@fugood/llama.node 0.3.16 → 0.4.0

package/src/llama.cpp/src/llama-kv-cache.h

@@ -2,96 +2,131 @@
 
 #include "llama.h"
 #include "llama-io.h"
+#include "llama-graph.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_sbatch;
+struct llama_model;
+struct llama_context;
 
 struct llama_kv_cache : public llama_memory_i {
-    using llama_memory_i::llama_memory_i;
+    virtual ~llama_kv_cache() = default;
 
-    virtual int32_t  get_n_tokens()   const = 0;
-    virtual uint32_t get_used_cells() const = 0; // TODO: remove, this is too-specific to the unified cache
+    // call if batch processing fails - restores the cache state
+    virtual void restore() = 0;
 
-    virtual bool get_can_shift() const = 0;
+    // call after successful batch processing - clears any pending state
+    virtual void commit()  = 0;
+
+    // process any pending defrag/shift/etc. operations
+    // optionally call once before processing a new batch
+    virtual bool update(llama_context & lctx) = 0;
+
+    // schedule a defrag if the fragmentation threshold is exceeded. otherwise, do nothing
+    virtual void defrag_sched(float thold) = 0;
+
+    // simulate full cache, used for allocating worst-case compute buffers
+    virtual void set_full() = 0;
+
+    //
+    // batch processing
+    //
+
+    virtual llama_sbatch sbatch_init(const llama_batch & batch, bool logits_all) = 0;
+
+    // different KV caches require different batch splitting strategies
+    virtual llama_ubatch ubatch_next(llama_sbatch & sbatch, uint32_t n_ubatch, bool embd_pooled) const = 0;
+
+    // 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;
 
     bool get_can_edit() const override { return get_can_shift(); }
+
+    //
+    // state write/read
+    //
+
+    virtual void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1) const = 0;
+    virtual void state_read (llama_io_read_i  & io, llama_seq_id seq_id = -1) = 0;
 };
 
-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;
+//
+// llama_kv_cache_guard
+//
 
-    std::set<llama_seq_id> seq_id;
+struct llama_kv_cache_guard {
+    llama_kv_cache_guard(llama_kv_cache * kv) : kv(kv) {}
 
-    bool has_seq_id(const llama_seq_id & id) const {
-        return seq_id.find(id) != seq_id.end();
+    ~llama_kv_cache_guard() {
+        kv->restore();
     }
 
-    bool is_empty() const {
-        return seq_id.empty();
+    void commit() {
+        kv->commit();
     }
 
-    bool is_same_seq(const llama_kv_cell & other) const {
-        return seq_id == other.seq_id;
-    }
+private:
+    llama_kv_cache * kv;
 };
 
-// a structure holds information about the slot found in llama_kv_cache_find_slot
-struct llama_kv_cache_slot_info {
-    std::pair<uint32_t, uint32_t> boundaries; // slot boundaries [begin, end)
-    bool found = false;                       // the slot was found
-
-    explicit llama_kv_cache_slot_info(bool found_) : found{found_} {}
-    llama_kv_cache_slot_info(uint32_t begin, uint32_t end) : boundaries{begin, end}, found{true} {}
-
-    operator bool() const { return found; }
-};
+//
+// llama_kv_cache_unified
+//
 
-// 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<ggml_tensor * (uint32_t n_ctx_per_seq, int il)> get_rope_factors;
-    };
+    struct kv_cell {
+        llama_pos pos   = -1;
+        llama_pos delta =  0;
 
-    llama_kv_cache_unified(
-            const llama_hparams & hparams,
-            callbacks             cbs);
+        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();
+        }
 
-    virtual ~llama_kv_cache_unified() = default;
+        bool is_same_seq(const kv_cell & other) const {
+            return seq_id == other.seq_id;
+        }
+    };
 
-    // TODO: become constructor
-    bool init(
-            const llama_model & model,   // TODO: do not reference the model
-          const llama_cparams & cparams,
+    static uint32_t get_padding(const llama_cparams & cparams);
+
+    llama_kv_cache_unified(
+            const llama_model & model,
                     ggml_type   type_k,
                     ggml_type   type_v,
+                         bool   v_trans,
+                         bool   offload,
                      uint32_t   kv_size,
-                         bool   offload);
+                     uint32_t   padding);
 
-    int32_t  get_n_tokens()   const override;
-    uint32_t get_used_cells() const override;
+    ~llama_kv_cache_unified() = default;
 
-    size_t total_size() const;
-
-    // TODO: better data structures to reduce the cost of this operation
-    llama_pos pos_max() const;
+    //
+    // llama_memory_i
+    //
 
     void clear() override;
-    void defrag() 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;
@@ -99,77 +134,120 @@ public:
     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) override;
+    llama_pos seq_pos_max(llama_seq_id seq_id) const override;
 
-    bool get_can_shift() const override;
+    //
+    // llama_kv_cache
+    //
 
-    // find an empty slot of size "n_tokens" in the cache
-    // updates the cache head
-    // returns a structure holding information about the slot found
-    // Note: On success, it's important that cache.head points
-    // to the first cell of the slot.
-    llama_kv_cache_slot_info find_slot(const llama_ubatch & batch);
+    void restore() override;
+    void commit()  override;
 
