File size: 13,179 Bytes
61b850a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
#include "llama-batch.h"

#include <cstring>
#include <algorithm>

llama_ubatch llama_sbatch::reserve_ubatch(size_t n_ubatch, bool has_embd) {
    // clear empty sequences
    // the previous ubatch is assumed to be gone,
    // so nothing should refer to values in these sequences anymore.
    for (size_t i = seq.size(); i-- > 0;) {
        if (seq[i].length == 0) {
            seq.pop_back();
        } else {
            break;
        }
    }
    ubatch_token.resize(!has_embd ? n_ubatch : 0);
    ubatch_embd.resize(has_embd ? n_embd * n_ubatch : 0);
    ubatch_pos.resize(n_ubatch);
    ubatch_n_seq_id.resize(n_ubatch);
    ubatch_seq_id.resize(n_ubatch);
    ubatch_output.resize(n_ubatch);
    llama_ubatch ubatch = {
        /*equal_seqs   =*/ true,
        /*n_tokens     =*/ 0,
        /*n_seq_tokens =*/ 0,
        /*n_seqs       =*/ 0,
        /*token        =*/ !has_embd ? ubatch_token.data() : nullptr,
        /*embd         =*/ has_embd  ? ubatch_embd.data()  : nullptr,
        /*pos          =*/ ubatch_pos.data(),
        /*n_seq_id     =*/ ubatch_n_seq_id.data(),
        /*seq_id       =*/ ubatch_seq_id.data(),
        /*output       =*/ ubatch_output.data(),
    };
    return ubatch;
}

void llama_sbatch::add_seq_to_ubatch(llama_ubatch & ubatch, llama_sbatch_seq & seq, size_t length) {
    GGML_ASSERT(batch != nullptr);
    GGML_ASSERT(length <= seq.length);
    // Can only add sequences of equal lengths to a batch,
    // otherwise it isn't clear to which sequence a token belongs
    GGML_ASSERT(seq.n_seq_id == 0 || ubatch.n_seqs == 0 || length == (size_t) ubatch.n_tokens / ubatch.n_seqs);
    GGML_ASSERT((seq.n_seq_id != 0) == ubatch.equal_seqs);
    // NOTE: loops are separated for cache-friendliness
    if (batch->token) {
        if (ubatch.equal_seqs) {
            for (size_t i = 0; i < length; ++i) {
                ubatch.token[ubatch.n_tokens + i] = batch->token[ids[seq.offset + i]];
            }
        } else {
            // simple split
            ubatch.token = batch->token + seq.offset;
        }
    } else {
        ubatch.token = nullptr;
    }
    if (batch->embd) {
        if (ubatch.equal_seqs) {
            for (size_t i = 0; i < length; ++i) {
                memcpy(
                        ubatch.embd + (n_embd * (ubatch.n_tokens + i)),
                        batch->embd + (n_embd * ids[seq.offset + i]),
                        n_embd * sizeof(float)
                      );
            }
        } else {
            // simple split
            ubatch.embd = batch->embd + (n_embd * seq.offset);
        }
    } else {
        ubatch.embd = nullptr;
    }
    if (ubatch.equal_seqs) {
        for (size_t i = 0; i < length; ++i) {
            ubatch.pos[ubatch.n_tokens + i] = batch->pos[ids[seq.offset + i]];
        }
    } else {
        // simple split
        ubatch.pos = batch->pos + seq.offset;
    }
    if (ubatch.equal_seqs) {
        ubatch.n_seq_id[ubatch.n_seqs] = seq.n_seq_id;
        if (seq.seq_id) {
            ubatch.seq_id[ubatch.n_seqs] = seq.seq_id;
        }
    } else {
        // simple split
        if (batch->n_seq_id) {
            ubatch.n_seq_id = batch->n_seq_id + seq.offset;
        } else {
            for (size_t i = 0; i < length; ++i) {
                ubatch.n_seq_id[ubatch.n_seqs + i] = 1;
            }
        }
        if (batch->seq_id) {
            ubatch.seq_id = batch->seq_id + seq.offset;
        }
    }
    if (logits_all) {
        for (size_t i = 0; i < length; ++i) {
            ubatch.output[ubatch.n_tokens + i] = 1;
            out_ids.push_back(ids[seq.offset + i]);
        }
    } else if (batch->logits) {
        if (ubatch.equal_seqs) {
            for (size_t i = 0; i < length; ++i) {
                size_t id = ids[seq.offset + i];
                int8_t is_output = batch->logits[id];
                ubatch.output[ubatch.n_tokens + i] = is_output;
                if (is_output) { out_ids.push_back(id); }
            }
        } else {
            // simple split
            ubatch.output = batch->logits + seq.offset;
            for (size_t i = 0; i < length; ++i) {
                if (ubatch.output[i] != 0) { out_ids.push_back(seq.offset + i); }
            }
        }
    } else {
        // only get last output
        for (size_t i = 0; i < length; ++i) {
            size_t id = ids[seq.offset + i];
            int8_t is_last = id == ids.size() - 1;
            ubatch.output[ubatch.n_tokens + i] = is_last;
            if (is_last) { out_ids.push_back(id); }
        }
    }
    if (ubatch.n_tokens == 0 && ubatch.n_seqs == 0) {
        ubatch.n_seq_tokens = ubatch.equal_seqs ? length : 1;
    }
    ubatch.n_tokens += length;
    ubatch.n_seqs += ubatch.equal_seqs ? 1 : length; // virtual sequences for simple splits
    seq.offset += length;
    seq.length -= length;
    n_tokens -= length;
    GGML_ASSERT(ubatch.n_tokens == ubatch.n_seq_tokens * ubatch.n_seqs);
}

