whispercpp 1.3.3 → 1.3.5

data/ext/sources/examples/talk-llama/llama-kv-cache.h

@@ -1,44 +1,390 @@
 #pragma once
 
-#include "llama.h"
-#include "llama-io.h"
+#include "llama-batch.h"
+#include "llama-graph.h"
+#include "llama-kv-cells.h"
 #include "llama-memory.h"
 
-struct llama_kv_cache : public llama_memory_i {
-    virtual ~llama_kv_cache() = default;
+#include <unordered_map>
+#include <vector>
 
-    // split the input batch into a set of ubatches and verify that they can fit into the cache
-    // return a state object containing the ubatches and KV cache state required to process them
-    // check the llama_memory_state_i::get_status() for the result
-    virtual llama_memory_state_ptr init_batch(
-            const llama_batch & batch,
+struct llama_cparams;
+struct llama_hparams;
+struct llama_model;
+struct llama_context;
+
+//
+// llama_kv_cache
+//
+
+class llama_kv_cache : public llama_memory_i {
+public:
+    struct stream_copy_info {
+        bool empty() const {
+            assert(ssrc.size() == sdst.size());
+            return ssrc.empty();
+        }
+
+        std::vector<uint32_t> ssrc;
+        std::vector<uint32_t> sdst;
+    };
+
+    // for each ubatch, create a slot_info that contains information about where the ubatch should be inserted in the
+    //   KV cells. for example, cell indices for each token, such that: token[i] -> goes to cells[idxs[i]]
+    struct slot_info {
+        // data for ggml_set_rows
+        using idx_vec_t = std::vector<uint32_t>;
+
+        // number of streams: ns = s1 - s0 + 1
+        uint32_t s0;
+        uint32_t s1;
+
+        std::vector<llama_seq_id> strm; // [ns]
+        std::vector<idx_vec_t>    idxs; // [ns]
+
+        uint32_t head() const {
+            GGML_ASSERT(idxs.size() == 1);
+            GGML_ASSERT(!idxs[0].empty());
+
+            return idxs[0][0];
+        }
+
+        void resize(size_t n) {
+            strm.resize(n);
+            idxs.resize(n);
+        }
+
+        size_t size() const {
+            GGML_ASSERT(idxs.size() == strm.size());
+            GGML_ASSERT(!idxs.empty());
+
+            return idxs[0].size();
+        }
+
+        size_t n_stream() const {
+            return strm.size();
+        }
+
+        bool empty() const {
+            return idxs.empty();
+        }
+
+        void clear() {
+            idxs.clear();
+        }
+
+        // check if indices are contiguous starting from head()
+        bool is_contiguous() const {
+            if (idxs.empty() || idxs[0].empty()) {
+                return true;
+            }
+            if (idxs.size() > 1) {
+                return false;
+            }
+            const uint32_t h = idxs[0][0];
+            for (size_t i = 0; i < idxs[0].size(); ++i) {
+                if (idxs[0][i] != h + i) {
+                    return false;
+                }
+            }
+            return true;
+        }
+    };
+
+    using slot_info_vec_t = std::vector<slot_info>;
+
+    llama_kv_cache(
+            const llama_model & model,
+                    ggml_type   type_k,
+                    ggml_type   type_v,