-    // TODO: maybe not needed
-    uint32_t get_padding(const llama_cparams & cparams) const;
+    bool update(llama_context & ctx) override;
 
-    // find how many cells are currently in use
-    uint32_t cell_max() const;
+    void defrag_sched(float thold) override;
 
-    size_t size_k_bytes() const;
-    size_t size_v_bytes() const;
-
-    // defrag
-
-    struct {
-        std::vector<uint32_t> ids;
-    } defrag_info;
+    void set_full() override;
 
-    // return true if cells have been moved
-    bool defrag_prepare(int32_t n_max_nodes);
+    llama_sbatch sbatch_init(const llama_batch & batch, bool logits_all) override;
 
-    // state save/load
+    llama_ubatch ubatch_next(llama_sbatch & sbatch, uint32_t n_ubatch, bool embd_pooled) const override;
 
-    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);
+    // 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) override;
 
-    // members
+    int32_t get_n_tokens()   const override;
+    int32_t get_used_cells() const override;
 
-    const llama_hparams & hparams;
+    // TODO: better data structures to reduce the cost of this operation
+    llama_pos get_pos_max() const override;
 
-    callbacks cbs;
+    bool get_can_shift() const override;
 
-    bool has_shift = false;
-    bool do_defrag = false;
+    // state write/load
 
-    // 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
+    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;
 
-    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 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
     uint32_t n = 0;
 
-    std::vector<llama_kv_cell> cells;
+    std::vector<kv_cell> cells;
 
     std::vector<ggml_tensor *> k_l; // per layer
     std::vector<ggml_tensor *> v_l;
 
 private:
+    const llama_model & model;
+    const llama_hparams & hparams;
+
+    bool has_shift = false;
+    bool do_defrag = false;
+
+    bool v_trans   = true;  // the value tensor is transposed
+    bool can_shift = false;
+
+    // required padding
+    uint32_t padding = 1;
+
     ggml_type type_k = GGML_TYPE_F16;
     ggml_type type_v = GGML_TYPE_F16;
 
     std::vector<ggml_context_ptr>        ctxs;
     std::vector<ggml_backend_buffer_ptr> bufs;
 
+    // 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;
+
+    // find how many cells are currently in use
+    uint32_t cell_max() const;
+
+    size_t total_size() const;
+
+    size_t size_k_bytes() const;
+    size_t size_v_bytes() const;
+
+    ggml_tensor * build_rope_shift(
+            const llama_cparams & cparams,
+                   ggml_context * ctx,
+                    ggml_tensor * cur,
+                    ggml_tensor * shift,
+                    ggml_tensor * factors,
+                          float   freq_base,
+                          float   freq_scale) const;
+
+    llm_graph_result_ptr build_graph_shift(
+            const llama_cparams & cparams,
+                   ggml_context * ctx,
+                    ggml_cgraph * gf) const;
+
+    llm_graph_result_ptr build_graph_defrag(
+            const llama_cparams & cparams,
+                   ggml_context * ctx,
+                    ggml_cgraph * gf) const;
+
     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;
 
@@ -177,106 +255,140 @@ private:
     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 restore
+// llama_kv_cache_recurrent
 //
 
-// saves the kv_cache state for future recovery.
-// used to rollback llama_kv_cache_find_slot changes.
-struct llama_kv_slot_restorer {
-    struct llama_kv_cache_state {
-        uint32_t head = 0;
-        uint32_t n    = 0;
-    } old_state;
-
-    // for non-recurrent models only
-    // list of slots to restore
-    std::vector<std::pair<uint32_t, uint32_t>> slot_boundaries;
-
-    bool do_restore = false;
+class llama_kv_cache_recurrent : public llama_kv_cache {
+public:
+    struct kv_cell {
+        llama_pos pos  = -1;
+        int32_t   src  = -1; // used to copy states
+        int32_t   tail = -1;
 
-    llama_kv_cache_unified & cache;
+        std::set<llama_seq_id> seq_id;
 
-    explicit llama_kv_slot_restorer(llama_kv_cache_unified & cache) : cache(cache) {
-        old_state.head = cache.head;
-        old_state.n    = cache.n;
-    }
+        bool has_seq_id(const llama_seq_id & id) const {
+            return seq_id.find(id) != seq_id.end();
+        }
 
-    // saves a slot information for future restoration
-    void save(const llama_kv_cache_slot_info & slot) {
-        if (slot) {
-            do_restore = true;
-            if (slot.boundaries.first != slot.boundaries.second) {
-                slot_boundaries.push_back(slot.boundaries);
-            }
+        bool is_empty() const {
+            return seq_id.empty();
         }
-    }
 
