cui-llama.rn 1.5.0 → 1.6.0

package/ios/rnllama.xcframework/ios-arm64_x86_64-simulator/rnllama.framework/Headers/llama-hparams.h

@@ -0,0 +1,156 @@
+#pragma once
+
+#include "llama.h"
+
+#include <array>
+
+// bump if necessary
+#define LLAMA_MAX_LAYERS  512
+#define LLAMA_MAX_EXPERTS 256  // DeepSeekV3
+
+enum llama_expert_gating_func_type {
+    LLAMA_EXPERT_GATING_FUNC_TYPE_NONE    = 0,
+    LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX = 1,
+    LLAMA_EXPERT_GATING_FUNC_TYPE_SIGMOID = 2,
+};
+
+struct llama_hparams_posnet {
+    uint32_t n_embd;
+    uint32_t n_layer;
+};
+
+struct llama_hparams_convnext {
+    uint32_t n_embd;
+    uint32_t n_layer;
+};
+
+struct llama_hparams {
+    bool vocab_only;
+    bool rope_finetuned;
+    bool use_par_res;
+    bool swin_norm;
+
+    uint32_t n_ctx_train; // context size the model was trained on
+    uint32_t n_embd;
+    uint32_t n_embd_features = 0;
+    uint32_t n_layer;
+    uint32_t n_rot;
+    uint32_t n_swa = 0; // sliding window attention (SWA)
+    uint32_t n_swa_pattern = 1; // by default, all layers use non-sliding-window attention
+    uint32_t n_embd_head_k; // dimension of keys (d_k). d_q is assumed to be the same, but there are n_head q heads, and only n_head_kv k-v heads
+    uint32_t n_embd_head_v; // dimension of values (d_v) aka n_embd_head
+    uint32_t n_expert = 0;
+    uint32_t n_expert_used = 0;
+    uint32_t n_rel_attn_bkts = 0;
+
+    // for WavTokenizer
+    struct llama_hparams_posnet   posnet;
+    struct llama_hparams_convnext convnext;
+
+    std::array<uint32_t, LLAMA_MAX_LAYERS> n_head_arr;
+    std::array<uint32_t, LLAMA_MAX_LAYERS> n_head_kv_arr;
+    std::array<uint32_t, LLAMA_MAX_LAYERS> n_ff_arr;
+
+    uint32_t n_layer_dense_lead = 0;
+    uint32_t n_lora_q           = 0;
+    uint32_t n_lora_kv          = 0;
+    uint32_t n_ff_exp           = 0;
+    uint32_t n_ff_shexp         = 0;
+    uint32_t n_expert_shared    = 0;
+    uint32_t n_norm_groups      = 0;
+
+    float    expert_weights_scale = 0.0;
+    bool     expert_weights_norm  = false;
+    uint32_t expert_gating_func   = LLAMA_EXPERT_GATING_FUNC_TYPE_NONE;
+
+    float f_norm_eps;
+    float f_norm_rms_eps;
+    float f_norm_group_eps;
+
+    float f_attn_logit_softcapping  = 50.0f;
+    float f_final_logit_softcapping = 30.0f;
+
+    // for RWKV
+    uint32_t rescale_every_n_layers = 0;
+    uint32_t time_mix_extra_dim     = 0;
+    uint32_t time_decay_extra_dim   = 0;
+    uint32_t wkv_head_size          = 0;
+    uint32_t token_shift_count      = 2;
+    uint32_t n_lora_decay           = 0;
+    uint32_t n_lora_iclr            = 0;
+    uint32_t n_lora_value_res_mix   = 0;
+    uint32_t n_lora_gate            = 0;
+
+    float    rope_attn_factor = 1.0f;
+    float    rope_freq_base_train;
+    float    rope_freq_base_train_swa;
+    float    rope_freq_scale_train;
+    float    rope_freq_scale_train_swa;
+    uint32_t n_ctx_orig_yarn;
+    float    rope_yarn_log_mul;
+
+    std::array<int, 4> rope_sections;
+
+    // for State Space Models
+    uint32_t ssm_d_conv  = 0;
+    uint32_t ssm_d_inner = 0;
+    uint32_t ssm_d_state = 0;
+    uint32_t ssm_dt_rank = 0;
+
+    bool ssm_dt_b_c_rms = false;
+
+    float f_clamp_kqv      = 0.0f;
+    float f_max_alibi_bias = 0.0f;
+    float f_logit_scale    = 0.0f;
+
+    // Additional scale factors (Granite/Granite MoE)
+    float f_residual_scale  = 0.0f;
+    float f_embedding_scale = 0.0f;
+    float f_attention_scale = 0.0f;
+
+    bool causal_attn   = true;
+    bool use_alibi     = false;
+    bool attn_soft_cap = false;
+
+    uint32_t n_moe_layer_step        = 0;
+    bool     use_kq_norm             = true;
+    uint32_t n_attn_chunk            = 0;
+    // values below seems to be fixed on llama4
+    uint32_t n_no_rope_layer_step    = 4;
+    uint32_t n_attn_temp_floor_scale = 8192;
+    float    f_attn_temp_scale       = 0.1;
+
+    // needed by encoder-decoder models (e.g. T5, FLAN-T5)
+    // ref: https://github.com/ggerganov/llama.cpp/pull/8141
+    llama_token dec_start_token_id = LLAMA_TOKEN_NULL;
+
+    enum llama_pooling_type      pooling_type            = LLAMA_POOLING_TYPE_NONE;
+    enum llama_rope_type         rope_type               = LLAMA_ROPE_TYPE_NONE;
+    enum llama_rope_scaling_type rope_scaling_type_train = LLAMA_ROPE_SCALING_TYPE_NONE;
+
+    uint32_t n_head(uint32_t il = 0) const;
+
+    uint32_t n_head_kv(uint32_t il = 0) const;
+
+    uint32_t n_ff(uint32_t il = 0) const;
+
+    uint32_t n_gqa(uint32_t il = 0) const;
+
+    // dimension of key embeddings across all k-v heads
+    uint32_t n_embd_k_gqa(uint32_t il = 0) const;
+
+    // dimension of value embeddings across all k-v heads
+    uint32_t n_embd_v_gqa(uint32_t il = 0) const;
+
+    // dimension of the rolling state embeddings
+    // corresponds to Mamba's conv_states size or RWKV's token_shift states size
+    uint32_t n_embd_k_s() const;
+
+    // dimension of the recurrent state embeddings
+    uint32_t n_embd_v_s() const;
+
+    bool is_swa(uint32_t il) const;
+};
+
+static_assert(std::is_trivially_copyable<llama_hparams>::value, "llama_hparams must be trivially copyable");
+