llama_ubatch llama_sbatch::split_simple(size_t n_ubatch) {
    n_ubatch = n_tokens < n_ubatch ? n_tokens : n_ubatch;
    llama_ubatch ubatch = reserve_ubatch(n_ubatch, /* has_embd */ batch->embd != nullptr);
    ubatch.equal_seqs = false;
    if (!seq.empty()) {
        llama_sbatch_seq & s = seq[0];
        size_t length = s.length < n_ubatch ? s.length : n_ubatch;
        GGML_ASSERT(seq.size() == 1 && s.n_seq_id == 0); // don't mix with other splits
        add_seq_to_ubatch(ubatch, s, length);
    }
    return ubatch;
}

llama_ubatch llama_sbatch::split_equal(size_t n_ubatch) {
    n_ubatch = n_tokens < n_ubatch ? n_tokens : n_ubatch;
    llama_ubatch ubatch = reserve_ubatch(n_ubatch, /* has_embd */ batch->embd != nullptr);
    if (!seq.empty()) {
        size_t length = 0;
        size_t n_tokens_in_ubatch = 0;
        GGML_ASSERT(seq[0].n_seq_id > 0); // should not be mixed with simple splits
                                          // smallest first, because it's easier to split this way;
                                          // starting from the end to pop in constant time.
        for (size_t i = seq.size(); i-- > 0;) {
            llama_sbatch_seq & s = seq[i];
            GGML_ASSERT(s.length > 0);
            if (length == 0) {
                length = s.length < n_ubatch ? s.length : n_ubatch;
            }
            add_seq_to_ubatch(ubatch, s, length);
            n_tokens_in_ubatch += length;
            // shared prompts can't be mixed with any of their sequences,
            // so it's safer to compute them in their own ubatch
            if (s.n_seq_id > 1) { break; }
            // stop when there isn't enough space for another sequence
            if (length + n_tokens_in_ubatch > n_ubatch) { break; }
        }
    }
    return ubatch;
}

llama_ubatch llama_sbatch::split_seq(size_t n_ubatch) {
    n_ubatch = n_tokens < n_ubatch ? n_tokens : n_ubatch;
    llama_ubatch ubatch = reserve_ubatch(n_ubatch, /* has_embd */ batch->embd != nullptr);
    if (!seq.empty()) {
        llama_sbatch_seq & s = seq[seq.size() - 1];
        size_t length = s.length < n_ubatch ? s.length : n_ubatch;
        GGML_ASSERT(s.n_seq_id > 0); // should not be mixed with simple splits
        add_seq_to_ubatch(ubatch, s, length);
    }
    return ubatch;
}

void llama_sbatch::from_batch(const llama_batch & batch, size_t n_embd, bool simple_split, bool logits_all) {
    GGML_ASSERT(batch.n_tokens >= 0);
    this->batch = &batch;
    this->n_embd = n_embd;
    this->logits_all = logits_all;

    n_tokens = batch.n_tokens;
    ids.resize(n_tokens);
    out_ids.clear();
    // TODO: reserve out_ids and seq

    for (size_t i = 0; i < n_tokens; ++i) {
        ids[i] = i;
    }
    if (simple_split) {
        seq.resize(1);
        llama_sbatch_seq & s = seq[0];
        s.n_seq_id = 0;
        s.seq_id = nullptr;
        s.offset = 0;
        s.length = n_tokens;
        return;
    }
    std::sort(ids.begin(), ids.end(),
            [&batch](size_t a, size_t b) {
                int32_t n_seq_a = batch.n_seq_id ? batch.n_seq_id[a] : 1;
                int32_t n_seq_b = batch.n_seq_id ? batch.n_seq_id[b] : 1;
                // sort by seq_id, then by pos
                if (n_seq_a == n_seq_b) {
                    if (batch.seq_id) {
                        for (int32_t i = 0; i < n_seq_a; ++i) {
                            llama_seq_id seq_id_a = batch.seq_id[a][i];
                            llama_seq_id seq_id_b = batch.seq_id[b][i];
                            // smaller seq_ids go first
                            if (seq_id_a != seq_id_b) {
                                return seq_id_a < seq_id_b;
                            }
                        }
                    }
                    // when all else is equal, sort by pos
                    if (batch.pos) {
                        return batch.pos[a] < batch.pos[b];
                    }
                    // no pos, sort by id
                    return a < b;
                }
                // shared prompts go first
                return n_seq_a > n_seq_b;
            }
    );
    // init seq
    llama_sbatch_seq * last_seq = nullptr;

