File size: 7,391 Bytes
31dd200
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
#include "arg.h"
#include "common.h"
#include "llama.h"

#include <vector>
#include <cstdio>


int main(int argc, char ** argv) {
    common_params params;

    params.prompt = "The quick brown fox";
    params.sampling.seed = 1234;

    const std::string_view state_file = "dump_state.bin";

    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON)) {
        return 1;
    }

    if (params.n_parallel == 1) {
        // the example uses 2 sequences, so when n_parallel == 1, we need to enable unified kv cache
        printf("%s: n_parallel == 1, enabling unified kv cache\n", __func__);
        params.kv_unified = true;
    }

    common_init();

    if (params.n_predict < 0) {
        params.n_predict = 16;
    }

    auto n_past = 0;

    std::string result0;
    std::string result1;
    std::string result2;

    // init
    auto llama_init = common_init_from_params(params);

    auto * model = llama_init->model();
    auto * ctx   = llama_init->context();

    if (model == nullptr || ctx == nullptr) {
        fprintf(stderr, "%s : failed to init\n", __func__);
        return 1;
    }

    auto sparams = llama_sampler_chain_default_params();

    llama_sampler * smpl = llama_sampler_chain_init(sparams);

    llama_sampler_chain_add(smpl, llama_sampler_init_dist(params.sampling.seed));

    // tokenize prompt
    auto tokens = common_tokenize(ctx, params.prompt, true);

    const bool save_state = true;
    if (!common_prompt_batch_decode(ctx, tokens, n_past, params.n_batch, state_file, save_state)) {
        return 1;
    }

    // first run
    printf("\nfirst run: %s", params.prompt.c_str());

    llama_batch batch = llama_batch_init(1, 0, 1);

    for (auto i = 0; i < params.n_predict; i++) {
        auto next_token     = llama_sampler_sample(smpl, ctx, -1);
        auto next_token_str = common_token_to_piece(ctx, next_token);

        printf("%s", next_token_str.c_str());
        result0 += next_token_str;

        common_batch_clear(batch);
        common_batch_add(batch, next_token, n_past, {0}, true);

        if (llama_decode(ctx, batch)) {
            fprintf(stderr, "\n%s : failed to evaluate\n", __func__);
            llama_batch_free(batch);
            return 1;
        }
        n_past += 1;
    }

    printf("\n\n");

    // make new context
    llama_context * ctx2 = llama_init_from_model(model, common_context_params_to_llama(params));

    llama_sampler * smpl2 = llama_sampler_chain_init(sparams);

    llama_sampler_chain_add(smpl2, llama_sampler_init_dist(params.sampling.seed));

    printf("\nsecond run: %s", params.prompt.c_str());

    // load state from file
    std::vector<llama_token> unused_sts(tokens.size()); // unused session tokens.
    size_t n_token_count_out = 0;

    if (!llama_state_load_file(ctx2, state_file.data(), unused_sts.data(), unused_sts.size(), &n_token_count_out)) {
        fprintf(stderr, "\n%s : failed to load state\n", __func__);
        return 1;
    }

    fprintf(stderr, "%s : loaded state with %zu tokens\n", __func__, n_token_count_out);

    // restore state (last tokens)
    n_past = n_token_count_out;
    if (!common_replay_last_token(ctx2, tokens.back(), n_past)) {
        return 1;
    }
    ++n_past;

    // second run
    for (auto i = 0; i < params.n_predict; i++) {
        auto next_token     = llama_sampler_sample(smpl2, ctx2, -1);
        auto next_token_str = common_token_to_piece(ctx2, next_token);

        printf("%s", next_token_str.c_str());
        result1 += next_token_str;

        common_batch_clear(batch);
        common_batch_add(batch, next_token, n_past, {0}, true);

        if (llama_decode(ctx2, batch)) {
            fprintf(stderr, "\n%s : failed to evaluate\n", __func__);
            llama_batch_free(batch);
            return 1;
        }
        n_past += 1;
    }

    printf("\n\n");

    if (result0 != result1) {
        fprintf(stderr, "\n%s : error : the 2 generations are different\n", __func__);
        return 1;
    }

    // make new context
    auto params_ctx3 = common_context_params_to_llama(params);
    params_ctx3.n_seq_max = 2;
    llama_context * ctx3 = llama_init_from_model(model, params_ctx3);

    llama_sampler * smpl3 = llama_sampler_chain_init(sparams);

    llama_sampler_chain_add(smpl3, llama_sampler_init_dist(params.sampling.seed));

    printf("\nsingle seq run: %s", params.prompt.c_str());

    // load state (rng, logits, embedding and kv_cache) from file
    n_token_count_out = 0;

    if (!llama_state_load_file(ctx3, state_file.data(), unused_sts.data(), unused_sts.size(), &n_token_count_out)) {
        fprintf(stderr, "\n%s : failed to load state\n", __func__);
        return 1;
    }

    fprintf(stderr, "%s : loaded state with %zu tokens\n", __func__, n_token_count_out);

    // restore state (last tokens)
    n_past = n_token_count_out;
    if (!common_replay_last_token(ctx3, tokens.back(), n_past)) {
        return 1;
    }
    ++n_past;

    // save seq 0 and load into seq 1
    {
        // save kv of seq 0
        std::vector<uint8_t> seq_store(llama_state_seq_get_size(ctx3, 0));
        const size_t ncopy = llama_state_seq_get_data(ctx3, seq_store.data(), seq_store.size(), 0);
        if (ncopy != seq_store.size()) {
            fprintf(stderr, "\n%s : seq copy data length %zd does not match expected length %zd\n", __func__, ncopy, seq_store.size());
            return 1;
        }
        fprintf(stderr, "%s : seq 0 copied, %zd bytes\n", __func__, ncopy);

        // erase whole kv
        llama_memory_clear(llama_get_memory(ctx3), true);
        fprintf(stderr, "%s : kv cache cleared\n", __func__);

        // restore kv into seq 1
        const size_t nset = llama_state_seq_set_data(ctx3, seq_store.data(), seq_store.size(), 1);
        if (nset != seq_store.size()) {
            fprintf(stderr, "\n%s : seq set data length %zd does not match expected length %zd\n", __func__, nset, seq_store.size());
            return 1;
        }
        fprintf(stderr, "%s : seq 1 restored, %zd bytes\n", __func__, nset);
    }

    // third run with seq 1 instead of 0
    for (auto i = 0; i < params.n_predict; i++) {
        auto next_token     = llama_sampler_sample(smpl3, ctx3, -1);
        auto next_token_str = common_token_to_piece(ctx3, next_token);

        printf("%s", next_token_str.c_str());
        result2 += next_token_str;

        common_batch_clear(batch);
        common_batch_add(batch, next_token, n_past, {1}, true);

        if (llama_decode(ctx3, batch)) {
            fprintf(stderr, "\n%s : failed to evaluate\n", __func__);
            llama_batch_free(batch);
            return 1;
        }
        n_past += 1;
    }

    printf("\n");

    llama_sampler_free(smpl);
    llama_sampler_free(smpl2);
    llama_sampler_free(smpl3);

    llama_batch_free(batch);

    // this one is managed by common_init_result
    //llama_free(ctx);

    llama_free(ctx2);
    llama_free(ctx3);

    if (result0 != result2) {
        fprintf(stderr, "\n%s : error : the seq restore generation is different\n", __func__);
        return 1;
    }

    fprintf(stderr, "\n%s : success\n", __func__);

    return 0;
}