssm_dt  = nullptr;
+    struct lm_ggml_tensor * ssm_out = nullptr;
+
+    // mamba
+    struct lm_ggml_tensor * ssm_conv1d = nullptr;
+    struct lm_ggml_tensor * ssm_a      = nullptr;
+    struct lm_ggml_tensor * ssm_d      = nullptr;
+
+    // mamba bias
+    struct lm_ggml_tensor * ssm_conv1d_b = nullptr;
+    struct lm_ggml_tensor * ssm_dt_b     = nullptr;
+
+    // rwkv
+    struct lm_ggml_tensor * time_mix_w1         = nullptr;
+    struct lm_ggml_tensor * time_mix_w2         = nullptr;
+    struct lm_ggml_tensor * time_mix_lerp_x     = nullptr;
+    struct lm_ggml_tensor * time_mix_lerp_w     = nullptr;
+    struct lm_ggml_tensor * time_mix_lerp_k     = nullptr;
+    struct lm_ggml_tensor * time_mix_lerp_v     = nullptr;
+    struct lm_ggml_tensor * time_mix_lerp_r     = nullptr;
+    struct lm_ggml_tensor * time_mix_lerp_g     = nullptr;
+    struct lm_ggml_tensor * time_mix_lerp_fused = nullptr;
+
+    struct lm_ggml_tensor * time_mix_first        = nullptr;
+    struct lm_ggml_tensor * time_mix_decay        = nullptr;
+    struct lm_ggml_tensor * time_mix_decay_w1     = nullptr;
+    struct lm_ggml_tensor * time_mix_decay_w2     = nullptr;
+    struct lm_ggml_tensor * time_mix_key          = nullptr;
+    struct lm_ggml_tensor * time_mix_key_b        = nullptr;
+    struct lm_ggml_tensor * time_mix_value        = nullptr;
+    struct lm_ggml_tensor * time_mix_value_b      = nullptr;
+    struct lm_ggml_tensor * time_mix_receptance   = nullptr;
+    struct lm_ggml_tensor * time_mix_receptance_b = nullptr;
+    struct lm_ggml_tensor * time_mix_gate         = nullptr;
+
+    // rwkv7
+    struct lm_ggml_tensor * time_mix_w0         = nullptr;
+    struct lm_ggml_tensor * time_mix_a0         = nullptr;
+    struct lm_ggml_tensor * time_mix_a1         = nullptr;
+    struct lm_ggml_tensor * time_mix_a2         = nullptr;
+    struct lm_ggml_tensor * time_mix_v0         = nullptr;
+    struct lm_ggml_tensor * time_mix_v1         = nullptr;
+    struct lm_ggml_tensor * time_mix_v2         = nullptr;
+    struct lm_ggml_tensor * time_mix_g1         = nullptr;
+    struct lm_ggml_tensor * time_mix_g2         = nullptr;
+    struct lm_ggml_tensor * time_mix_k_k        = nullptr;
+    struct lm_ggml_tensor * time_mix_k_a        = nullptr;
+    struct lm_ggml_tensor * time_mix_r_k        = nullptr;
+
+    struct lm_ggml_tensor * time_mix_ln     = nullptr;
+    struct lm_ggml_tensor * time_mix_ln_b   = nullptr;
+    struct lm_ggml_tensor * time_mix_output = nullptr;
+
+    struct lm_ggml_tensor * channel_mix_lerp_k = nullptr;
+    struct lm_ggml_tensor * channel_mix_lerp_r = nullptr;
+
+    struct lm_ggml_tensor * channel_mix_key        = nullptr;
+    struct lm_ggml_tensor * channel_mix_receptance = nullptr;
+    struct lm_ggml_tensor * channel_mix_value      = nullptr;
+
+    // long rope factors
+    struct lm_ggml_tensor * rope_long  = nullptr;
+    struct lm_ggml_tensor * rope_short = nullptr;
+    struct lm_ggml_tensor * rope_freqs = nullptr;
+
+    // bitnet scale
+    struct lm_ggml_tensor * wq_scale       = nullptr;
+    struct lm_ggml_tensor * wk_scale       = nullptr;
+    struct lm_ggml_tensor * wv_scale       = nullptr;
+    struct lm_ggml_tensor * wo_scale       = nullptr;
+    struct lm_ggml_tensor * ffn_gate_scale = nullptr;
+    struct lm_ggml_tensor * ffn_up_scale   = nullptr;
+    struct lm_ggml_tensor * ffn_down_scale = nullptr;
+
+    struct llama_layer_posnet posnet;
+
+    struct llama_layer_convnext convnext;
+};
+
+struct llama_model {
+    llm_type type = LLM_TYPE_UNKNOWN;
+    llm_arch arch = LLM_ARCH_UNKNOWN;
+
+    std::string name = "n/a";
+
+    llama_hparams hparams = {};
+    llama_vocab   vocab;
+
+    struct lm_ggml_tensor * tok_embd   = nullptr;
