cui-llama.rn 1.6.1 → 1.7.0

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

@@ -8,6 +8,7 @@
 #include "ggml-cpp.h"
 
 #include <set>
+#include <unordered_map>
 #include <vector>
 
 struct llama_cparams;
@@ -40,6 +41,9 @@ struct llama_kv_cache : public llama_memory_i {
     // batch processing
     //
 
+    // =============================================================================================================
+    // TODO: refactor  and simplify this
+
     virtual llama_sbatch sbatch_init(const llama_batch & batch, bool logits_all) = 0;
 
     // different KV caches require different batch splitting strategies
@@ -48,11 +52,10 @@ struct llama_kv_cache : public llama_memory_i {
     // find an empty slot of size "n_tokens" in the cache
     virtual bool find_slot(const llama_ubatch & batch) = 0;
 
+    // =============================================================================================================
+
     // getters
-    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 llama_pos get_pos_max()    const = 0;
-    virtual bool      get_can_shift()  const = 0;
+    virtual bool get_can_shift() const = 0;
 
     bool get_can_edit() const override { return get_can_shift(); }
 
@@ -87,38 +90,25 @@ private:
 // llama_kv_cache_unified
 //
 
-// TODO: add notion of max sequences
 class llama_kv_cache_unified : public llama_kv_cache {
 public:
-    struct kv_cell {
-        llama_pos pos   = -1;
-        llama_pos delta =  0;
-
-        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 kv_cell & other) const {
-            return seq_id == other.seq_id;
-        }
-    };
-
     static uint32_t get_padding(const llama_cparams & cparams);
 
+    // this callback is used to filter out layers that should not be included in the cache
+    using layer_filter_cb = std::function<bool(int32_t il)>;
+
     llama_kv_cache_unified(
-            const llama_model & model,
-                    lm_ggml_type   type_k,
-                    lm_ggml_type   type_v,
-                         bool   v_trans,
-                         bool   offload,
-                     uint32_t   kv_size,
-                     uint32_t   padding);
+            const llama_model &  model,
+              layer_filter_cb && filter,
+                    lm_ggml_type    type_k,
+                    lm_ggml_type    type_v,
+                         bool    v_trans,
+                         bool    offload,
+                     uint32_t    kv_size,
+                     uint32_t    n_seq_max,
+                     uint32_t    n_pad,
+                     uint32_t    n_swa,
+               llama_swa_type    swa_type);
 
     ~llama_kv_cache_unified() = default;
 
@@ -130,10 +120,11 @@ public:
 
     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_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_min(llama_seq_id seq_id) const override;
     llama_pos seq_pos_max(llama_seq_id seq_id) const override;
 
     //
@@ -150,7 +141,6 @@ public:
     void set_full() override;
 
     llama_sbatch sbatch_init(const llama_batch & batch, bool logits_all) override;
-
     llama_ubatch ubatch_next(llama_sbatch & sbatch, uint32_t n_ubatch, bool embd_pooled) const override;
 
     // updates the cache head
@@ -158,53 +148,106 @@ public:
     // to the first cell of the slot.
     bool find_slot(const llama_ubatch & batch) override;
 
-    int32_t get_n_tokens()   const override;
-    int32_t get_used_cells() const override;
-
-    // TODO: better data structures to reduce the cost of this operation
-    llama_pos get_pos_max() const override;
-
     bool get_can_shift() const override;
 
     // state write/load
 
     void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1) const override;
-    void state_read (llama_io_read_i  & io, llama_seq_id seq_id = -1) override;
+    void state_read (llama_io_read_i  & io, llama_seq_id seq_id = -1)       override;
 
-    // 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)
+    //
+    // llama_kv_cache_unified specific API
+    //
 
-    // computed before each graph build
-    uint32_t n = 0;
+    uint32_t get_n() const;
+    uint32_t get_size() const;
 
-    std::vector<kv_cell> cells;
+    // get views of the current state of the cache
+    lm_ggml_tensor * get_k(lm_ggml_context * ctx, int32_t il) const;
+    lm_ggml_tensor * get_v(lm_ggml_context * ctx, int32_t il) const;
 
