@novastera-oss/llamarn 0.2.5 → 0.2.7

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

@@ -0,0 +1,308 @@
+#pragma once
+
+#include "llama-batch.h"
+#include "llama-graph.h"
+#include "llama-kv-cells.h"
+#include "llama-memory.h"
+
+#include <unordered_map>
+#include <vector>
+
+struct llama_cparams;
+struct llama_hparams;
+struct llama_model;
+struct llama_context;
+
+//
+// llama_kv_cache_unified
+//
+
+class llama_kv_cache_unified : public llama_memory_i {
+public:
+    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)>;
+
+    using ubatch_heads = std::vector<uint32_t>;
+
+    struct defrag_info {
+        bool empty() const {
+            return ids.empty();
+        }
+
+        // contains information about which cell moves where:
+        //  - cell i moves to ids[i]
+        //  - if ids[i] == i || ids[i] == ids.size(), then cell i is not moved
+        std::vector<uint32_t> ids;
+    };
+
+    llama_kv_cache_unified(
+            const llama_model &  model,
+              layer_filter_cb && filter,
+                    ggml_type    type_k,
+                    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;
+
+    //
+    // llama_memory_i
+    //
+
+    llama_memory_state_ptr init_batch(
+            const llama_batch & batch,
+            uint32_t n_ubatch,
+            bool embd_all) override;
+
+    llama_memory_state_ptr init_full() override;
+
+    llama_memory_state_ptr init_update(llama_context * lctx, bool optimize) override;
+
+    bool get_can_shift() const override;
+
+    void clear(bool data) 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 shift) 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;
+
+    // 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 specific API
+    //
+
+    uint32_t get_size() const;
+
+    bool get_has_shift() const;
+
+    //
+    // graph_build API
+    //
+
+    uint32_t get_n_kv() const;
+
+    // get views of the current state of the cache
+    ggml_tensor * get_k(ggml_context * ctx, int32_t il, uint32_t n_kv) const;
+    ggml_tensor * get_v(ggml_context * ctx, int32_t il, uint32_t n_kv) const;
+
+    // store k_cur and v_cur in the cache based on the provided head location
+    ggml_tensor * cpy_k(ggml_context * ctx, ggml_tensor * k_cur, int32_t il, uint32_t head_cur) const;
+    ggml_tensor * cpy_v(ggml_context * ctx, ggml_tensor * v_cur, int32_t il, uint32_t head_cur) const;
+
+    //
+    // preparation API
+    //
+
+    // find places for the provided ubatches in the cache, returns the head locations
+    // return empty vector on failure
+    ubatch_heads prepare(const std::vector<llama_ubatch> & ubatches);
+
+    bool update(llama_context * lctx, bool do_shift, const defrag_info & dinfo);
+
+    // return the cell position where we can insert the ubatch
+    // return -1 on failure to find a contiguous slot of kv cells
+    int32_t find_slot(const llama_ubatch & ubatch) const;
+
+    // emplace the ubatch context into slot: [head_cur, head_cur + ubatch.n_tokens)
+    void apply_ubatch(uint32_t head_cur, const llama_ubatch & ubatch);
+
+    //
+    // set_input API
+    //
+
+    void set_input_kq_mask   (ggml_tensor * dst, const llama_ubatch * ubatch, bool causal_attn) const;
+    void set_input_k_shift   (ggml_tensor * dst) const;
+    void set_input_pos_bucket(ggml_tensor * dst, const llama_ubatch * ubatch) const;
+
+private:
+    const llama_model & model;
+    const llama_hparams & hparams;
+
+    struct kv_layer {
+        // layer index in the model
+        // note: can be different from the layer index in the KV cache
+        uint32_t il;
+
+        ggml_tensor * k;
+        ggml_tensor * v;
+    };
+
+    bool v_trans = true;  // the value tensor is transposed
+
+    // the current index from where we start searching for a free slot in the ring buffer of KV cells (see find_slot())
+    // note: this is not part of the KV state and it's only used to speed-up the find_slot() method
+    uint32_t head = 0;
+
+    const uint32_t n_seq_max = 1;
+
+    // required padding
+    const uint32_t n_pad = 1;
+
+    // SWA
+    const uint32_t n_swa = 0;
+
+    int debug = 0;
+
+    const llama_swa_type swa_type = LLAMA_SWA_TYPE_NONE;
+
+    std::vector<ggml_context_ptr>        ctxs;
+    std::vector<ggml_backend_buffer_ptr> bufs;
+
+    llama_kv_cells_unified cells;
+
+    std::vector<kv_layer> layers;
+
+    // model layer id -> KV cache layer id
+    std::unordered_map<int32_t, int32_t> map_layer_ids;
+
+    // return non-empty vector if cells have been moved
+    defrag_info defrag_prepare(int32_t n_max_nodes) const;
+
+    size_t total_size() const;
+
+    size_t size_k_bytes() const;
+    size_t size_v_bytes() const;
+
+    bool is_masked_swa(llama_pos p0, llama_pos p1) 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 defrag_info & dinfo) 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);
+};
+
+class llama_kv_cache_unified_state : public llama_memory_state_i {
+public:
+    // some shorthands
+    using ubatch_heads = llama_kv_cache_unified::ubatch_heads;
+    using defrag_info  = llama_kv_cache_unified::defrag_info;
+
+    // used for errors
+    llama_kv_cache_unified_state(llama_memory_status status);
+
+    // used to create a full-cache state
+    llama_kv_cache_unified_state(
+            llama_kv_cache_unified * kv);
+
+    // used to create an update state
+    llama_kv_cache_unified_state(
+            llama_kv_cache_unified * kv,
+            llama_context * lctx,
+            bool do_shift,
+            defrag_info dinfo);
+
+    // used to create a decode state from a batch
+    llama_kv_cache_unified_state(
+            llama_kv_cache_unified * kv,
+            llama_sbatch sbatch,
+            ubatch_heads heads,
+            std::vector<llama_ubatch> ubatches);
+
+    virtual ~llama_kv_cache_unified_state();
+
+    //
+    // llama_memory_state_i
+    //
+
+    bool next()  override;
+    bool apply() override;
+
+    std::vector<int64_t> & out_ids() override;
+
+    llama_memory_status  get_status() const override;
+    const llama_ubatch & get_ubatch() const override;
+
+    //
+    // llama_kv_cache_unified_state specific API
+    //
+
+    uint32_t get_n_kv() const;
+
+    // get views of the current state of the cache
+    ggml_tensor * get_k(ggml_context * ctx, int32_t il) const;
+    ggml_tensor * get_v(ggml_context * ctx, int32_t il) const;
+
+    // store k_cur and v_cur in the cache based on the provided head location
+    ggml_tensor * cpy_k(ggml_context * ctx, ggml_tensor * k_cur, int32_t il) const;
+    ggml_tensor * cpy_v(ggml_context * ctx, ggml_tensor * v_cur, int32_t il) const;
+
+    void set_input_k_shift(ggml_tensor * dst) const;
+
+    void set_input_kq_mask   (ggml_tensor * dst, const llama_ubatch * ubatch, bool causal_attn) const;
+    void set_input_pos_bucket(ggml_tensor * dst, const llama_ubatch * ubatch) const;
+
+private:
+    llama_memory_status status;
+
+    llama_kv_cache_unified * kv;
+    llama_context * lctx;
+
+    //
+    // update state
+    //
+
+    bool do_shift = false;
+
+    defrag_info dinfo;
+
+    //
+    // batch processing state
+    //
+
+    llama_sbatch sbatch;
+
+    // the index of the next ubatch to process
+    size_t i_next = 0;
+
+    ubatch_heads heads;
+
+    std::vector<llama_ubatch> ubatches;
+
+    //
+    // data needed for building the compute graph for the current ubatch:
+    //
+
+    // a heuristic, to avoid attending the full cache if it is not yet utilized
+    // as the cache gets filled, the benefit from this heuristic disappears
+    int32_t n_kv;
+
+    // the beginning of the current slot in which the ubatch will be inserted
+    int32_t head;
+};

