| 1 | #pragma once |
| 2 | |
| 3 | #include "llama-batch.h" |
| 4 | #include "llama-graph.h" |
| 5 | #include "llama-kv-cache.h" |
| 6 | #include "llama-memory.h" |
| 7 | #include "llama-memory-recurrent.h" |
| 8 | |
| 9 | #include <memory> |
| 10 | #include <vector> |
| 11 | |
| 12 | // |
| 13 | // llama_memory_hybrid |
| 14 | // |
| 15 | |
| 16 | // utilizes instances of llama_memory_recurrent and llama_kv_cache to |
| 17 | // support models where each layer may be either attention-based or recurrent |
| 18 | |
| 19 | class llama_memory_hybrid : public llama_memory_i { |
| 20 | public: |
| 21 | llama_memory_hybrid( |
| 22 | const llama_model & model, |
| 23 | /* attn */ |
| 24 | ggml_type type_k, |
| 25 | ggml_type type_v, |
| 26 | bool v_trans, |
| 27 | uint32_t kv_size, |
| 28 | uint32_t n_pad, |
| 29 | uint32_t n_swa, |
| 30 | llama_swa_type swa_type, |
| 31 | /* recurrent */ |
| 32 | ggml_type type_r, |
| 33 | ggml_type type_s, |
| 34 | uint32_t rs_size, |
| 35 | /* common */ |
| 36 | uint32_t n_seq_max, |
| 37 | bool offload, |
| 38 | bool unified, |
| 39 | /* layer filters */ |
| 40 | const layer_filter_cb & filter_attn = nullptr, |
| 41 | const layer_filter_cb & filter_recr = nullptr); |
| 42 | |
| 43 | ~llama_memory_hybrid() = default; |
| 44 | |
| 45 | // |
| 46 | // llama_memory_i |
| 47 | // |
| 48 | |
| 49 | llama_memory_context_ptr init_batch( |
| 50 | llama_batch_allocr & balloc, |
| 51 | uint32_t n_ubatch, |
| 52 | bool embd_all) override; |
| 53 | |
| 54 | llama_memory_context_ptr init_full() override; |
| 55 | |
| 56 | llama_memory_context_ptr init_update(llama_context * lctx, bool optimize) override; |
| 57 | |
| 58 | bool get_can_shift() const override; |
| 59 | |
| 60 | void clear(bool data) override; |
| 61 | |
| 62 | bool seq_rm (llama_seq_id seq_id, llama_pos p0, llama_pos p1) override; |
| 63 | void seq_cp (llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) override; |
| 64 | void seq_keep(llama_seq_id seq_id) override; |
| 65 | void seq_add (llama_seq_id seq_id, llama_pos p0, llama_pos p1, llama_pos shift) override; |
| 66 | void seq_div (llama_seq_id seq_id, llama_pos p0, llama_pos p1, int d) override; |
| 67 | |
| 68 | llama_pos seq_pos_min(llama_seq_id seq_id) const override; |
| 69 | llama_pos seq_pos_max(llama_seq_id seq_id) const override; |
| 70 | |
| 71 | std::map<ggml_backend_buffer_type_t, size_t> memory_breakdown() const override; |
| 72 | |
| 73 | // state write/load |
| 74 | |
| 75 | void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1, llama_state_seq_flags flags = 0) const override; |
| 76 | void state_read (llama_io_read_i & io, llama_seq_id seq_id = -1, llama_state_seq_flags flags = 0) override; |
| 77 | |
| 78 | // |
| 79 | // llama_memory_hybrid specific API |
| 80 | // |
| 81 | |
| 82 | llama_kv_cache * get_mem_attn() const; |
| 83 | llama_memory_recurrent * get_mem_recr() const; |
| 84 | |
| 85 | private: |
| 86 | const llama_hparams & hparams; |
| 87 | |
| 88 | const std::unique_ptr<llama_kv_cache> mem_attn; |
| 89 | const std::unique_ptr<llama_memory_recurrent> mem_recr; |
| 90 | }; |
| 91 | |
| 92 | class llama_memory_hybrid_context : public llama_memory_context_i { |
| 93 | public: |
| 94 | using slot_info_vec_t = llama_kv_cache::slot_info_vec_t; |
| 95 | |
| 96 | // init failure |
| 97 | explicit llama_memory_hybrid_context(llama_memory_status status); |
| 98 | |
| 99 | // init full |
| 100 | explicit llama_memory_hybrid_context(llama_memory_hybrid * mem); |
| 101 | |
| 102 | // init update |
| 103 | explicit llama_memory_hybrid_context( |
| 104 | llama_memory_hybrid * mem, |
| 105 | llama_context * lctx, |
| 106 | bool optimize); |
| 107 | |
| 108 | // init success |
| 109 | llama_memory_hybrid_context( |
| 110 | llama_memory_hybrid * mem, |
| 111 | slot_info_vec_t sinfos_attn, |
| 112 | std::vector<llama_ubatch> ubatches); |
| 113 | |
| 114 | ~llama_memory_hybrid_context() = default; |
| 115 | |
| 116 | bool next() override; |
| 117 | bool apply() override; |
| 118 | |
| 119 | llama_memory_status get_status() const override; |
| 120 | const llama_ubatch & get_ubatch() const override; |
| 121 | |
| 122 | // |
| 123 | // llama_memory_hybrid_context |
| 124 | // |
| 125 | |
| 126 | const llama_kv_cache_context * get_attn() const; |
| 127 | const llama_memory_recurrent_context * get_recr() const; |
| 128 | |
| 129 | private: |
| 130 | // the index of the next ubatch to process |
| 131 | size_t i_next = 0; |
| 132 | |
| 133 | std::vector<llama_ubatch> ubatches; |
| 134 | |
| 135 | const llama_memory_context_ptr ctx_attn; |
| 136 | const llama_memory_context_ptr ctx_recr; |
| 137 | |
| 138 | const llama_memory_status status; |
| 139 | }; |
| 140 | |