-    std::vector<lm_ggml_tensor *> k_l; // per layer
-    std::vector<lm_ggml_tensor *> v_l;
+    // store k_cur and v_cur in the cache based on the current head location
+    lm_ggml_tensor * cpy_k(lm_ggml_context * ctx, lm_ggml_tensor * k_cur, int32_t il) const;
+    lm_ggml_tensor * cpy_v(lm_ggml_context * ctx, lm_ggml_tensor * v_cur, int32_t il) const;
+
+    void prune_swa(llama_seq_id seq_id, llama_pos pmin, llama_pos pmax);
+
+    void set_input_kq_mask   (lm_ggml_tensor * dst, const llama_ubatch * ubatch, bool causal_attn) const;
+    void set_input_k_shift   (lm_ggml_tensor * dst) const;
+    void set_input_pos_bucket(lm_ggml_tensor * dst, const llama_ubatch * ubatch) const;
 
 private:
     const llama_model & model;
     const llama_hparams & hparams;
 
+    struct kv_cell {
+        llama_pos pos   = -1;
+        llama_pos delta =  0;
+
+        // TODO: replace with bitset uint64_t
+        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 kv_cell & other) const {
+            return seq_id == other.seq_id;
+        }
+    };
+
+    struct kv_layer {
+        // layer index in the model
+        // note: can be different from the layer index in the KV cache
+        uint32_t il;
+
+        lm_ggml_tensor * k;
+        lm_ggml_tensor * v;
+    };
+
     bool has_shift = false;
     bool do_defrag = false;
-
     bool v_trans   = true;  // the value tensor is transposed
-    bool can_shift = false;
+
+    uint32_t head = 0; // the location where the batch will be placed in the cache (see find_slot())
+    uint32_t size = 0; // total number of cells, shared across all sequences
+    uint32_t used = 0; // used cells (i.e. at least one seq_id) (TODO: add `struct kv_cells` and keep track automaticallt)
+
+    // computed before each graph build
+    uint32_t n = 0;
+
+    const uint32_t n_seq_max = 1;
 
     // required padding
-    uint32_t padding = 1;
+    const uint32_t n_pad = 1;
+
+    // SWA
+    const uint32_t n_swa = 0;
 
-    lm_ggml_type type_k = LM_GGML_TYPE_F16;
-    lm_ggml_type type_v = LM_GGML_TYPE_F16;
+    const llama_swa_type swa_type = LLAMA_SWA_TYPE_NONE;
 
     std::vector<lm_ggml_context_ptr>        ctxs;
     std::vector<lm_ggml_backend_buffer_ptr> bufs;
 
+    std::vector<kv_cell>  cells;  // TODO: replace with `struct kv_cells`
+    std::vector<kv_layer> layers;
+
+    // model layer id -> KV cache layer id
+    std::unordered_map<int32_t, int32_t> map_layer_ids;
+
+    // recovery information used to restore the KV cells to their original state in case of a failure
+    struct {
+        void clear() {
+            cells.clear();
+        }
+
+        std::unordered_map<uint32_t, kv_cell> cells;
+    } recovery;
+
     // defrag
     struct {
         std::vector<uint32_t> ids;
@@ -213,17 +256,6 @@ private:
     // 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;
-
     // find how many cells are currently in use
     uint32_t cell_max() const;
 
@@ -232,6 +264,8 @@ private:
     size_t size_k_bytes() const;
     size_t size_v_bytes() const;
 
+    bool is_masked_swa(llama_pos p0, llama_pos p1) const;
+
     lm_ggml_tensor * build_rope_shift(
             const llama_cparams & cparams,
                    lm_ggml_context * ctx,
@@ -258,6 +292,100 @@ private:
     bool state_read_data(llama_io_read_i & io, uint32_t cell_count);
 };
 
