| 1 | #pragma once |
| 2 | |
| 3 | #include "llama.h" |
| 4 | |
| 5 | #include <map> |
| 6 | #include <memory> |
| 7 | #include <functional> |
| 8 | |
| 9 | struct llama_ubatch; |
| 10 | |
| 11 | class llama_batch_allocr; |
| 12 | |
| 13 | class llama_io_write_i; |
| 14 | class llama_io_read_i; |
| 15 | |
| 16 | struct llama_memory_params { |
| 17 | // kv cache |
| 18 | ggml_type type_k; |
| 19 | ggml_type type_v; |
| 20 | |
| 21 | // use full-size SWA cache |
| 22 | bool swa_full; |
| 23 | }; |
| 24 | |
| 25 | enum llama_memory_status { |
| 26 | LLAMA_MEMORY_STATUS_SUCCESS = 0, |
| 27 | LLAMA_MEMORY_STATUS_NO_UPDATE, |
| 28 | LLAMA_MEMORY_STATUS_FAILED_PREPARE, |
| 29 | LLAMA_MEMORY_STATUS_FAILED_COMPUTE, |
| 30 | }; |
| 31 | |
| 32 | // helper function for combining the status of two memory contexts |
| 33 | // useful for implementing hybrid memory types (e.g. iSWA) |
| 34 | llama_memory_status llama_memory_status_combine(llama_memory_status s0, llama_memory_status s1); |
| 35 | |
| 36 | // helper function for checking if a memory status indicates a failure |
| 37 | bool llama_memory_status_is_fail(llama_memory_status status); |
| 38 | |
| 39 | // the interface for managing the memory context during batch processing |
| 40 | // this interface is implemented per memory type. see: |
| 41 | // - llama_kv_cache_context |
| 42 | // - llama_kv_cache_iswa_context |
| 43 | // ... |
| 44 | // |
| 45 | // the only method that should mutate the memory and the memory context is llama_memory_i::apply() |
| 46 | struct llama_memory_context_i { |
| 47 | virtual ~llama_memory_context_i() = default; |
| 48 | |
| 49 | // consume the current ubatch from the context and proceed to the next one |
| 50 | // return false if we are done |
| 51 | virtual bool next() = 0; |
| 52 | |
| 53 | // apply the memory state for the current ubatch to the memory object |
| 54 | // return false on failure |
| 55 | virtual bool apply() = 0; |
| 56 | |
| 57 | // get the current ubatch |
| 58 | virtual const llama_ubatch & get_ubatch() const = 0; |
| 59 | |
| 60 | // get the status of the memory context - used for error handling and checking if any updates would be applied |
| 61 | virtual llama_memory_status get_status() const = 0; |
| 62 | }; |
| 63 | |
| 64 | using llama_memory_context_ptr = std::unique_ptr<llama_memory_context_i>; |
| 65 | |
| 66 | // general concept of LLM memory |
| 67 | // the KV cache is a type of LLM memory, but there can be other types |
| 68 | struct llama_memory_i { |
| 69 | // this callback is used to filter out layers that should not be included in the cache |
| 70 | using layer_filter_cb = std::function<bool(int32_t il)>; |
| 71 | |
| 72 | // this callback is used to specify which layers should reuse memory from other layers |
| 73 | // return negative value to indicate that the layer il should not reuse memory |
| 74 | using layer_reuse_cb = std::function<int32_t(int32_t il)>; |
| 75 | |
| 76 | virtual ~llama_memory_i() = default; |
| 77 | |
| 78 | // split the input batch into a set of ubatches and verify that they can fit into the cache |
| 79 | // return a context object containing the ubatches and memory state required to process them |
| 80 | // check the llama_memory_context_i::get_status() for the result |
| 81 | virtual llama_memory_context_ptr init_batch( |
| 82 | llama_batch_allocr & balloc, |
| 83 | uint32_t n_ubatch, |
| 84 | bool embd_all) = 0; |
| 85 | |
| 86 | // simulate full cache, used for allocating worst-case compute buffers |
| 87 | virtual llama_memory_context_ptr init_full() = 0; |
| 88 | |
| 89 | // prepare for any pending memory updates, such as shifts, copies, etc. |
| 90 | // status == LLAMA_MEMORY_STATUS_NO_UPDATE if there is nothing to update |
| 91 | virtual llama_memory_context_ptr init_update(llama_context * lctx, bool optimize) = 0; |
| 92 | |
| 93 | // getters |
| 94 | virtual bool get_can_shift() const = 0; |
| 95 | |
| 96 | // |
| 97 | // ops |
| 98 | // |
| 99 | |
| 100 | // if data == true, the data buffers will also be cleared together with the metadata |
| 101 | virtual void clear(bool data) = 0; |
| 102 | |
| 103 | virtual bool seq_rm (llama_seq_id seq_id, llama_pos p0, llama_pos p1) = 0; |
| 104 | virtual void seq_cp (llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) = 0; |
| 105 | virtual void seq_keep(llama_seq_id seq_id) = 0; |
| 106 | virtual void seq_add (llama_seq_id seq_id, llama_pos p0, llama_pos p1, llama_pos shift) = 0; |
| 107 | virtual void seq_div (llama_seq_id seq_id, llama_pos p0, llama_pos p1, int d) = 0; |
| 108 | |
| 109 | virtual llama_pos seq_pos_min(llama_seq_id seq_id) const = 0; |
| 110 | virtual llama_pos seq_pos_max(llama_seq_id seq_id) const = 0; |
| 111 | |
| 112 | virtual std::map<ggml_backend_buffer_type_t, size_t> memory_breakdown() const = 0; |
| 113 | |
| 114 | // |
| 115 | // state write/read |
| 116 | // |
| 117 | |
| 118 | virtual void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1, llama_state_seq_flags flags = 0) const = 0; |
| 119 | virtual void state_read (llama_io_read_i & io, llama_seq_id seq_id = -1, llama_state_seq_flags flags = 0) = 0; |
| 120 | }; |
| 121 | |
| 122 | using llama_memory_ptr = std::unique_ptr<llama_memory_i>; |
| 123 | |