package/ios/rnllama.xcframework/ios-arm64_x86_64-simulator/rnllama.framework/Headers/llama-impl.h

@@ -0,0 +1,61 @@
+#pragma once
+
+#include "ggml.h" // for lm_ggml_log_level
+
+#include <string>
+#include <vector>
+
+#ifdef __GNUC__
+#    if defined(__MINGW32__) && !defined(__clang__)
+#        define LLAMA_ATTRIBUTE_FORMAT(...) __attribute__((format(gnu_printf, __VA_ARGS__)))
+#    else
+#        define LLAMA_ATTRIBUTE_FORMAT(...) __attribute__((format(printf, __VA_ARGS__)))
+#    endif
+#else
+#    define LLAMA_ATTRIBUTE_FORMAT(...)
+#endif
+
+//
+// logging
+//
+
+LLAMA_ATTRIBUTE_FORMAT(2, 3)
+void llama_log_internal        (lm_ggml_log_level level, const char * format, ...);
+void llama_log_callback_default(lm_ggml_log_level level, const char * text, void * user_data);
+
+#define LLAMA_LOG(...)       llama_log_internal(LM_GGML_LOG_LEVEL_NONE , __VA_ARGS__)
+#define LLAMA_LOG_INFO(...)  llama_log_internal(LM_GGML_LOG_LEVEL_INFO , __VA_ARGS__)
+#define LLAMA_LOG_WARN(...)  llama_log_internal(LM_GGML_LOG_LEVEL_WARN , __VA_ARGS__)
+#define LLAMA_LOG_ERROR(...) llama_log_internal(LM_GGML_LOG_LEVEL_ERROR, __VA_ARGS__)
+#define LLAMA_LOG_DEBUG(...) llama_log_internal(LM_GGML_LOG_LEVEL_DEBUG, __VA_ARGS__)
+#define LLAMA_LOG_CONT(...)  llama_log_internal(LM_GGML_LOG_LEVEL_CONT , __VA_ARGS__)
+
+//
+// helpers
+//
+
+template <typename T>
+struct no_init {
+    T value;
+    no_init() { /* do nothing */ }
+};
+
+struct time_meas {
+    time_meas(int64_t & t_acc, bool disable = false);
+    ~time_meas();
+
+    const int64_t t_start_us;
+
+    int64_t & t_acc;
+};
+
+void replace_all(std::string & s, const std::string & search, const std::string & replace);
+
+// TODO: rename to llama_format ?
+LLAMA_ATTRIBUTE_FORMAT(1, 2)
+std::string format(const char * fmt, ...);
+
+std::string llama_format_tensor_shape(const std::vector<int64_t> & ne);
+std::string llama_format_tensor_shape(const struct lm_ggml_tensor * t);
+
+std::string lm_gguf_kv_to_str(const struct lm_gguf_context * ctx_gguf, int i);

