1#pragma once
  2
  3#include "llama-batch.h"
  4#include "llama-graph.h"
  5#include "llama-kv-cells.h"
  6#include "llama-memory.h"
  7
  8#include <unordered_map>
  9#include <vector>
 10
 11struct llama_cparams;
 12struct llama_hparams;
 13struct llama_model;
 14struct llama_context;
 15
 16//
 17// llama_kv_cache
 18//
 19
 20class llama_kv_cache : public llama_memory_i {
 21public:
 22    struct stream_copy_info {
 23        bool empty() const {
 24            assert(ssrc.size() == sdst.size());
 25            return ssrc.empty();
 26        }
 27
 28        std::vector<uint32_t> ssrc;
 29        std::vector<uint32_t> sdst;
 30    };
 31
 32    // for each ubatch, create a slot_info that contains information about where the ubatch should be inserted in the
 33    //   KV cells. for example, cell indices for each token, such that: token[i] -> goes to cells[idxs[i]]
 34    struct slot_info {
 35        // data for ggml_set_rows
 36        using idx_vec_t = std::vector<uint32_t>;
 37
 38        // number of streams: ns = s1 - s0 + 1
 39        uint32_t s0;
 40        uint32_t s1;
 41
 42        std::vector<llama_seq_id> strm; // [ns]
 43        std::vector<idx_vec_t>    idxs; // [ns]
 44
 45        uint32_t head() const {
 46            GGML_ASSERT(idxs.size() == 1);
 47            GGML_ASSERT(!idxs[0].empty());
 48
 49            return idxs[0][0];
 50        }
 51
 52        void resize(size_t n) {
 53            strm.resize(n);
 54            idxs.resize(n);
 55        }
 56
 57        size_t size() const {
 58            GGML_ASSERT(idxs.size() == strm.size());
 59            GGML_ASSERT(!idxs.empty());
 60
 61            return idxs[0].size();
 62        }
 63
 64        size_t n_stream() const {
 65            return strm.size();
 66        }
 67
 68        bool empty() const {
 69            return idxs.empty();
 70        }
 71
 72        void clear() {
 73            idxs.clear();
 74        }
 75
 76        // check if indices are contiguous starting from head()
 77        bool is_contiguous() const {
 78            if (idxs.empty() || idxs[0].empty()) {
 79                return true;
 80            }
 81            if (idxs.size() > 1) {
 82                return false;
 83            }
 84            const uint32_t h = idxs[0][0];
 85            for (size_t i = 0; i < idxs[0].size(); ++i) {
 86                if (idxs[0][i] != h + i) {
 87                    return false;
 88                }
 89            }
 90            return true;
 91        }
 92    };
 93
 94    using slot_info_vec_t = std::vector<slot_info>;
 95
 96    llama_kv_cache(
 97            const llama_model & model,
 98                    ggml_type   type_k,
 99                    ggml_type   type_v,
100                         bool   v_trans,
101                         bool   offload,
102                         bool   unified,
103                     uint32_t   kv_size,
104                     uint32_t   n_seq_max,
105                     uint32_t   n_pad,
106                     uint32_t   n_swa,
107               llama_swa_type   swa_type,
108        const layer_filter_cb & filter,
109        const  layer_reuse_cb & reuse);
110
111    ~llama_kv_cache() = default;
112
113    //
114    // llama_memory_i
115    //
116
117    llama_memory_context_ptr init_batch(
118            llama_batch_allocr & balloc,
119            uint32_t n_ubatch,
120            bool embd_all) override;
121
122    llama_memory_context_ptr init_full() override;
123
124    llama_memory_context_ptr init_update(llama_context * lctx, bool optimize) override;
125
126    bool get_can_shift() const override;
127
128    void clear(bool data) override;
129
130    bool seq_rm  (llama_seq_id seq_id,                              llama_pos p0, llama_pos p1) override;
131    void seq_cp  (llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) override;
132    void seq_keep(llama_seq_id seq_id)                                                          override;
133    void seq_add (llama_seq_id seq_id,                              llama_pos p0, llama_pos p1, llama_pos shift) override;
134    void seq_div (llama_seq_id seq_id,                              llama_pos p0, llama_pos p1, int d) override;
135
136    llama_pos seq_pos_min(llama_seq_id seq_id) const override;
137    llama_pos seq_pos_max(llama_seq_id seq_id) const override;
138
139    std::map<ggml_backend_buffer_type_t, size_t> memory_breakdown() const override;
140
141    // state write/load
142
143    void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1, llama_state_seq_flags flags = 0) const override;
