1#include "ggml.h"
  2#include "llama.h"
  3
  4#ifdef NDEBUG
  5#undef NDEBUG
  6#endif
  7
  8#include <algorithm>
  9#include <cmath>
 10#include <string>
 11#include <vector>
 12
 13extern struct llama_sampler * llama_sampler_init_dry_testing(int32_t context_size, float dry_multiplier, float dry_base, int32_t dry_allowed_length, int32_t dry_penalty_last_n, const std::vector<std::vector<llama_token>>& seq_breakers);
 14
 15static void dump(const llama_token_data_array * cur_p) {
 16    for (size_t i = 0; i < cur_p->size; i++) {
 17        printf("%d: %f (%f)\n", cur_p->data[i].id, cur_p->data[i].p, cur_p->data[i].logit);
 18    }
 19}
 20
 21#define DUMP(__cur_p) do { printf("%s:%d (%s)\n", __FILE__, __LINE__, __func__); dump((__cur_p)); printf("-\n"); } while(0)
 22
 23struct sampler_tester {
 24    sampler_tester(size_t n_vocab) {
 25        cur.reserve(n_vocab);
 26        for (llama_token token_id = 0; token_id < (llama_token)n_vocab; token_id++) {
 27            const float logit = logf(token_id);
 28            cur.emplace_back(llama_token_data{token_id, logit, 0.0f});
 29        }
 30
 31        cur_p = llama_token_data_array { cur.data(), cur.size(), -1, false };
 32    }
 33
 34    sampler_tester(const std::vector<float> & probs, const std::vector<float> & probs_expected) : probs_expected(probs_expected) {
 35        cur.reserve(probs.size());
 36        for (llama_token token_id = 0; token_id < (llama_token)probs.size(); token_id++) {
 37            const float logit = logf(probs[token_id]);
 38            cur.emplace_back(llama_token_data{token_id, logit, probs[token_id]});
 39        }
 40
 41        cur_p = llama_token_data_array { cur.data(), cur.size(), -1, false };
 42    }
 43
 44    void apply(llama_sampler * sampler) {
 45        llama_sampler_apply(sampler, &cur_p);
 46        llama_sampler_free(sampler);
 47    }
 48
 49    void check() {
 50        GGML_ASSERT(cur_p.size == probs_expected.size());
 51        for (size_t i = 0; i < cur_p.size; i++) {
 52            GGML_ASSERT(fabs(cur_p.data[i].p - probs_expected[i]) < 1e-5);
 53        }
 54    }
 55
 56    llama_token_data_array cur_p;
 57
 58private:
 59    const std::vector<float> probs_expected;
 60
 61    std::vector<llama_token_data> cur;
 62};
 63
 64static void test_temp(const std::vector<float> & probs, const std::vector<float> & probs_expected, float temp) {
 65    sampler_tester tester(probs, probs_expected);
 66
 67    DUMP(&tester.cur_p);
 68    tester.apply(llama_sampler_init_temp(temp));
 69    tester.apply(llama_sampler_init_dist(0));
 70    DUMP(&tester.cur_p);
 71
 72    tester.check();
 73}
 74
 75static void test_temp_ext(const std::vector<float> & probs, const std::vector<float> & probs_expected, float temp, float delta, float exponent) {
 76    sampler_tester tester(probs, probs_expected);
 77
 78    DUMP(&tester.cur_p);
 79    tester.apply(llama_sampler_init_temp_ext(temp, delta, exponent));
 80    tester.apply(llama_sampler_init_dist (0));
 81    DUMP(&tester.cur_p);
 82
 83    tester.check();
 84}
 85
 86static void test_top_k(const std::vector<float> & probs, const std::vector<float> & probs_expected, int k) {
 87    sampler_tester tester(probs, probs_expected);
 88
 89    DUMP(&tester.cur_p);
 90    tester.apply(llama_sampler_init_top_k(k));
 91    tester.apply(llama_sampler_init_dist (0));
 92    DUMP(&tester.cur_p);
 93
 94    tester.check();
 95}
 96
 97static void test_top_p(const std::vector<float> & probs, const std::vector<float> & probs_expected, float p) {
 98    sampler_tester tester(probs, probs_expected);
 99
100    DUMP(&tester.cur_p);
101    tester.apply(llama_sampler_init_top_p(p, 0));