+    struct lm_ggml_tensor * type_embd  = nullptr;
+    struct lm_ggml_tensor * pos_embd   = nullptr;
+    struct lm_ggml_tensor * tok_norm   = nullptr;
+    struct lm_ggml_tensor * tok_norm_b = nullptr;
+
+    struct lm_ggml_tensor * output_norm     = nullptr;
+    struct lm_ggml_tensor * output_norm_b   = nullptr;
+    struct lm_ggml_tensor * output          = nullptr;
+    struct lm_ggml_tensor * output_b        = nullptr;
+    struct lm_ggml_tensor * output_norm_enc = nullptr;
+
+    // classifier
+    struct lm_ggml_tensor * cls       = nullptr;
+    struct lm_ggml_tensor * cls_b     = nullptr;
+    struct lm_ggml_tensor * cls_out   = nullptr;
+    struct lm_ggml_tensor * cls_out_b = nullptr;
+
+    struct lm_ggml_tensor * conv1d   = nullptr;
+    struct lm_ggml_tensor * conv1d_b = nullptr;
+
+    std::vector<llama_layer> layers;
+
+    llama_model_params params;
+
+    // gguf metadata
+    std::unordered_map<std::string, std::string> lm_gguf_kv;
+
+    // list of devices used in this model
+    std::vector<lm_ggml_backend_dev_t> devices;
+
+    // for quantize-stats only
+    std::vector<std::pair<std::string, struct lm_ggml_tensor *>> tensors_by_name;
+
+    int64_t t_load_us  = 0;
+    int64_t t_start_us = 0;
+
+    explicit llama_model(const struct llama_model_params & params);
+    ~llama_model();
+
+    void load_stats  (llama_model_loader & ml);
+    void load_arch   (llama_model_loader & ml);
+    void load_hparams(llama_model_loader & ml);
+    void load_vocab  (llama_model_loader & ml);
+    bool load_tensors(llama_model_loader & ml); // returns false if cancelled by progress_callback
+
+    std::string arch_name() const;
+    std::string type_name() const;
+
+    std::string desc() const;
+
+    size_t size() const;
+    size_t n_tensors() const;
+    size_t n_devices() const;
+
+    // total number of parameters in the model
+    uint64_t n_elements() const;
+
+    void print_info() const;
+
+    lm_ggml_backend_dev_t dev_layer(int il) const;
+    lm_ggml_backend_dev_t dev_output() const;
+
+    lm_ggml_backend_buffer_type_t select_buft(int il) const;
+
+    bool has_tensor_overrides() const;
+
+    const struct lm_ggml_tensor * get_tensor(const char * name) const;
+
+    // TODO: move this to new llm_arch_model_i interface
+    llama_memory_i * create_memory() const; // TODO: params
+
+    // TODO: move this to new llm_arch_model_i interface
+    llm_graph_result_ptr build_graph(
+            const llm_graph_params & params,
+                       lm_ggml_cgraph * gf,
+                    llm_graph_type   type) const;
+
+private:
+    struct impl;
+    std::unique_ptr<impl> pimpl;
+};
+
+const char * llm_type_name(llm_type type);
+
+// For internal test use
+// TODO: remove
+const std::vector<std::pair<std::string, lm_ggml_tensor *>> & llama_internal_get_tensor_map(const llama_model * model);

package/ios/rnllama.xcframework/ios-arm64/rnllama.framework/Headers/llama-sampling.h

@@ -0,0 +1,32 @@
+#pragma once
+
+// TODO: rename llama-sampling.h/.cpp to llama-sampler.h/.cpp ?
+
+#include "llama.h"
+
+#include <vector>
+
+struct llama_vocab;
+struct llama_grammar;
+
+// sampler chain
+
+struct llama_sampler_chain {
+    llama_sampler_chain_params params;
+
+    std::vector<struct llama_sampler *> samplers;
+
+    // timing
+
+    mutable int64_t t_sample_us;
+
+    mutable int32_t n_sample;
+};
+
+struct llama_sampler * llama_sampler_init_dry_testing(
+                         int32_t   context_size,
+                           float   dry_multiplier,
+                           float   dry_base,
+                         int32_t   dry_allowed_length,
+                         int32_t   dry_penalty_last_n,
+  const std::vector<std::vector<llama_token>>& seq_breakers);