144    void state_read (llama_io_read_i  & io, llama_seq_id seq_id = -1, llama_state_seq_flags flags = 0) override;
145
146    //
147    // llama_kv_cache specific API
148    //
149
150    uint32_t get_size()     const;
151    uint32_t get_n_stream() const;
152
153    bool get_has_shift() const;
154
155    //
156    // graph_build API
157    //
158
159    uint32_t get_n_kv(const slot_info & sinfo) const;
160
161    // get views of the current state of the cache
162    ggml_tensor * get_k(ggml_context * ctx, int32_t il, uint32_t n_kv, const slot_info & sinfo) const;
163    ggml_tensor * get_v(ggml_context * ctx, int32_t il, uint32_t n_kv, const slot_info & sinfo) const;
164
165    // store k_cur and v_cur in the cache based on the provided head location
166    ggml_tensor * cpy_k(ggml_context * ctx, ggml_tensor * k_cur, ggml_tensor * k_idxs, int32_t il, const slot_info & sinfo) const;
167    ggml_tensor * cpy_v(ggml_context * ctx, ggml_tensor * v_cur, ggml_tensor * v_idxs, int32_t il, const slot_info & sinfo) const;
168
169    //
170    // preparation API
171    //
172
173    // find places for the provided ubatches in the cache, returns the slot infos
174    // return empty vector on failure
175    slot_info_vec_t prepare(const std::vector<llama_ubatch> & ubatches);
176
177    bool update(llama_context * lctx, bool do_shift, const stream_copy_info & sc_info);
178
179    // find a slot of kv cells that can hold the ubatch
180    // if cont == true, then the slot must be continuous
181    // return empty slot_info on failure
182    slot_info find_slot(const llama_ubatch & ubatch, bool cont) const;
183
184    // emplace the ubatch context into slot: [sinfo.idxs[0...ubatch.n_tokens - 1]]
185    void apply_ubatch(const slot_info & sinfo, const llama_ubatch & ubatch);
186
187    //
188    // input API
189    //
190
191    ggml_tensor * build_input_k_idxs(ggml_context * ctx, const llama_ubatch & ubatch) const;
192    ggml_tensor * build_input_v_idxs(ggml_context * ctx, const llama_ubatch & ubatch) const;
193
194    void set_input_k_idxs(ggml_tensor * dst, const llama_ubatch * ubatch, const slot_info & sinfo) const;
195    void set_input_v_idxs(ggml_tensor * dst, const llama_ubatch * ubatch, const slot_info & sinfo) const;
196
197    void set_input_k_shift(ggml_tensor * dst) const;
198
199    void set_input_kq_mask   (ggml_tensor * dst, const llama_ubatch * ubatch, bool causal_attn) const;
200    void set_input_pos_bucket(ggml_tensor * dst, const llama_ubatch * ubatch) const;
201
202private:
203    const llama_model & model;
204    const llama_hparams & hparams;
205
206    struct kv_layer {
207        // layer index in the model
208        // note: can be different from the layer index in the KV cache
209        uint32_t il;
210
211        ggml_tensor * k;
212        ggml_tensor * v;
213
214        std::vector<ggml_tensor *> k_stream;
215        std::vector<ggml_tensor *> v_stream;
216    };
217
218    bool v_trans = true;  // the value tensor is transposed
219
220    const uint32_t n_seq_max = 1;
221    const uint32_t n_stream  = 1;
222
223    // required padding
224    const uint32_t n_pad = 1;
225
226    // SWA
227    const uint32_t n_swa = 0;
228
229    // env: LLAMA_KV_CACHE_DEBUG
230    int debug = 0;
231
232    // this is the SWA type of the cache - not to be confused with the model SWA type
233    const llama_swa_type swa_type = LLAMA_SWA_TYPE_NONE;
234
235    // ggml contexts for the KV cache along with the allocated backend buffers:
236    std::vector<std::pair<ggml_context_ptr, ggml_backend_buffer_ptr>> ctxs_bufs;
237
238    // the current index from where we start searching for a free slot in the ring buffer of KV cells (see find_slot())
239    // note: this is not part of the KV state and it's only used to speed-up the find_slot() method
240    std::vector<uint32_t> v_heads;
241
242    std::vector<llama_kv_cells> v_cells;
243
244    // maps from a sequence id to a stream id
245    std::vector<uint32_t> seq_to_stream;
246
247    // pending stream copies that will be applied during the next update
248    stream_copy_info sc_info;
249
250    std::vector<kv_layer> layers;
251
252    // model layer id -> KV cache layer id
253    std::unordered_map<int32_t, int32_t> map_layer_ids;
254
255    size_t total_size() const;
256
257    size_t size_k_bytes() const;
258    size_t size_v_bytes() const;
259
260    ggml_tensor * build_rope_shift(
261            const llama_cparams & cparams,
262                   ggml_context * ctx,
263                    ggml_tensor * cur,
264                    ggml_tensor * shift,
265                    ggml_tensor * factors,