102    tester.apply(llama_sampler_init_dist (0));
103    DUMP(&tester.cur_p);
104
105    tester.check();
106}
107
108static void test_min_p(const std::vector<float> & probs, const std::vector<float> & probs_expected, float p) {
109    sampler_tester tester(probs, probs_expected);
110
111    DUMP(&tester.cur_p);
112    tester.apply(llama_sampler_init_min_p(p, 0));
113    tester.apply(llama_sampler_init_dist (0));
114    DUMP(&tester.cur_p);
115
116    tester.check();
117}
118
119static void test_xtc(const std::vector<float> & probs, const std::vector<float> & probs_expected, float p, float t) {
120    sampler_tester tester(probs, probs_expected);
121
122    DUMP(&tester.cur_p);
123    tester.apply(llama_sampler_init_xtc(p, t, 0, 0));
124    DUMP(&tester.cur_p);
125
126    tester.check();
127}
128
129static void test_typical(const std::vector<float> & probs, const std::vector<float> & probs_expected, float p) {
130    sampler_tester tester(probs, probs_expected);
131
132    DUMP(&tester.cur_p);
133    tester.apply(llama_sampler_init_typical(p, 0));
134    DUMP(&tester.cur_p);
135
136    tester.check();
137}
138
139static void test_penalties(
140    const std::vector<float> & probs, const std::vector<llama_token> & last_tokens,
141    const std::vector<float> & probs_expected, float repeat_penalty, float alpha_frequency, float alpha_presence
142) {
143    GGML_ASSERT(probs.size() == probs_expected.size());
144
145    sampler_tester tester(probs, probs_expected);
146
147    auto * sampler = llama_sampler_init_penalties(last_tokens.size(), repeat_penalty, alpha_frequency, alpha_presence);
148
149    for (size_t i = 0; i < last_tokens.size(); i++) {
150        llama_sampler_accept(sampler, last_tokens[i]);
151    }
152
153    DUMP(&tester.cur_p);
154    tester.apply(sampler);
155    tester.apply(llama_sampler_init_dist(0));
156    DUMP(&tester.cur_p);
157
158    tester.check();
159}
160
161static void test_dry(
162    const std::vector<float> & probs, const std::vector<llama_token> & last_tokens,
163    const std::vector<float> & expected_probs, float dry_multiplier, float dry_base,
164    int dry_allowed_length, int dry_penalty_last_n,
165    const std::vector<std::vector<llama_token>> & seq_breakers
166) {
167    GGML_ASSERT(probs.size() == expected_probs.size());
168
169    sampler_tester tester(probs, expected_probs);
170
171    auto * sampler = llama_sampler_init_dry_testing(1024, dry_multiplier, dry_base, dry_allowed_length, dry_penalty_last_n, seq_breakers);
172
173    for (size_t i = 0; i < last_tokens.size(); i++) {
174        llama_sampler_accept(sampler, last_tokens[i]);
175    }
176
177    DUMP(&tester.cur_p);
178    tester.apply(sampler);
179    tester.apply(llama_sampler_init_dist(0));
180    DUMP(&tester.cur_p);
181    tester.check();
182}
183
184static void test_top_n_sigma(const std::vector<float> & probs, const std::vector<float> & probs_expected, int n) {
185    sampler_tester tester(probs, probs_expected);
186
187    DUMP(&tester.cur_p);
188    tester.apply(llama_sampler_init_top_n_sigma(n));
189    tester.apply(llama_sampler_init_dist (0));
190    DUMP(&tester.cur_p);
191
192    tester.check();
193}
194
195static void test_sampler_queue(const size_t n_vocab, const std::string & samplers_sequence, const int top_k, const float top_p, const float min_p
196) {
197    sampler_tester tester(n_vocab);
198
199          llama_token min_token_id = 0;
200    const llama_token max_token_id = n_vocab - 1;
201
202    for (auto s : samplers_sequence) {
203        switch (s) {
204            case 'k': tester.apply(llama_sampler_init_top_k(top_k)); break;