+                         bool   v_trans,
+                         bool   offload,
+                         bool   unified,
+                     uint32_t   kv_size,
+                     uint32_t   n_seq_max,
+                     uint32_t   n_pad,
+                     uint32_t   n_swa,
+               llama_swa_type   swa_type,
+        const layer_filter_cb & filter,
+        const  layer_reuse_cb & reuse);
+
+    ~llama_kv_cache() = default;
+
+    //
+    // llama_memory_i
+    //
+
+    llama_memory_context_ptr init_batch(
+            llama_batch_allocr & balloc,
             uint32_t n_ubatch,
-            bool embd_pooled,
-            bool logits_all) = 0;
+            bool embd_all) override;
+
+    llama_memory_context_ptr init_full() override;
+
+    llama_memory_context_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;
+
+    std::map<ggml_backend_buffer_type_t, size_t> memory_breakdown() const override;
+
+    // state write/load
+
+    void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1, llama_state_seq_flags flags = 0) const override;
+    void state_read (llama_io_read_i  & io, llama_seq_id seq_id = -1, llama_state_seq_flags flags = 0) override;
+
+    //
+    // llama_kv_cache specific API
+    //
+
+    uint32_t get_size()     const;
+    uint32_t get_n_stream() const;
+
+    bool get_has_shift() const;
+
+    //
+    // graph_build API
+    //
+
+    uint32_t get_n_kv(const slot_info & sinfo) 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 slot_info & sinfo) const;
+    ggml_tensor * get_v(ggml_context * ctx, int32_t il, uint32_t n_kv, const slot_info & sinfo) 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, ggml_tensor * k_idxs, int32_t il, const slot_info & sinfo) const;
+    ggml_tensor * cpy_v(ggml_context * ctx, ggml_tensor * v_cur, ggml_tensor * v_idxs, int32_t il, const slot_info & sinfo) const;
+
+    //
+    // preparation API
+    //
+
+    // find places for the provided ubatches in the cache, returns the slot infos
+    // return empty vector on failure
+    slot_info_vec_t prepare(const std::vector<llama_ubatch> & ubatches);
+
+    bool update(llama_context * lctx, bool do_shift, const stream_copy_info & sc_info);
+
+    // find a slot of kv cells that can hold the ubatch
+    // if cont == true, then the slot must be continuous
+    // return empty slot_info on failure
+    slot_info find_slot(const llama_ubatch & ubatch, bool cont) const;
+
+    // emplace the ubatch context into slot: [sinfo.idxs[0...ubatch.n_tokens - 1]]
+    void apply_ubatch(const slot_info & sinfo, const llama_ubatch & ubatch);
+
+    //
+    // input API
+    //
+
+    ggml_tensor * build_input_k_idxs(ggml_context * ctx, const llama_ubatch & ubatch) const;
+    ggml_tensor * build_input_v_idxs(ggml_context * ctx, const llama_ubatch & ubatch) const;
+
+    void set_input_k_idxs(ggml_tensor * dst, const llama_ubatch * ubatch, const slot_info & sinfo) const;
+    void set_input_v_idxs(ggml_tensor * dst, const llama_ubatch * ubatch, const slot_info & sinfo) const;
+
+    void set_input_k_shift(ggml_tensor * dst) const;
 