package/ios/rnllama.xcframework/ios-arm64_x86_64-simulator/rnllama.framework/Headers/llama-io.h

@@ -0,0 +1,35 @@
+#pragma once
+
+#include <cstddef>
+#include <cstdint>
+#include <string>
+
+struct lm_ggml_tensor;
+
+class llama_io_write_i {
+public:
+    llama_io_write_i() = default;
+    virtual ~llama_io_write_i() = default;
+
+    virtual void write(const void * src, size_t size) = 0;
+    virtual void write_tensor(const lm_ggml_tensor * tensor, size_t offset, size_t size) = 0;
+
+    // bytes written so far
+    virtual size_t n_bytes() = 0;
+
+    void write_string(const std::string & str);
+};
+
+class llama_io_read_i {
+public:
+    llama_io_read_i() = default;
+    virtual ~llama_io_read_i() = default;
+
+    virtual const uint8_t * read(size_t size) = 0;
+    virtual void read_to(void * dst, size_t size) = 0;
+
+    // bytes read so far
+    virtual size_t n_bytes() = 0;
+
+    void read_string(std::string & str);
+};

package/ios/rnllama.xcframework/ios-arm64_x86_64-simulator/rnllama.framework/Headers/llama-kv-cache.h

@@ -0,0 +1,213 @@
+#pragma once
+
+#include "llama.h"
+#include "llama-io.h"
+#include "llama-memory.h"
+
+#include "ggml-cpp.h"
+
+#include <functional>
+#include <set>
+#include <vector>
+
+struct llama_cparams;
+struct llama_hparams;
+struct llama_ubatch;
+
+struct llama_kv_cache : public llama_memory_i {
+    using llama_memory_i::llama_memory_i;
+
+    virtual void restore() = 0; // call if batch processing fails - restores the cache state
+    virtual void commit() = 0;  // call after successful batch processing - clears any pending state
+
+    virtual int32_t get_n_tokens()   const = 0;
+    virtual int32_t get_used_cells() const = 0; // TODO: remove, this is too-specific to the unified cache
+
+    virtual bool get_can_shift() const = 0;
+
+    bool get_can_edit() const override { return get_can_shift(); }
+};
+
+struct llama_kv_cache_guard {
+    llama_kv_cache_guard(llama_kv_cache * kv) : kv(kv) {}
+
+    ~llama_kv_cache_guard() {
+        kv->restore();
+    }
+
+    void commit() {
+        kv->commit();
+    }
+
+private:
+    llama_kv_cache * kv;
+};
+
+struct llama_kv_cell {
+    llama_pos pos   = -1;
+    llama_pos delta =  0;
+    int32_t   src   = -1; // used by recurrent state models to copy states
+    int32_t   tail  = -1;
+
+    std::set<llama_seq_id> seq_id;
+
+    bool has_seq_id(const llama_seq_id & id) const {
+        return seq_id.find(id) != seq_id.end();
+    }
+
+    bool is_empty() const {
+        return seq_id.empty();
+    }
+
+    bool is_same_seq(const llama_kv_cell & other) const {
+        return seq_id == other.seq_id;
+    }
+};
+
+// ring-buffer of cached KV data
+// TODO: pimpl
+// TODO: add notion of max sequences
+class llama_kv_cache_unified : public llama_kv_cache {
+public:
+    // can be used to query data from the model if needed
+    struct callbacks {
+        std::function<lm_ggml_tensor * (uint32_t n_ctx_per_seq, int il)> get_rope_factors;
+    };
+
+    llama_kv_cache_unified(
+            const llama_hparams & hparams,
+            callbacks             cbs);
+
+    virtual ~llama_kv_cache_unified() = default;
+
+    // TODO: become constructor
+    bool init(
+            const llama_model & model,   // TODO: do not reference the model
+          const llama_cparams & cparams,
+                    lm_ggml_type   type_k,
+                    lm_ggml_type   type_v,
+                     uint32_t   kv_size,
+                         bool   offload);
+
+    int32_t get_n_tokens()   const override;
+    int32_t get_used_cells() const override;
+
+    size_t total_size() const;
+
+    // TODO: better data structures to reduce the cost of this operation
+    llama_pos pos_max() const;
+
+    void clear() override;
+    void defrag() override;
+
+    virtual void restore() override;