205            case 'y': GGML_ABORT("typical test not implemented");
206            case 'p': tester.apply(llama_sampler_init_top_p(top_p, 1)); break;
207            case 'm': tester.apply(llama_sampler_init_min_p(min_p, 1)); break;
208            case 't': GGML_ABORT("temperature test not implemented");
209            default : GGML_ABORT("Unknown sampler");
210        }
211
212        tester.apply(llama_sampler_init_dist(0));
213
214        auto & cur_p = tester.cur_p;
215
216        const int size = cur_p.size;
217
218        if (s == 'k') {
219            const int expected_size = std::min(size, top_k);
220            min_token_id = std::max(min_token_id, (llama_token)(n_vocab - top_k));
221
222            GGML_ASSERT(size == expected_size);
223            GGML_ASSERT(cur_p.data[0].id == max_token_id);
224            GGML_ASSERT(cur_p.data[expected_size-1].id == min_token_id);
225        } else if (s == 'p') {
226            const int softmax_divisor = n_vocab * (n_vocab-1) / 2 - min_token_id * (min_token_id-1) / 2;
227            const int softmax_numerator_target = ceilf(top_p * softmax_divisor);
228
229                min_token_id  = n_vocab;
230            int expected_size = 0;
231            int cumsum        = 0;
232            do { // do-while because always at least one token is sampled
233                min_token_id--;
234                expected_size++;
235
236                cumsum += min_token_id;
237            } while (cumsum < softmax_numerator_target);
238
239            // token 0 has p == 0, need special consideration for cumsum because top_p immediately returns
240            if (min_token_id == 1) {
241                min_token_id--;
242                expected_size += 1;
243            }
244
245            GGML_ASSERT(size == expected_size);
246            GGML_ASSERT(!cur_p.sorted || cur_p.data[0].id == max_token_id);
247            GGML_ASSERT(!cur_p.sorted || cur_p.data[expected_size-1].id == min_token_id);
248        } else if (s == 'm') {
249            int expected_size = ceilf((1.0f - min_p) * n_vocab);
250            expected_size = std::max(expected_size, 1);
251            expected_size = std::min(expected_size, size);
252
253            min_token_id = floorf(min_p * n_vocab);
254            min_token_id = std::max(min_token_id, 1);
255            min_token_id = std::max(min_token_id, (llama_token)(n_vocab - size));
256            min_token_id = std::min(min_token_id, (llama_token)(n_vocab - 1));
257
258            GGML_ASSERT(size == expected_size);
259            GGML_ASSERT(!cur_p.sorted || cur_p.data[0].id == max_token_id);
260            GGML_ASSERT(!cur_p.sorted || cur_p.data[expected_size-1].id == min_token_id);
261        } else {
262            GGML_ABORT("fatal error");
263        }
264    }
265
266    printf("Sampler queue %3s OK with n_vocab=%05zu top_k=%5d top_p=%f min_p=%f\n",
267           samplers_sequence.c_str(), n_vocab, top_k, top_p, min_p);
268}
269
270static void bench(llama_sampler * cnstr, const char * cnstr_name, const std::vector<llama_token_data> & data, int n_iter) {
271    std::vector<llama_token_data> cur(data.size());
272    std::copy(data.begin(), data.end(), cur.begin());
273    llama_token_data_array cur_p = { cur.data(), cur.size(), -1, false };
274    llama_sampler_apply(cnstr, &cur_p);
275    llama_sampler_reset(cnstr);
276    const int64_t t_start = ggml_time_us();
277    for (int i = 0; i < n_iter; i++) {
278        std::copy(data.begin(), data.end(), cur.begin());
279        llama_token_data_array cur_p = { cur.data(), cur.size(), -1, false };
280        llama_sampler_apply(cnstr, &cur_p);
281        llama_sampler_reset(cnstr);
282    }
