char-fork/llama.cpp

git clone https://git.t4t.associates/char-fork/llama.cpp

Daniel Hanmodel: add Qwen3.8-Flash-Next (qwen4exp) (#27742)6c84c7d5d

char/aliceai-t5-moe
brancheschar/aliceai-t5-moeall refs →
6.1 KiB195 linesraw
1#pragma once
2
3#include "llama-batch.h"
4#include "llama-graph.h"
5#include "llama-memory.h"
6
7#include <map>
8#include <set>
9#include <vector>
10
11//
12// llama_memory_recurrent
13//
14
15// TODO: extract the cache state used for graph computation into llama_memory_recurrent_context_i
16//       see the implementation of llama_kv_cache_context_i for an example how to do it
17class llama_memory_recurrent : public llama_memory_i {
18public:
19    llama_memory_recurrent(
20            const llama_model & model,
21                    ggml_type   type_r,
22                    ggml_type   type_s,
23                         bool   offload,
24                     uint32_t   mem_size,
25                     uint32_t   n_seq_max,
26                     uint32_t   n_rs_seq,
27        const layer_filter_cb & filter);
28
29    ~llama_memory_recurrent() = default;
30
31    //
32    // llama_memory_i
33    //
34
35    llama_memory_context_ptr init_batch(
36            llama_batch_allocr & balloc,
37            uint32_t n_ubatch,
38            bool embd_all) override;
39
40    llama_memory_context_ptr init_full() override;
41
42    llama_memory_context_ptr init_update(llama_context * lctx, bool optimize) override;
43
44    void clear(bool data) override;
45
46    bool seq_rm  (llama_seq_id seq_id,                              llama_pos p0, llama_pos p1) override;
47    void seq_cp  (llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) override;
48    void seq_keep(llama_seq_id seq_id)                                                          override;
49    void seq_add (llama_seq_id seq_id,                              llama_pos p0, llama_pos p1, llama_pos shift) override;
50    void seq_div (llama_seq_id seq_id,                              llama_pos p0, llama_pos p1, int d) override;
51
52    llama_pos seq_pos_min(llama_seq_id seq_id) const override;
53    llama_pos seq_pos_max(llama_seq_id seq_id) const override;
54
55    std::map<ggml_backend_buffer_type_t, size_t> memory_breakdown() const override;
56
57    bool prepare(const std::vector<llama_ubatch> & ubatches);
58
59    // find a contiguous slot of memory cells and emplace the ubatch there
60    bool find_slot(const llama_ubatch & ubatch);
61
62    bool get_can_shift() const override;
63
64    // state write/load
65
66    void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1, llama_state_seq_flags flags = 0) const override;
67    void state_read (llama_io_read_i  & io, llama_seq_id seq_id = -1, llama_state_seq_flags flags = 0) override;
68
69    uint32_t head = 0; // the location where the batch will be placed in the cache (see find_slot())
70    uint32_t size = 0; // total number of cells, shared across all sequences
71    uint32_t used = 0; // used cells (i.e. at least one seq_id)
72
73    // number of recurrent-state snapshots per seq for rollback; tensors are widened to (1 + n_rs_seq) groups
74    uint32_t n_rs_seq = 0;
75
76    // per-seq rollback index
77    std::vector<uint32_t> rs_idx;
78
79    void set_rs_idx(llama_seq_id seq_id, uint32_t idx);
80
81    // computed before each graph build
82    uint32_t n = 0;
83
84    // first zero-ed state
85    int32_t rs_z = -1;
86
87    // TODO: optimize for recurrent state needs
88    struct mem_cell {
89        llama_pos pos  = -1;
90        int32_t   src  = -1; // used to know where states should be copied from
91        int32_t   src0 = -1; // like src, but only used when setting the inputs (allowing to copy once)
92        int32_t   tail = -1;
93
94        std::set<llama_seq_id> seq_id;
95
96        bool has_seq_id(const llama_seq_id & id) const {
97            return seq_id.find(id) != seq_id.end();
98        }
99
100        bool is_empty() const {
101            return seq_id.empty();
102        }
103
104        bool is_same_seq(const mem_cell & other) const {
105            return seq_id == other.seq_id;
106        }
107    };
108
109    std::vector<mem_cell> cells;
110
111    // per layer
112    std::vector<ggml_tensor *> r_l;
113    std::vector<ggml_tensor *> s_l;
114    // a second conv history that must stay replicated across devices, so it cannot share the r row
115    std::vector<ggml_tensor *> p_l;
116
117private:
118    //const llama_model & model;
119    const llama_hparams & hparams;
120
121    const uint32_t n_seq_max = 1;
122
123    // ggml contexts for the KV cache along with the allocated backend buffers:
124    std::vector<std::pair<ggml_context_ptr, ggml_backend_buffer_ptr>> ctxs_bufs;
125
126    size_t total_size() const;
127
128    size_t size_r_bytes() const;
129    size_t size_s_bytes() const;
130    size_t size_p_bytes() const;
131
132    void state_write_meta(llama_io_write_i & io, const std::vector<std::pair<uint32_t, uint32_t>> & cell_ranges, llama_seq_id seq_id = -1) const;
133    void state_write_data(llama_io_write_i & io, const std::vector<std::pair<uint32_t, uint32_t>> & cell_ranges) const;
134
135    bool state_read_meta(llama_io_read_i & io, uint32_t cell_count, llama_seq_id dest_seq_id = -1);
136    bool state_read_data(llama_io_read_i & io, uint32_t cell_count);
137};
138
139class llama_memory_recurrent_context : public llama_memory_context_i {
140public:
141    // used for errors
142    llama_memory_recurrent_context(llama_memory_status status);
143
144    // used to create a full-cache or update context
145    llama_memory_recurrent_context(
146            llama_memory_recurrent * mem);
147
148    // used to create a batch processing context from a batch
149    llama_memory_recurrent_context(
150            llama_memory_recurrent * mem,
151            std::vector<llama_ubatch> ubatches);
152
153    virtual ~llama_memory_recurrent_context();
154
155    //
156    // llama_memory_context_i
157    //
158
159    bool next()  override;
160    bool apply() override;
161
162    llama_memory_status  get_status() const override;
163    const llama_ubatch & get_ubatch() const override;
164
165    //
166    // llama_memory_recurrent_context specific API
167    //
168
169    uint32_t get_n_rs() const;
170    uint32_t get_head() const;
171    int32_t  get_rs_z() const;
172    uint32_t get_size() const;
173
174    ggml_tensor * get_r_l(int32_t il) const;
175    ggml_tensor * get_s_l(int32_t il) const;
176    ggml_tensor * get_p_l(int32_t il) const;
177
178    int32_t s_copy(int i) const;
179
180private:
181    const llama_memory_status status;
182
183    llama_memory_recurrent * mem;
184
185    size_t i_next = 0;
186
187    std::vector<llama_ubatch> ubatches;
188
189    //
190    // data needed for building the compute graph for the current ubatch:
191    // TODO: extract all the state like `head` and `n` here
192    //
193
194    const bool is_full = false;
195};