266                          float   freq_base,
267                          float   freq_scale) const;
268
269    ggml_cgraph * build_graph_shift(
270               llm_graph_result * res,
271                  llama_context * lctx) const;
272
273    struct cell_ranges_t {
274        uint32_t strm;
275
276        std::vector<std::pair<uint32_t, uint32_t>> data; // ranges, from inclusive, to exclusive
277    };
278
279    void state_write_meta(llama_io_write_i & io, const cell_ranges_t & cr, llama_seq_id seq_id = -1) const;
280    void state_write_data(llama_io_write_i & io, const cell_ranges_t & cr) const;
281
282    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);
283    bool state_read_data(llama_io_read_i & io, uint32_t strm, uint32_t cell_count, const slot_info & sinfo);
284};
285
286class llama_kv_cache_context : public llama_memory_context_i {
287public:
288    // some shorthands
289    using slot_info_vec_t  = llama_kv_cache::slot_info_vec_t;
290    using stream_copy_info = llama_kv_cache::stream_copy_info;
291
292    // used for errors
293    llama_kv_cache_context(llama_memory_status status);
294
295    // used to create a full-cache context
296    llama_kv_cache_context(
297            llama_kv_cache * kv);
298
299    // used to create an update context
300    llama_kv_cache_context(
301            llama_kv_cache * kv,
302            llama_context * lctx,
303            bool do_shift,
304            stream_copy_info sc_info);
305
306    // used to create a batch processing context from a batch
307    llama_kv_cache_context(
308            llama_kv_cache * kv,
309            slot_info_vec_t sinfos,
310            std::vector<llama_ubatch> ubatches);
311
312    virtual ~llama_kv_cache_context();
313
314    //
315    // llama_memory_context_i
316    //
317
318    bool next()  override;
319    bool apply() override;
320
321    llama_memory_status  get_status() const override;
322    const llama_ubatch & get_ubatch() const override;
323
324    //
325    // llama_kv_cache_context specific API
326    //
327
328    uint32_t get_n_kv() const;
329
330    // get views of the current state of the cache
331    ggml_tensor * get_k(ggml_context * ctx, int32_t il) const;
332    ggml_tensor * get_v(ggml_context * ctx, int32_t il) const;
333
334    // store k_cur and v_cur in the cache based on the provided head location
335    // note: the heads in k_cur and v_cur should be layed out contiguously in memory
336    //   - k_cur  [n_embd_head_k, n_head_k, n_tokens]
337    //   - k_idxs [n_tokens]
338    //   - v_cur  [n_embd_head_v, n_head_v, n_tokens]
339    //   - v_idxs [n_tokens] or [n_tokens*n_embd_v_gqa] depending if V cache is transposed
340    ggml_tensor * cpy_k(ggml_context * ctx, ggml_tensor * k_cur, ggml_tensor * k_idxs, int32_t il) const;
341    ggml_tensor * cpy_v(ggml_context * ctx, ggml_tensor * v_cur, ggml_tensor * v_idxs, int32_t il) const;
342
343    // create destination indices for each head of the current batch for where it would be written in the KV cache
344    // the indices address the global KV cache (not per stream) - this is not relevant for the user of this API, but
345    //   helps understand the implementation logic of cpy_k and cpy_v
346    ggml_tensor * build_input_k_idxs(ggml_context * ctx, const llama_ubatch & ubatch) const;
347    ggml_tensor * build_input_v_idxs(ggml_context * ctx, const llama_ubatch & ubatch) const;
348
349    void set_input_k_idxs(ggml_tensor * dst, const llama_ubatch * ubatch) const;
350    void set_input_v_idxs(ggml_tensor * dst, const llama_ubatch * ubatch) const;
351
352    void set_input_k_shift   (ggml_tensor * dst) const;
353    void set_input_kq_mask   (ggml_tensor * dst, const llama_ubatch * ubatch, bool causal_attn) const;
354    void set_input_pos_bucket(ggml_tensor * dst, const llama_ubatch * ubatch) const;
355
356private:
357    llama_memory_status status;
358
359    llama_kv_cache * kv;
360    llama_context * lctx;
361
362    //
363    // update context
364    //
365
366    bool do_shift = false;
367
368    stream_copy_info sc_info;
369
370    //
371    // batch processing context
372    //
373
374    // the index of the cur ubatch to process
375    size_t i_cur = 0;
376
377    slot_info_vec_t sinfos;
378
379    std::vector<llama_ubatch> ubatches;
380
381    //
382    // data needed for building the compute graph for the current ubatch:
383    //
384
385    // a heuristic, to avoid attending the full cache if it is not yet utilized
386    // as the cache gets filled, the benefit from this heuristic disappears
387    int32_t n_kv;
388};