-    // simulate full cache, used for allocating worst-case compute buffers
-    virtual llama_memory_state_ptr init_full() = 0;
+    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;
 
-    // process any pending defrag/shift/etc. operations
-    // optionally call once before processing a new batch
-    // return true if any operations were performed
-    virtual bool update(llama_context & lctx) = 0;
+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;
+
+        std::vector<ggml_tensor *> k_stream;
+        std::vector<ggml_tensor *> v_stream;
+    };
+
+    bool v_trans = true;  // the value tensor is transposed
+
+    const uint32_t n_seq_max = 1;
+    const uint32_t n_stream  = 1;
+
+    // required padding
+    const uint32_t n_pad = 1;
+
+    // SWA
+    const uint32_t n_swa = 0;
+
+    // env: LLAMA_KV_CACHE_DEBUG
+    int debug = 0;
+
+    // this is the SWA type of the cache - not to be confused with the model SWA type
+    const llama_swa_type swa_type = LLAMA_SWA_TYPE_NONE;
+
+    // ggml contexts for the KV cache along with the allocated backend buffers:
+    std::vector<std::pair<ggml_context_ptr, ggml_backend_buffer_ptr>> ctxs_bufs;
+
+    // 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
+    std::vector<uint32_t> v_heads;
+
+    std::vector<llama_kv_cells> v_cells;
+
+    // maps from a sequence id to a stream id
+    std::vector<uint32_t> seq_to_stream;
+
+    // pending stream copies that will be applied during the next update
+    stream_copy_info sc_info;
+
+    std::vector<kv_layer> layers;
+
+    // model layer id -> KV cache layer id
+    std::unordered_map<int32_t, int32_t> map_layer_ids;
+
+    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;
+
+    ggml_cgraph * build_graph_shift(
+               llm_graph_result * res,
+                  llama_context * lctx) const;
+
+    struct cell_ranges_t {
+        uint32_t strm;
+
+        std::vector<std::pair<uint32_t, uint32_t>> data; // ranges, from inclusive, to exclusive
+    };
+
+    void state_write_meta(llama_io_write_i & io, const cell_ranges_t & cr, llama_seq_id seq_id = -1) const;
+    void state_write_data(llama_io_write_i & io, const cell_ranges_t & cr) const;
+
+    bool state_read_meta(llama_io_read_i & io, uint32_t strm, uint32_t cell_count,       slot_info & sinfo, llama_seq_id dest_seq_id = -1);
+    bool state_read_data(llama_io_read_i & io, uint32_t strm, uint32_t cell_count, const slot_info & sinfo);
+};
+
+class llama_kv_cache_context : public llama_memory_context_i {
+public:
+    // some shorthands
+    using slot_info_vec_t  = llama_kv_cache::slot_info_vec_t;
+    using stream_copy_info = llama_kv_cache::stream_copy_info;
+
+    // used for errors
+    llama_kv_cache_context(llama_memory_status status);
+
+    // used to create a full-cache context
+    llama_kv_cache_context(
+            llama_kv_cache * kv);
+
+    // used to create an update context
+    llama_kv_cache_context(
+            llama_kv_cache * kv,
+            llama_context * lctx,
+            bool do_shift,
+            stream_copy_info sc_info);
+
+    // used to create a batch processing context from a batch
+    llama_kv_cache_context(
+            llama_kv_cache * kv,
+            slot_info_vec_t sinfos,
+            std::vector<llama_ubatch> ubatches);
+
+    virtual ~llama_kv_cache_context();
+
+    //
+    // llama_memory_context_i
+    //
+
+    bool next()  override;
+    bool apply() override;
+
+    llama_memory_status  get_status() const override;
+    const llama_ubatch & get_ubatch() const override;
 
-    // schedule a defrag if the fragmentation threshold is exceeded. otherwise, do nothing
-    // TODO: change to
-    //   llama_memory_state_ptr init_defrag(float thold) = 0;
     //
-    virtual void defrag_sched(float thold) = 0;
+    // llama_kv_cache_context 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
+    // note: the heads in k_cur and v_cur should be layed out contiguously in memory
+    //   - k_cur  [n_embd_head_k, n_head_k, n_tokens]
+    //   - k_idxs [n_tokens]
+    //   - v_cur  [n_embd_head_v, n_head_v, n_tokens]
+    //   - v_idxs [n_tokens] or [n_tokens*n_embd_v_gqa] depending if V cache is transposed
+    ggml_tensor * cpy_k(ggml_context * ctx, ggml_tensor * k_cur, ggml_tensor * k_idxs, int32_t il) const;
+    ggml_tensor * cpy_v(ggml_context * ctx, ggml_tensor * v_cur, ggml_tensor * v_idxs, int32_t il) const;
+
+    // create destination indices for each head of the current batch for where it would be written in the KV cache
+    // the indices address the global KV cache (not per stream) - this is not relevant for the user of this API, but
+    //   helps understand the implementation logic of cpy_k and cpy_v
+    ggml_tensor * build_input_k_idxs(ggml_context * ctx, const llama_ubatch & ubatch) const;
+    ggml_tensor * build_input_v_idxs(ggml_context * ctx, const llama_ubatch & ubatch) const;
+
+    void set_input_k_idxs(ggml_tensor * dst, const llama_ubatch * ubatch) const;
+    void set_input_v_idxs(ggml_tensor * dst, const llama_ubatch * ubatch) 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 * kv;
+    llama_context * lctx;
+
+    //
+    // update context
+    //
+
+    bool do_shift = false;
+
+    stream_copy_info sc_info;
+
+    //
+    // batch processing context
+    //
+
+    // the index of the cur ubatch to process
+    size_t i_cur = 0;
 
-    // getters
-    virtual bool get_can_shift() const = 0;
+    slot_info_vec_t sinfos;
 
-    bool get_can_edit() const override { return get_can_shift(); }
+    std::vector<llama_ubatch> ubatches;
 