-    // must be explicitly called to restore the kv_cache state
-    // and rollback changes from all llama_kv_cache_find_slot calls
-    void restore() {
-        if (do_restore) {
-            cache.head = old_state.head;
-            cache.n    = old_state.n;
-
-            if (cache.recurrent) { // recurrent models like Mamba or RWKV can't have a state partially erased
-                cache.seq_rm(-1, -1, -1);
-            } else {
-                for (auto & slot : slot_boundaries) {
-                    cache.seq_rm(-1, slot.first, slot.second);
-                }
-            }
+        bool is_same_seq(const kv_cell & other) const {
+            return seq_id == other.seq_id;
         }
-    }
-};
+    };
+
+    llama_kv_cache_recurrent(
+            const llama_model & model,
+                    ggml_type   type_k,
+                    ggml_type   type_v,
+                         bool   offload,
+                     uint32_t   kv_size);
+
+    ~llama_kv_cache_recurrent() = 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;
 
-// TODO: maybe become part of the public llama_kv_cache in the future
-int32_t llama_kv_cache_n_tokens(const llama_kv_cache * kv);
+    llama_pos seq_pos_max(llama_seq_id seq_id) const override;
 
-int32_t llama_kv_cache_used_cells(const llama_kv_cache * kv);
+    //
+    // llama_kv_cache
+    //
 
-void llama_kv_cache_clear(llama_kv_cache * kv);
+    void restore() override;
+    void commit()  override;
 
-bool llama_kv_cache_seq_rm(
-        llama_kv_cache * kv,
-          llama_seq_id   seq_id,
-             llama_pos   p0,
-             llama_pos   p1);
+    bool update(llama_context & lctx) override;
 
-void llama_kv_cache_seq_cp(
-        llama_kv_cache * kv,
-          llama_seq_id   seq_id_src,
-          llama_seq_id   seq_id_dst,
-             llama_pos   p0,
-             llama_pos   p1);
+    void defrag_sched(float thold) override;
 
-void llama_kv_cache_seq_keep(llama_kv_cache * kv, llama_seq_id seq_id);
+    void set_full() override;
 
-void llama_kv_cache_seq_add(
-        llama_kv_cache * kv,
-          llama_seq_id   seq_id,
-             llama_pos   p0,
-             llama_pos   p1,
-             llama_pos   delta);
+    llama_sbatch sbatch_init(const llama_batch & batch, bool logits_all) override;
 
-void llama_kv_cache_seq_div(
-        llama_kv_cache * kv,
-          llama_seq_id   seq_id,
-             llama_pos   p0,
-             llama_pos   p1,
-                   int   d);
+    llama_ubatch ubatch_next(llama_sbatch & sbatch, uint32_t n_ubatch, bool embd_pooled) const override;
 
-llama_pos llama_kv_cache_seq_pos_max(llama_kv_cache * kv, llama_seq_id seq_id);
+    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;
 
-void llama_kv_cache_defrag(llama_kv_cache * kv);
+    // TODO: temporary methods - they are not really const as they do const_cast<>, fix this
+    int32_t s_copy(int i) const;
+    float   s_mask(int i) const;
+
+    // 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;
+
+    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
+    uint32_t n = 0;
+
+    std::vector<kv_cell> cells;
+
+    std::vector<ggml_tensor *> k_l; // per layer
+    std::vector<ggml_tensor *> v_l;
+
+private:
+    //const llama_model & model;
+    const llama_hparams & hparams;
+
+    // commit/restore cache
+    // TODO: rework for recurrent 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;
+
+    ggml_type type_k = GGML_TYPE_F16;
+    ggml_type type_v = GGML_TYPE_F16;
+
+    std::vector<ggml_context_ptr>        ctxs;
+    std::vector<ggml_backend_buffer_ptr> bufs;
+
+    // find how many cells are currently in use
+    uint32_t cell_max() const;
+
+    size_t total_size() const;
+
+    size_t size_k_bytes() const;
+    size_t size_v_bytes() const;
+
+    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);
+};
 
-bool llama_kv_cache_can_shift(const llama_kv_cache * kv);
 
 //
 // kv cache view

package/src/llama.cpp/src/llama-memory.h

@@ -2,12 +2,22 @@
 
 #include "llama.h"
 
+struct llama_memory_params {
+    // kv cache
+    ggml_type type_k;
+    ggml_type type_v;
+
+    // parameters for other types of memory
+    // ...
+};
+
 // 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 ~llama_memory_i() = default;
+
     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;
@@ -15,7 +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_max(llama_seq_id seq_id) = 0;
+    virtual llama_pos seq_pos_max(llama_seq_id seq_id) const = 0;
 
     virtual bool get_can_edit() const = 0;
 };

package/src/llama.cpp/src/llama-mmap.cpp

@@ -476,7 +476,7 @@ struct llama_mlock::impl {
 
         char* errmsg = std::strerror(errno);
         bool suggest = (errno == ENOMEM);
-#if defined(TARGET_OS_VISION) || defined(TARGET_OS_TV)
+#if defined(TARGET_OS_VISION) || defined(TARGET_OS_TV) || defined(_AIX)
         // visionOS/tvOS dont't support RLIMIT_MEMLOCK
         // Skip resource limit checks on visionOS/tvOS
         suggest = false;