+//
+// llama_kv_cache_unified_iswa
+//
+
+// utilizes two instances of llama_kv_cache_unified
+//   the first instance is for the non-SWA layers of the model and the second instance is for the SWA layers
+//   upon successful commit, the SWA cache removes old tokens outside the n_swa window
+
+class llama_kv_cache_unified_iswa : public llama_kv_cache {
+public:
+    llama_kv_cache_unified_iswa(
+            const llama_model & model,
+                    lm_ggml_type   type_k,
+                    lm_ggml_type   type_v,
+                         bool   v_trans,
+                         bool   offload,
+                         bool   swa_full,
+                     uint32_t   kv_size,
+                     uint32_t   n_seq_max,
+                     uint32_t   n_batch,
+                     uint32_t   n_pad);
+
+    ~llama_kv_cache_unified_iswa() = default;
+
+    //
+    // llama_memory_i
+    //
+
+    void clear() 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_min(llama_seq_id seq_id) const override;
+    llama_pos seq_pos_max(llama_seq_id seq_id) const override;
+
+    //
+    // llama_kv_cache
+    //
+
+    void restore() override;
+    void commit()  override;
+
+    bool update(llama_context & ctx) override;
+
+    void defrag_sched(float thold) override;
+
+    void set_full() override;
+
+    llama_sbatch sbatch_init(const llama_batch & batch, bool logits_all) override;
+    llama_ubatch ubatch_next(llama_sbatch & sbatch, uint32_t n_ubatch, bool embd_pooled) const override;
+
+    bool find_slot(const llama_ubatch & batch) override;
+
+    bool get_can_shift() const override;
+
+    // state write/load
+
+    void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1) const override;
+    void state_read (llama_io_read_i  & io, llama_seq_id seq_id = -1)       override;
+
+    //
+    // llama_kv_cache_unified_iswa specific API
+    //
+
+    llama_kv_cache_unified * get_kv_base() const;
+    llama_kv_cache_unified * get_kv_swa () const;
+
+private:
+    const llama_hparams & hparams;
+
+    bool do_prune = true;
+
+    struct {
+        struct entry {
+            llama_pos pmin;
+            llama_pos pmax;
+        };
+
+        void clear() {
+            pos.clear();
+        }
+
+        // used to perform SWA pruning of old tokens
+        std::unordered_map<llama_seq_id, entry> pos;
+    } pending;
+
+    std::unique_ptr<llama_kv_cache_unified> kv_base;
+    std::unique_ptr<llama_kv_cache_unified> kv_swa;
+};
+
 //
 // llama_kv_cache_recurrent
 //
@@ -289,7 +417,8 @@ public:
                     lm_ggml_type   type_k,
                     lm_ggml_type   type_v,
                          bool   offload,
-                     uint32_t   kv_size);
+                     uint32_t   kv_size,
+                     uint32_t   n_seq_max);
 
     ~llama_kv_cache_recurrent() = default;
 
@@ -301,10 +430,11 @@ public:
 
     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_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_min(llama_seq_id seq_id) const override;
     llama_pos seq_pos_max(llama_seq_id seq_id) const override;
 
     //
@@ -314,24 +444,17 @@ public:
     void restore() override;
     void commit()  override;
 
-    bool update(llama_context & lctx) override;
+    bool update(llama_context & ctx) override;
 
     void defrag_sched(float thold) override;
 
     void set_full() override;
 
     llama_sbatch sbatch_init(const llama_batch & batch, bool logits_all) override;
-
     llama_ubatch ubatch_next(llama_sbatch & sbatch, uint32_t n_ubatch, bool embd_pooled) const override;
 
     bool find_slot(const llama_ubatch & batch) override;
 
-    int32_t get_n_tokens()   const override;
-    int32_t get_used_cells() const override;
-
-    // TODO: better data structures to reduce the cost of this operation
-    llama_pos get_pos_max() const override;
-
     bool get_can_shift() const override;
 
     // TODO: temporary methods - they are not really const as they do const_cast<>, fix this
@@ -343,11 +466,8 @@ public:
     void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1) const override;
     void state_read (llama_io_read_i  & io, llama_seq_id seq_id = -1) override;
 
-    // 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 head = 0; // the location where the batch will be placed in the cache (see find_slot())
+    uint32_t size = 0; // total number of cells, shared across all sequences
     uint32_t used = 0; // used cells (i.e. at least one seq_id)
 
     // computed before each graph build
@@ -374,8 +494,7 @@ private:
         std::vector<slot_range> ranges;
     } pending;
 