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

@@ -23,7 +23,7 @@ public:
 
         used.clear();
 
-        for (uint32_t s = 0; s < LLAMA_MAX_PARALLEL_SEQUENCES; ++s) {
+        for (uint32_t s = 0; s < LLAMA_MAX_SEQ; ++s) {
             seq_pos[s].clear();
         }
     }
@@ -68,12 +68,6 @@ public:
     // the index of the last cell that is used + 1
     // return 0 if no cells are used
     uint32_t used_max_p1() const {
-#if 0
-        if (!seq_pos[0].empty()) printf("kv_cells: min[0] = %5d, max[0] = %5d\n", *seq_pos[0].begin(), *seq_pos[0].rbegin());
-        if (!seq_pos[1].empty()) printf("kv_cells: min[1] = %5d, max[1] = %5d\n", *seq_pos[1].begin(), *seq_pos[1].rbegin());
-        if (!seq_pos[2].empty()) printf("kv_cells: min[2] = %5d, max[2] = %5d\n", *seq_pos[2].begin(), *seq_pos[2].rbegin());
-#endif
-
         return used.empty() ? 0 : *used.rbegin() + 1;
     }
 
@@ -86,6 +80,9 @@ public:
         assert(isrc < pos.size());
         assert(idst < pos.size());
 
+        assert(pos[idst] == -1);
+        assert(pos[isrc] != -1);
+
         pos  [idst] = pos  [isrc];
         shift[idst] = shift[isrc];
         seq  [idst] = seq  [isrc];
@@ -144,6 +141,20 @@ public:
         }
     }
 
