| 1 | #pragma once |
| 2 | |
| 3 | #include "llama-kv-cache.h" |
| 4 | |
| 5 | #include <vector> |
| 6 | |
| 7 | // |
| 8 | // llama_kv_cache_iswa |
| 9 | // |
| 10 | |
| 11 | // utilizes two instances of llama_kv_cache |
| 12 | // the first instance is for the non-SWA layers of the model and the second instance is for the SWA layers |
| 13 | |
| 14 | class llama_kv_cache_iswa : public llama_memory_i { |
| 15 | public: |
| 16 | llama_kv_cache_iswa( |
| 17 | const llama_model & model, |
| 18 | ggml_type type_k, |
| 19 | ggml_type type_v, |
| 20 | bool v_trans, |
| 21 | bool offload, |
| 22 | bool swa_full, |
| 23 | bool unified, |
| 24 | uint32_t kv_size, |
| 25 | uint32_t n_seq_max, |
| 26 | uint32_t n_ubatch, |
| 27 | uint32_t n_pad, |
| 28 | const layer_filter_cb & filter, |
| 29 | const layer_reuse_cb & reuse); |
| 30 | |
| 31 | ~llama_kv_cache_iswa() = default; |
| 32 | |
| 33 | // |
| 34 | // llama_memory_i |
| 35 | // |
| 36 | |
| 37 | llama_memory_context_ptr init_batch( |
| 38 | llama_batch_allocr & balloc, |
| 39 | uint32_t n_ubatch, |
| 40 | bool embd_all) override; |
| 41 | |
| 42 | llama_memory_context_ptr init_full() override; |
| 43 | |
| 44 | llama_memory_context_ptr init_update(llama_context * lctx, bool optimize) override; |
| 45 | |
| 46 | bool get_can_shift() const override; |
| 47 | |
| 48 | void clear(bool data) override; |
| 49 | |
| 50 | bool seq_rm (llama_seq_id seq_id, llama_pos p0, llama_pos p1) override; |
| 51 | void seq_cp (llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) override; |
| 52 | void seq_keep(llama_seq_id seq_id) override; |
| 53 | void seq_add (llama_seq_id seq_id, llama_pos p0, llama_pos p1, llama_pos shift) override; |
| 54 | void seq_div (llama_seq_id seq_id, llama_pos p0, llama_pos p1, int d) override; |
| 55 | |
| 56 | llama_pos seq_pos_min(llama_seq_id seq_id) const override; |
| 57 | llama_pos seq_pos_max(llama_seq_id seq_id) const override; |
| 58 | |
| 59 | std::map<ggml_backend_buffer_type_t, size_t> memory_breakdown() const override; |
| 60 | |
| 61 | // state write/load |
| 62 | |
| 63 | void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1, llama_state_seq_flags flags = 0) const override; |
| 64 | void state_read (llama_io_read_i & io, llama_seq_id seq_id = -1, llama_state_seq_flags flags = 0) override; |
| 65 | |
| 66 | // |
| 67 | // llama_kv_cache_iswa specific API |
| 68 | // |
| 69 | |
| 70 | llama_kv_cache * get_base() const; |
| 71 | llama_kv_cache * get_swa () const; |
| 72 | |
| 73 | private: |
| 74 | const llama_hparams & hparams; |
| 75 | |
| 76 | const bool unified; |
| 77 | |
| 78 | std::unique_ptr<llama_kv_cache> kv_base; |
| 79 | std::unique_ptr<llama_kv_cache> kv_swa; |
| 80 | }; |
| 81 | |
| 82 | class llama_kv_cache_iswa_context : public llama_memory_context_i { |
| 83 | public: |
| 84 | using slot_info_vec_t = llama_kv_cache::slot_info_vec_t; |
| 85 | |
| 86 | // used for errors |
| 87 | llama_kv_cache_iswa_context(llama_memory_status status); |
| 88 | |
| 89 | // used to create a full-cache context |
| 90 | llama_kv_cache_iswa_context( |
| 91 | llama_kv_cache_iswa * kv); |
| 92 | |
| 93 | // used to create an update context |
| 94 | llama_kv_cache_iswa_context( |
| 95 | llama_kv_cache_iswa * kv, |
| 96 | llama_context * lctx, |
| 97 | bool optimize); |
| 98 | |
| 99 | // used to create a batch processing context from a batch |
| 100 | llama_kv_cache_iswa_context( |
| 101 | llama_kv_cache_iswa * kv, |
| 102 | slot_info_vec_t sinfos_base, |
| 103 | slot_info_vec_t sinfos_swa, |
| 104 | std::vector<llama_ubatch> ubatches); |
| 105 | |
| 106 | virtual ~llama_kv_cache_iswa_context(); |
| 107 | |
| 108 | // |
| 109 | // llama_memory_context_i |
| 110 | // |
| 111 | |
| 112 | bool next() override; |
| 113 | bool apply() override; |
| 114 | |
| 115 | llama_memory_status get_status() const override; |
| 116 | const llama_ubatch & get_ubatch() const override; |
| 117 | |
| 118 | // |
| 119 | // llama_kv_cache_iswa_context specific API |
| 120 | // |
| 121 | |
| 122 | const llama_kv_cache_context * get_base() const; |
| 123 | const llama_kv_cache_context * get_swa() const; |
| 124 | |
| 125 | private: |
| 126 | //llama_kv_cache_iswa * kv; |
| 127 | |
| 128 | // the index of the next ubatch to process |
| 129 | size_t i_next = 0; |
| 130 | |
| 131 | std::vector<llama_ubatch> ubatches; |
| 132 | |
| 133 | const llama_memory_context_ptr ctx_base; |
| 134 | const llama_memory_context_ptr ctx_swa; |
| 135 | |
| 136 | const llama_memory_status status; |
| 137 | }; |
| 138 | |