     //
-    // state write/read
+    // data needed for building the compute graph for the current ubatch:
     //
 
-    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;
+    // 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;
 };

data/ext/sources/examples/talk-llama/llama-kv-cells.h

@@ -5,17 +5,36 @@
 
 #include <bitset>
 #include <cassert>
-#include <vector>
-#include <set>
+#include <cstring>
 #include <map>
+#include <set>
+#include <vector>
+
+struct llama_kv_cell_ext {
+    // 2D spatial positions, typically used for M-RoPE
+    llama_pos x = 0;
+    llama_pos y = 0;
+
+    // return true if the current 2D spatial position is greater than other
+    bool is_2d_gt(llama_pos ox, llama_pos oy) const {
+        return (y > oy) || (y == oy && x > ox);
+    }
+
+    void reset() {
+        static_assert(std::is_trivially_copyable_v<llama_kv_cell_ext>);
+
+        memset(this, 0, sizeof(*this));
+    }
+};
 
 // meta information about KV cells that can be part of multiple sequences at the same time
 // TODO: add unit tests
-class llama_kv_cells_unified {
+class llama_kv_cells {
 public:
     void reset() {
         for (uint32_t i = 0; i < pos.size(); ++i) {
             pos[i]   = -1;
+            ext[i].reset();
             shift[i] =  0;
             seq[i].reset();
         }
@@ -43,6 +62,7 @@ public:
 
     void resize(uint32_t n) {
         pos.resize(n);
+        ext.resize(n);
         shift.resize(n);
         seq.resize(n);
 
@@ -77,68 +97,124 @@ public:
     }
 
     // move cell isrc to idst (used during defrag)
-    void mv(uint32_t isrc, uint32_t idst) {
-        assert(isrc < pos.size());
-        assert(idst < pos.size());
+    //void mv(uint32_t isrc, uint32_t idst) {
+    //    assert(isrc < pos.size());
+    //    assert(idst < pos.size());
 
-        assert(pos[idst] == -1);
-        assert(pos[isrc] != -1);
+    //    assert(pos[idst] == -1);
+    //    assert(pos[isrc] != -1);
 
-        pos  [idst] = pos  [isrc];
-        shift[idst] = shift[isrc];
-        seq  [idst] = seq  [isrc];
+    //    pos  [idst] = pos  [isrc];
+    //    shift[idst] = shift[isrc];
+    //    seq  [idst] = seq  [isrc];
 
-        pos  [isrc] = -1;
-        shift[isrc] =  0;
-        seq  [isrc].reset();
+    //    pos  [isrc] = -1;
+    //    shift[isrc] =  0;
+    //    seq  [isrc].reset();
 
-        used.erase (isrc);
-        used.insert(idst);
-    }
+    //    used.erase (isrc);
+    //    used.insert(idst);
+    //}
 
     // copy the state of cells [i, i + n) (used for save/restore the state of the cells)
-    llama_kv_cells_unified cp(uint32_t i, uint32_t n) const {
+    llama_kv_cells cp(uint32_t i, uint32_t n) const {
         assert(i + n <= pos.size());
 
-        llama_kv_cells_unified res;
+        llama_kv_cells res;
 
         res.resize(n);
 
         for (uint32_t j = 0; j < n; ++j) {
-            res.pos[j] = pos[i + j];
-            res.seq[j] = seq[i + j];
+            const auto idx = i + j;
+
+            res.pos[j] = pos[idx];
+            res.ext[j] = ext[idx];
+            res.seq[j] = seq[idx];
+
+            assert(shift[idx] == 0);
+        }
+
+        return res;
+    }
+
+    // copy the state of cells [idxs[0], idxs[1], ..., idxs[idxs.size() - 1])
+    llama_kv_cells cp(const std::vector<uint32_t> & idxs) const {
+        llama_kv_cells res;
+
+        res.resize(idxs.size());
 