+    // clear a non-empty cell
+    void rm(uint32_t i) {
+        assert(i < pos.size());
+        assert(pos[i] != -1);
+
+        seq_pos_rm(i);
+        seq[i].reset();
+
+        pos[i] = -1;
+        shift[i] = 0;
+
+        used.erase(i);
+    }
+
     // note: call only if the cell has seq_id
     // return true if the cell becomes empty
     bool seq_rm(uint32_t i, llama_seq_id seq_id) {
@@ -157,6 +168,7 @@ public:
 
         if (seq[i].none()) {
             pos[i] = -1;
+            shift[i] = 0;
 
             used.erase(i);
 
@@ -185,6 +197,7 @@ public:
             seq[i].reset();
 
             pos[i] = -1;
+            shift[i] = 0;
 
             used.erase(i);
 
@@ -196,6 +209,15 @@ public:
         return false;
     }
 
+    // number of different sequences in the cell
+    int seq_count(uint32_t i) const {
+        assert(i < pos.size());
+        assert(pos[i] != -1);
+
+        return seq[i].count();
+    }
+
+    // check if the cell contains seq_id
     bool seq_has(uint32_t i, llama_seq_id seq_id) const {
         assert(i < pos.size());
         assert(seq_id >= 0);
@@ -213,11 +235,25 @@ public:
         seq_pos[seq_id].insert(pos[i]);
     }
 
+    // return the sequence id of this cell
+    // note: call only for cells with exactly one sequence
+    llama_seq_id seq_get(uint32_t i) const {
+        assert(seq[i].count() == 1);
+
+        for (int s = 0; s < LLAMA_MAX_SEQ; ++s) {
+            if (seq[i].test(s)) {
+                return s;
+            }
+        }
+
+        return -1;
+    }
+
     // the minimum position of sequence seq_id currently present in any of the cells
     // return -1 if the sequence is not present
     llama_pos seq_pos_min(llama_seq_id seq_id) const {
         assert(seq_id >= 0);
-        assert(seq_id < LLAMA_MAX_PARALLEL_SEQUENCES);
+        assert(seq_id < LLAMA_MAX_SEQ);
 
         if (seq_pos[seq_id].empty()) {
             return -1;
@@ -230,7 +266,7 @@ public:
     // return -1 if the sequence is not present
     llama_pos seq_pos_max(llama_seq_id seq_id) const {
         assert(seq_id >= 0);
-        assert(seq_id < LLAMA_MAX_PARALLEL_SEQUENCES);
+        assert(seq_id < LLAMA_MAX_SEQ);
 
         if (seq_pos[seq_id].empty()) {
             return -1;
@@ -268,6 +304,7 @@ public:
     void pos_set(uint32_t i, llama_pos p) {
         assert(i < pos.size());
         assert(pos[i] == -1);
+        assert(seq[i].none());
 
         pos[i] = p;
 
@@ -286,21 +323,20 @@ public:
         pos[i]   += d;
         shift[i] += d;
 
-        seq_pos_add(i);
-
         has_shift = true;
 
         if (pos[i] < 0) {
-            seq_pos_rm(i);
-
             seq[i].reset();
             pos[i] = -1;
+            shift[i] = 0;
 
             used.erase(i);
 
             return true;
         }
 
+        seq_pos_add(i);
+
         return false;
     }
 
@@ -348,20 +384,20 @@ private:
     //
     std::vector<llama_pos> shift;
 
-    using bits_t = std::bitset<LLAMA_MAX_PARALLEL_SEQUENCES>;
+    using bits_t = std::bitset<LLAMA_MAX_SEQ>;
 
     // the bitset seq[i] tells us which sequences are currently occupying the i-th cell
     std::vector<bits_t> seq;
 
     // the set seq_pos[s] tells us which positions are currently present for sequence s
     // this way seq_pos[s].begin() and seq_pos[s].rbegin() give us the min/max positions currently in the cache
-    std::set<llama_pos> seq_pos[LLAMA_MAX_PARALLEL_SEQUENCES];
+    std::set<llama_pos> seq_pos[LLAMA_MAX_SEQ];
 
     // helper functions for updating `seq_pos`, once cell at a time:
 
     // remove cell i
     void seq_pos_rm(uint32_t i) {
-        for (int s = 0; s < LLAMA_MAX_PARALLEL_SEQUENCES; ++s) {
+        for (int s = 0; s < LLAMA_MAX_SEQ; ++s) {
             if (seq[i].test(s)) {
                 seq_pos[s].erase(pos[i]);
             }
@@ -370,7 +406,7 @@ private:
 
     // add cell i
     void seq_pos_add(uint32_t i) {
-        for (int s = 0; s < LLAMA_MAX_PARALLEL_SEQUENCES; ++s) {
+        for (int s = 0; s < LLAMA_MAX_SEQ; ++s) {
             if (seq[i].test(s)) {
                 seq_pos[s].insert(pos[i]);
             }