-    lm_ggml_type type_k = LM_GGML_TYPE_F16;
-    lm_ggml_type type_v = LM_GGML_TYPE_F16;
+    const uint32_t n_seq_max = 1;
 
     std::vector<lm_ggml_context_ptr>        ctxs;
     std::vector<lm_ggml_backend_buffer_ptr> bufs;
@@ -394,12 +513,3 @@ private:
     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);
 };
-
-
-//
-// 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/rnllama.framework/Headers/llama-memory.h

@@ -7,8 +7,8 @@ struct llama_memory_params {
     lm_ggml_type type_k;
     lm_ggml_type type_v;
 
-    // parameters for other types of memory
-    // ...
+    // use full-size SWA cache
+    bool swa_full;
 };
 
 // general concept of LLM memory
@@ -25,6 +25,7 @@ public:
     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_min(llama_seq_id seq_id) const = 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/rnllama.framework/Headers/llama-model-saver.h

@@ -0,0 +1,37 @@
+#pragma once
+
+#include "llama.h"
+#include "llama-arch.h"
+
+#include <vector>
+
+struct llama_model_saver {
+    struct lm_gguf_context * lm_gguf_ctx = nullptr;
+    const struct llama_model & model;
+    const struct LLM_KV llm_kv;
+
+    llama_model_saver(const struct llama_model & model);
+    ~llama_model_saver();
+
+    void add_kv(enum llm_kv key, uint32_t     value);
+    void add_kv(enum llm_kv key, int32_t      value);
+    void add_kv(enum llm_kv key, float        value);
+    void add_kv(enum llm_kv key, bool         value);
+    void add_kv(enum llm_kv key, const char * value);
+
+    [[noreturn]]
+    void add_kv(enum llm_kv key, char value); // needed to make the template below compile
+
+    template <typename Container>
+    void add_kv(enum llm_kv key, const Container & value, bool per_layer = false);
+
+    void add_kv(enum llm_kv key, const std::vector<std::string> & value);
+
+    void add_tensor(const struct lm_ggml_tensor * tensor);
+
+    void add_kv_from_model();
+
+    void add_tensors_from_model();
+
+    void save(const std::string & path_model);
+};

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

@@ -76,6 +76,7 @@ enum llm_type {
     LLM_TYPE_236B,
     LLM_TYPE_290B,
     LLM_TYPE_314B,
+    LLM_TYPE_405B,
     LLM_TYPE_671B,
     LLM_TYPE_SMALL,
     LLM_TYPE_MEDIUM,
@@ -95,6 +96,8 @@ enum llm_type {
     LLM_TYPE_235B_A22B,
 };
 
+std::string llama_rope_scaling_type_name(llama_rope_scaling_type rope_scaling_type);
+
 struct llama_layer_posnet {
     // resnet
     struct lm_ggml_tensor * norm1   = nullptr;
@@ -395,7 +398,10 @@ struct llama_model {
 
     const struct lm_ggml_tensor * get_tensor(const char * name) const;
 
-    lm_ggml_tensor * get_rope_factors(uint32_t n_ctx_per_seq, int il) const;
+    float get_rope_freq_base (const llama_cparams & cparams, int il) const;
+    float get_rope_freq_scale(const llama_cparams & cparams, int il) const;
+
+    lm_ggml_tensor * get_rope_factors(const llama_cparams & cparams, int il) const;
 
     // note: can mutate `cparams`
     // TODO: move this to new llm_arch_model_i interface

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

@@ -21,6 +21,9 @@ struct llama_vocab {
 
     void load(llama_model_loader & ml, const LLM_KV & kv);
 
+    std::string get_tokenizer_model() const;
+    std::string get_tokenizer_pre() const;
+
     enum llama_vocab_type     get_type()     const;
     enum llama_vocab_pre_type get_pre_type() const;
 
@@ -80,6 +83,9 @@ struct llama_vocab {
     int max_token_len() const;
 
     int find_bpe_rank(const std::string & token_left, const std::string & token_right) const;
+    std::vector<std::string> get_bpe_merges() const;
+
+    std::vector<char> get_precompiled_charsmap() const;
 
     int32_t tokenize(
                    const char * text,