-            assert(shift[i + j] == 0);
+        for (uint32_t j = 0; j < idxs.size(); ++j) {
+            const auto idx = idxs[j];
+
+            res.pos[j] = pos[idx];
+            res.ext[j] = ext[idx];
+            res.seq[j] = seq[idx];
+
+            assert(shift[idx] == 0);
         }
 
         return res;
     }
 
     // set the state of cells [i, i + other.pos.size()) (used for save/restore the state of the cells)
-    void set(uint32_t i, const llama_kv_cells_unified & other) {
+    void set(uint32_t i, const llama_kv_cells & other) {
         assert(i + other.pos.size() <= pos.size());
 
         for (uint32_t j = 0; j < other.pos.size(); ++j) {
-            if (pos[i + j] == -1 && other.pos[j] != -1) {
+            const auto idx = i + j;
+
+            if (pos[idx] == -1 && other.pos[j] != -1) {
                 used.insert(i + j);
             }
 
-            if (pos[i + j] != -1 && other.pos[j] == -1) {
+            if (pos[idx] != -1 && other.pos[j] == -1) {
                 used.erase(i + j);
             }
 
-            if (pos[i + j] != -1) {
+            if (pos[idx] != -1) {
                 seq_pos_rm(i + j);
             }
 
-            pos[i + j] = other.pos[j];
-            seq[i + j] = other.seq[j];
+            pos[idx] = other.pos[j];
+            ext[idx] = other.ext[j];
+            seq[idx] = other.seq[j];
 
-            if (pos[i + j] != -1) {
+            if (pos[idx] != -1) {
                 seq_pos_add(i + j);
             }
 
-            assert(shift[i + j] == 0);
+            assert(shift[idx] == 0);
+        }
+    }
+
+    // set the state of cells [idxs[0], idxs[1], ..., idxs[idxs.size() - 1])
+    void set(const std::vector<uint32_t> & idxs, const llama_kv_cells & other) {
+        assert(idxs.size() == other.pos.size());
+
+        for (uint32_t j = 0; j < other.pos.size(); ++j) {
+            const auto idx = idxs[j];
+
+            if (pos[idx] == -1 && other.pos[j] != -1) {
+                used.insert(idx);
+            }
+
+            if (pos[idx] != -1 && other.pos[j] == -1) {
+                used.erase(idx);
+            }
+
+            if (pos[idx] != -1) {
+                seq_pos_rm(idx);
+            }
+
+            pos[idx] = other.pos[j];
+            ext[idx] = other.ext[j];
+            seq[idx] = other.seq[j];
+
+            if (pos[idx] != -1) {
+                seq_pos_add(idx);
+            }
+
+            assert(shift[idx] == 0);
         }
     }
 
@@ -151,6 +227,7 @@ public:
         seq[i].reset();
 
         pos[i] = -1;
+        ext[i].reset();
         shift[i] = 0;
 
         used.erase(i);
@@ -169,6 +246,7 @@ public:
 
         if (seq[i].none()) {
             pos[i] = -1;
+            ext[i].reset();
             shift[i] = 0;
 
             used.erase(i);
@@ -198,6 +276,7 @@ public:
             seq[i].reset();
 
             pos[i] = -1;
+            ext[i].reset();
             shift[i] = 0;
 
             used.erase(i);
@@ -288,6 +367,13 @@ public:
         return pos[i];
     }
 
+    const llama_kv_cell_ext & ext_get(uint32_t i) const {
+        assert(i < pos.size());
+        assert(pos[i] != -1);
+
+        return ext[i];
+    }
+
     // note: call only if the cell is not empty
     llama_pos get_shift(uint32_t i) const {
         assert(i < pos.size());
@@ -316,6 +402,11 @@ public:
         used.insert(i);
     }
 
+    void ext_set(uint32_t i, llama_kv_cell_ext p) {
+        assert(i < ext.size());
+        ext[i] = p;
+    }
+
     // pos[i] = pos[i] + d
     // sets "has_shift" to true
     // note: call only if the cell is not empty
@@ -372,6 +463,9 @@ private:
 
     std::vector<llama_pos> pos;
 
+    // stores extra info per cell
+    std::vector<llama_kv_cell_ext> ext;
+
     // this array accumulates any applied shifts to the pos array since the last reset_shift() call
     // this is used to queue multiple updates to the pos array, which in the end can be applied in one go:
     //