+    virtual void commit() override;
+
+    bool seq_rm  (llama_seq_id seq_id,                              llama_pos p0, llama_pos p1) override;
+    void seq_cp  (llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) override;
+    void seq_keep(llama_seq_id seq_id) override;
+    void seq_add (llama_seq_id seq_id,                              llama_pos p0, llama_pos p1, llama_pos delta) override;
+    void seq_div (llama_seq_id seq_id,                              llama_pos p0, llama_pos p1, int d) override;
+
+    llama_pos seq_pos_max(llama_seq_id seq_id) const override;
+
+    bool get_can_shift() const override;
+
+    // find an empty slot of size "n_tokens" in the cache
+    // updates the cache head
+    // Note: On success, it's important that cache.head points
+    // to the first cell of the slot.
+    bool find_slot(const llama_ubatch & batch);
+
+    // TODO: maybe not needed
+    uint32_t get_padding(const llama_cparams & cparams) const;
+
+    // find how many cells are currently in use
+    uint32_t cell_max() const;
+
+    size_t size_k_bytes() const;
+    size_t size_v_bytes() const;
+
+    // defrag
+
+    struct {
+        std::vector<uint32_t> ids;
+    } defrag_info;
+
+    // return true if cells have been moved
+    bool defrag_prepare(int32_t n_max_nodes);
+
+    // commit/restore cache
+
+    struct slot_range {
+        uint32_t c0 = 0; // note: these are cell indices, not sequence positions
+        uint32_t c1 = 0;
+    };
+
+    // pending cell updates that are not yet committed
+    struct {
+        std::vector<slot_range> ranges;
+    } pending;
+
+    // state write/load
+
+    void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1) const;
+    void state_read (llama_io_read_i  & io, llama_seq_id seq_id = -1);
+
+    // members
+
+    const llama_hparams & hparams;
+
+    callbacks cbs;
+
+    bool has_shift = false;
+    bool do_defrag = false;
+
+    // TODO: remove this and implement llama_kv_cache_recurrent instead
+    bool recurrent = false; // with recurrent state models, a cell can hold the state for more than one past token
+
+    bool v_trans   = true;  // the value tensor is transposed
+    bool can_shift = false;
+
+    // Note: The value of head isn't only used to optimize searching
+    // for a free KV slot. llama_decode_impl also uses it, so it
+    // cannot be freely changed after a slot has been allocated.
+    uint32_t head = 0;
+    uint32_t size = 0;
+    uint32_t used = 0; // used cells (i.e. at least one seq_id)
+
+    // computed before each graph build
+    uint32_t n = 0;
+
+    std::vector<llama_kv_cell> cells;
+
+    std::vector<lm_ggml_tensor *> k_l; // per layer
+    std::vector<lm_ggml_tensor *> v_l;
+
+private:
+    lm_ggml_type type_k = LM_GGML_TYPE_F16;
+    lm_ggml_type type_v = LM_GGML_TYPE_F16;
+
+    std::vector<lm_ggml_context_ptr>        ctxs;
+    std::vector<lm_ggml_backend_buffer_ptr> bufs;
+
+    void state_write_meta(llama_io_write_i & io, const std::vector<std::pair<uint32_t, uint32_t>> & cell_ranges, llama_seq_id seq_id = -1) const;
+    void state_write_data(llama_io_write_i & io, const std::vector<std::pair<uint32_t, uint32_t>> & cell_ranges) const;
+
+    bool state_read_meta(llama_io_read_i & io, uint32_t cell_count, llama_seq_id dest_seq_id = -1);
+    bool state_read_data(llama_io_read_i & io, uint32_t cell_count);
+};
+
+// TODO: temporary reusing llama_kv_cache_unified -- implement recurrent cache and simplify llama_kv_cache_unified
+//class llama_kv_cache_recurrent : public llama_kv_cache_unified {
+//public:
+//    using llama_kv_cache_unified::llama_kv_cache_unified;
+//};
+
+//
+// kv cache view
+//
+
+llama_kv_cache_view llama_kv_cache_view_init(const llama_kv_cache & kv, int32_t n_seq_max);
+
+void llama_kv_cache_view_update(llama_kv_cache_view * view, const llama_kv_cache * kv);