283    const int64_t t_end = ggml_time_us();
284    llama_sampler_free(cnstr);
285    printf("%-43s: %8.3f us/iter\n", cnstr_name, (t_end - t_start) / (float)n_iter);
286}
287
288#define BENCH(__cnstr, __data, __n_iter) bench((__cnstr), #__cnstr, (__data), (__n_iter))
289
290static void test_perf() {
291    const int n_vocab = 1 << 17;
292
293    std::vector<llama_token_data> data;
294
295    data.reserve(n_vocab);
296    for (int i = 0; i < n_vocab; i++) {
297        const float logit = 2.0f*((double)(rand())/RAND_MAX - 0.5);
298        data.emplace_back(llama_token_data{i, logit, 0.0f});
299    }
300
301    BENCH(llama_sampler_init_top_k  (40),                     data, 32);
302    BENCH(llama_sampler_init_top_p  (0.8f, 1),                data, 32);
303    BENCH(llama_sampler_init_min_p  (0.2f, 1),                data, 32);
304    BENCH(llama_sampler_init_typical(0.5f, 1),                data, 32);
305    BENCH(llama_sampler_init_xtc    (1.0f, 0.1f, 1, 1),       data, 32);
306}
307
308int main(void) {
309    ggml_time_init();
310
311    test_temp({0.1f, 0.2f, 0.3f, 0.4f}, {0.1f, 0.2f, 0.3f, 0.4f}, 1.0f);
312    test_temp({0.1f, 0.2f, 0.3f, 0.4f}, {0.0f, 0.0f, 0.0f, 1.0f}, 0.0f);
313
314    test_temp_ext({0.1f, 0.2f, 0.3f, 0.4f}, {0.1f, 0.2f, 0.3f, 0.4f}, 1.0f, 0.0f, 1.0f);
315    test_temp_ext({0.1f, 0.2f, 0.3f, 0.4f}, {0.0f, 0.0f, 0.0f, 1.0f}, 0.0f, 0.0f, 1.0f);
316
317    test_top_k({0.1f, 0.2f, 0.3f, 0.4f}, {1.0f}, 1);
318    test_top_k({0.1f, 0.2f, 0.3f, 0.4f}, {0.44444f, 0.33333f, 0.22222f}, 3);
319    test_top_k({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f, 0.3f, 0.2f, 0.1f}, 4);
320    test_top_k({0.1f, 0.2f, 0.3f, 0.4f}, {0.1f, 0.2f, 0.3f, 0.4f}, 0);
321
322    test_top_p({0.1f, 0.2f, 0.3f, 0.4f}, {1.0f}, 0);
323    test_top_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.571429f, 0.428571f}, 0.7f);
324    test_top_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.44444f, 0.33333f, 0.22222f}, 0.8f);
325    test_top_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.1f, 0.2f, 0.3f, 0.4f}, 1.0f);
326
327    test_min_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.1f/1.0f, 0.2f/1.0f, 0.3f/1.0f, 0.4f/1.0f}, 0.00f);
328    test_min_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.1f/1.0f, 0.2f/1.0f, 0.3f/1.0f, 0.4f/1.0f}, 0.24f);
329    test_min_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.2f/0.9f, 0.3f/0.9f, 0.4f/0.9f},            0.26f);
330    test_min_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.2f/0.9f, 0.3f/0.9f, 0.4f/0.9f},            0.49f);
331    test_min_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.3f/0.7f, 0.4f/0.7f},                       0.51f);
332    test_min_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.3f/0.7f, 0.4f/0.7f},                       0.74f);
333    test_min_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f/0.4f},                                  0.76f);
334    test_min_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f/0.4f},                                  1.00f);
335    test_min_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f/0.4f},                                  1.05f);
336
337    printf("XTC should:\n");
338    test_xtc({0.4f, 0.3f, 0.2f, 0.1f},   {0.1f},                                0.99f, 0.09f);
339    test_xtc({0.4f, 0.3f, 0.2f, 0.1f},   {0.2f, 0.1f},                          0.99f, 0.19f);
340    test_xtc({0.4f, 0.3f, 0.2f, 0.1f},   {0.3f, 0.2f, 0.1f},                    0.99f, 0.29f);
341
342    printf("XTC should not:\n");
343    test_xtc({0.4f, 0.3f, 0.2f, 0.1f},   {0.4f, 0.3f, 0.2f, 0.1f},              0.99f, 0.39f);
344