    for (size_t i = 0; i < n_tokens; ++i) {
        const size_t bi = ids[i];
        const int32_t n_seqs = batch.n_seq_id[bi];
        llama_seq_id * seq_ids = batch.seq_id[bi];
        if (last_seq != nullptr) {
            bool same = n_seqs == last_seq->n_seq_id;
            for (int32_t j = 0; same && j < n_seqs; ++j) {
                if (seq_ids[j] != last_seq->seq_id[j]) {
                    same = false;
                }
            }
            if (same) {
                last_seq->length += 1;
                continue;
            }
        }
        llama_sbatch_seq new_seq = {n_seqs, seq_ids, i, 1};
        seq.push_back(new_seq);
        last_seq = &seq.back();
    }
    // keep shared prompts first at the end, then sort by length descending.
    std::sort(seq.begin(), seq.end(),
            [](llama_sbatch_seq & a, llama_sbatch_seq & b) {
                if (a.n_seq_id == b.n_seq_id) {
                    return a.length > b.length;
                }
                return a.n_seq_id < b.n_seq_id;
            }
            );
}

llama_batch_allocr::llama_batch_allocr(struct llama_batch in_batch, llama_pos p0) {
    batch = in_batch;
    GGML_ASSERT(batch.n_tokens > 0);
    if (!batch.pos) {
        pos.resize(batch.n_tokens);
        for (int32_t i = 0; i < batch.n_tokens; i++) {
            pos[i] = i + p0;
        }
        batch.pos = pos.data();
    }
    if (!batch.n_seq_id) {
        n_seq_id.resize(batch.n_tokens);
        for (int32_t i = 0; i < batch.n_tokens; i++) {
            n_seq_id[i] = seq_id_0.size();
        }
        batch.n_seq_id = n_seq_id.data();
    }
    if (!batch.seq_id) {
        seq_id.resize(batch.n_tokens + 1);
        seq_id[batch.n_tokens] = NULL;
        for (int32_t i = 0; i < batch.n_tokens; i++) {
            seq_id[i] = seq_id_0.data();
        }
        batch.seq_id = seq_id.data();
    }
    if (!batch.logits) {
        logits.resize(batch.n_tokens);
        logits[logits.size() - 1] = true;
        batch.logits = logits.data();
    }
}

//
// interface implementation
//

struct llama_batch llama_batch_get_one(
             llama_token * tokens,
                 int32_t   n_tokens) {
    return {
        /*n_tokens       =*/ n_tokens,
        /*tokens         =*/ tokens,
        /*embd           =*/ nullptr,
        /*pos            =*/ nullptr,
        /*n_seq_id       =*/ nullptr,
        /*seq_id         =*/ nullptr,
        /*logits         =*/ nullptr,
    };
}

struct llama_batch llama_batch_init(int32_t n_tokens_alloc, int32_t embd, int32_t n_seq_max) {
    llama_batch batch = {
        /*n_tokens       =*/ 0,
        /*tokens         =*/ nullptr,
        /*embd           =*/ nullptr,
        /*pos            =*/ nullptr,
        /*n_seq_id       =*/ nullptr,
        /*seq_id         =*/ nullptr,
        /*logits         =*/ nullptr,
    };

    if (embd) {
        batch.embd = (float *) malloc(sizeof(float) * n_tokens_alloc * embd);
    } else {
        batch.token = (llama_token *) malloc(sizeof(llama_token) * n_tokens_alloc);
    }

    batch.pos      = (llama_pos *)     malloc(sizeof(llama_pos)      * n_tokens_alloc);
    batch.n_seq_id = (int32_t *)       malloc(sizeof(int32_t)        * n_tokens_alloc);
    batch.seq_id   = (llama_seq_id **) malloc(sizeof(llama_seq_id *) * (n_tokens_alloc + 1));
    for (int i = 0; i < n_tokens_alloc; ++i) {
        batch.seq_id[i] = (llama_seq_id *) malloc(sizeof(llama_seq_id) * n_seq_max);
    }
    batch.seq_id[n_tokens_alloc] = nullptr;

    batch.logits   = (int8_t *)        malloc(sizeof(int8_t)         * n_tokens_alloc);

    return batch;
}

void llama_batch_free(struct llama_batch batch) {
    if (batch.token)    free(batch.token);
    if (batch.embd)     free(batch.embd);
    if (batch.pos)      free(batch.pos);
    if (batch.n_seq_id) free(batch.n_seq_id);
    if (batch.seq_id) {
        for (int i = 0; batch.seq_id[i] != nullptr; ++i) {
            free(batch.seq_id[i]);
        }
        free(batch.seq_id);
    }
    if (batch.logits)   free(batch.logits);
}