package/ios/rnllama.xcframework/ios-arm64_x86_64-simulator/rnllama.framework/Headers/llama-memory.h

@@ -0,0 +1,21 @@
+#pragma once
+
+#include "llama.h"
+
+// general concept of LLM memory
+// the KV cache is a type of LLM memory, but there can be other types
+class llama_memory_i {
+public:
+    virtual void clear() = 0;
+    virtual void defrag() = 0;
+
+    virtual bool seq_rm  (llama_seq_id seq_id,                              llama_pos p0, llama_pos p1) = 0;
+    virtual void seq_cp  (llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) = 0;
+    virtual void seq_keep(llama_seq_id seq_id) = 0;
+    virtual void seq_add (llama_seq_id seq_id,                              llama_pos p0, llama_pos p1, llama_pos delta) = 0;
+    virtual void seq_div (llama_seq_id seq_id,                              llama_pos p0, llama_pos p1, int d) = 0;
+
+    virtual llama_pos seq_pos_max(llama_seq_id seq_id) const = 0;
+
+    virtual bool get_can_edit() const = 0;
+};

package/ios/rnllama.xcframework/ios-arm64_x86_64-simulator/rnllama.framework/Headers/llama-mmap.h

@@ -0,0 +1,68 @@
+#pragma once
+
+#include <cstdint>
+#include <memory>
+#include <vector>
+
+struct llama_file;
+struct llama_mmap;
+struct llama_mlock;
+
+using llama_files  = std::vector<std::unique_ptr<llama_file>>;
+using llama_mmaps  = std::vector<std::unique_ptr<llama_mmap>>;
+using llama_mlocks = std::vector<std::unique_ptr<llama_mlock>>;
+
+struct llama_file {
+    llama_file(const char * fname, const char * mode);
+    ~llama_file();
+
+    size_t tell() const;
+    size_t size() const;
+
+    int file_id() const; // fileno overload
+
+    void seek(size_t offset, int whence) const;
+
+    void read_raw(void * ptr, size_t len) const;
+    uint32_t read_u32() const;
+
+    void write_raw(const void * ptr, size_t len) const;
+    void write_u32(uint32_t val) const;
+
+private:
+    struct impl;
+    std::unique_ptr<impl> pimpl;
+};
+
+struct llama_mmap {
+    llama_mmap(const llama_mmap &) = delete;
+    llama_mmap(struct llama_file * file, size_t prefetch = (size_t) -1, bool numa = false);
+    ~llama_mmap();
+
+    size_t size() const;
+    void * addr() const;
+
+    void unmap_fragment(size_t first, size_t last);
+
+    static const bool SUPPORTED;
+
+private:
+    struct impl;
+    std::unique_ptr<impl> pimpl;
+};
+
+struct llama_mlock {
+    llama_mlock();
+    ~llama_mlock();
+
+    void init(void * ptr);
+    void grow_to(size_t target_size);
+
+    static const bool SUPPORTED;
+
+private:
+    struct impl;
+    std::unique_ptr<impl> pimpl;
+};
+
+size_t llama_path_max();