345    test_typical({0.97f, 0.01f, 0.01f, 0.01f}, {0.97f},            0.5f);
346    test_typical({0.4f, 0.2f, 0.2f, 0.2f},     {0.2f, 0.2f, 0.2f}, 0.5f);
347
348    test_penalties({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0}, {0, 0.25f, 0.25f, 0.25f, 0.25f},   50.0f, 0.0f, 0.0f);
349    test_penalties({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 2}, {0, 0, 0, 0.5f, 0.5f},       50.0f, 0.0f, 0.0f);
350    test_penalties({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 2, 0, 0}, {0, 0, 0, 0.5f, 0.5f}, 50.0f, 0.0f, 0.0f);
351
352    test_penalties({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0},             {0.000011f, 0.249997f, 0.249997f, 0.249997f, 0.249997f}, 1.0f, 5.0f, 5.0f);
353    test_penalties({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 2},       {0.000023f, 0.000023f, 0.000023f, 0.499966f, 0.499966f}, 1.0f, 5.0f, 5.0f);
354    test_penalties({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 2, 0, 0}, {0.000000f, 0.000023f, 0.000023f, 0.499977f, 0.499977f}, 1.0f, 5.0f, 5.0f);
355
356
357    test_dry({0.25f, 0.25f, 0.25f, 0.25f}, {0, 1}, {0.25f, 0.25f, 0.25f, 0.25f}, 1.0f, 1.1f, 2, 4, {});
358    test_dry({0.25f, 0.25f, 0.25f, 0.25f}, {0, 1, 2, 0, 1}, {0.296923f, 0.296923f, 0.109232f, 0.296923f}, 1.0f, 1.1f, 2, 5, {});
359    test_dry({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 3, 4, 0, 1}, {0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, 1.0f, 1.1f, 2, 6, {{3}});
360    test_dry({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 2, 0, 1}, {0.241818f, 0.241818f, 0.032727f, 0.241818f, 0.241818f}, 2.0f, 1.1f, 2, 5, {});
361    test_dry({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 2, 3, 4, 0, 1}, {0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, 1.0f, 1.1f, 4, 7, {});
362
363    test_top_n_sigma({0.1f, 0.2f, 0.3f, 0.4f}, {0.571429f, 0.428571f, 0.0f, 0.0f}, 1.00f);
364    test_top_n_sigma({0.1f, 0.2f, 0.3f, 0.4f}, {0.1f, 0.2f, 0.3f, 0.4f}, 0.00f); // top_n_sigma == 0 now represents a no-op rather than greedy decoding as of PR#13345
365    test_top_n_sigma({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f, 0.3f, 0.2f, 0.1f}, 3.00f);
366
367    test_sampler_queue(10000, "k", 10000, 1.0f, 1.0f);
368    test_sampler_queue(10000, "k",     1, 1.0f, 1.0f);
369    test_sampler_queue(10000, "p", 10000, 1.0f, 1.0f);
370    test_sampler_queue(10000, "p", 10000, 0.0f, 1.0f);
371    test_sampler_queue(10000, "m", 10000, 1.0f, 1.0f);
372    test_sampler_queue(10000, "m", 10000, 1.0f, 1e-12);
373
374    test_sampler_queue(10000, "k",   100, 1.0000f, 1.0f);
375    test_sampler_queue(10000, "p", 10000, 0.0003f, 1.0f);
376    test_sampler_queue(10000, "p", 10000, 0.8000f, 1.0f);
377    test_sampler_queue(10000, "m", 10000, 1.0000f, 9997.9f/9999.0f);
378    test_sampler_queue(10000, "m", 10000, 1.0000f, 0.1f);
379
380    test_sampler_queue(10000, "kp", 100, 0.8f, 0.1f);
381    test_sampler_queue(10000, "km", 100, 0.8f, 0.1f);
382    test_sampler_queue(10000, "pk", 100, 0.8f, 0.1f);
383    test_sampler_queue(10000, "pm", 100, 0.8f, 0.1f);
384    test_sampler_queue(10000, "mk", 100, 0.8f, 0.1f);
385    test_sampler_queue(10000, "mp", 100, 0.8f, 9997.9f/9999.0f);
386    test_sampler_queue(10000, "mp", 100, 0.8f, 0.1f);
387
388    test_sampler_queue(10000, "kpm", 100, 0.8f, 0.1f);
389    test_sampler_queue(10000, "kmp", 100, 0.8f, 0.1f);
390    test_sampler_queue(10000, "pkm", 100, 0.8f, 0.1f);
391    test_sampler_queue(10000, "pmk", 100, 0.8f, 0.1f);
392    test_sampler_queue(10000, "mkp", 100, 0.8f, 0.1f);
393    test_sampler_queue(10000, "mpk", 100, 0.8f, 0.1f);
394
395    printf("OK\n");
396
397    test_perf();
398
399    return 0;
400}