@fugood/llama.node 0.3.1 → 0.3.3

package/src/llama.cpp/tests/test-sampling.cpp

@@ -10,181 +10,208 @@
 #include <string>
 #include <vector>
 
-static void dump(const llama_token_data_array * candidates) {
-    for (size_t i = 0; i < candidates->size; i++) {
-        printf("%d: %f (%f)\n", candidates->data[i].id, candidates->data[i].p, candidates->data[i].logit);
+extern 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);
+
+static void dump(const llama_token_data_array * cur_p) {
+    for (size_t i = 0; i < cur_p->size; i++) {
+        printf("%d: %f (%f)\n", cur_p->data[i].id, cur_p->data[i].p, cur_p->data[i].logit);
     }
 }
 
-#define DUMP(__candidates) do { printf("%s:%d (%s)\n", __FILE__, __LINE__, __func__); dump((__candidates)); printf("-\n"); } while(0)
+#define DUMP(__cur_p) do { printf("%s:%d (%s)\n", __FILE__, __LINE__, __func__); dump((__cur_p)); printf("-\n"); } while(0)
 
-static void test_top_k(const std::vector<float> & probs, const std::vector<float> & expected_probs, int k) {
-    const size_t n_vocab = probs.size();
-    std::vector<llama_token_data> candidates;
-    candidates.reserve(n_vocab);
-    for (llama_token token_id = 0; token_id < (llama_token)n_vocab; token_id++) {
-        const float logit = logf(probs[token_id]);
-        candidates.emplace_back(llama_token_data{token_id, logit, 0.0f});
+struct sampler_tester {
+    sampler_tester(size_t n_vocab) {
+        cur.reserve(n_vocab);
+        for (llama_token token_id = 0; token_id < (llama_token)n_vocab; token_id++) {
+            const float logit = logf(token_id);
+            cur.emplace_back(llama_token_data{token_id, logit, 0.0f});
+        }
+
+        cur_p = llama_token_data_array { cur.data(), cur.size(), -1, false };
     }
 
-    llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
-    llama_sample_softmax(nullptr, &candidates_p);
-    DUMP(&candidates_p);
-    llama_sample_top_k(nullptr, &candidates_p, k, 1);
-    DUMP(&candidates_p);
+    sampler_tester(const std::vector<float> & probs, const std::vector<float> & probs_expected) : probs_expected(probs_expected) {
+        cur.reserve(probs.size());
+        for (llama_token token_id = 0; token_id < (llama_token)probs.size(); token_id++) {
+            const float logit = logf(probs[token_id]);
+            cur.emplace_back(llama_token_data{token_id, logit, probs[token_id]});
+        }
 
-    GGML_ASSERT(candidates_p.size == expected_probs.size());
-    for (size_t i = 0; i < candidates_p.size; i++) {
-        GGML_ASSERT(fabs(candidates_p.data[i].p - expected_probs[i]) < 1e-5);
+        cur_p = llama_token_data_array { cur.data(), cur.size(), -1, false };
     }
-}
 
-static void test_top_p(const std::vector<float> & probs, const std::vector<float> & expected_probs, float p) {
-    const size_t n_vocab = probs.size();
-    std::vector<llama_token_data> candidates;
-    candidates.reserve(n_vocab);
-    for (llama_token token_id = 0; token_id < (llama_token)n_vocab; token_id++) {
-        const float logit = logf(probs[token_id]);
-        candidates.emplace_back(llama_token_data{token_id, logit, 0.0f});
+    void apply(llama_sampler * sampler) {
+        llama_sampler_apply(sampler, &cur_p);
+        llama_sampler_free(sampler);
     }
 
-    llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
-    llama_sample_softmax(nullptr, &candidates_p);
-    DUMP(&candidates_p);
-    llama_sample_top_p(nullptr, &candidates_p, p, 1);
-    DUMP(&candidates_p);
-
-    GGML_ASSERT(candidates_p.size == expected_probs.size());
-    for (size_t i = 0; i < candidates_p.size; i++) {
-        GGML_ASSERT(fabs(candidates_p.data[i].p - expected_probs[i]) < 1e-3);
+    void check() {
+        GGML_ASSERT(cur_p.size == probs_expected.size());
+        for (size_t i = 0; i < cur_p.size; i++) {
+            GGML_ASSERT(fabs(cur_p.data[i].p - probs_expected[i]) < 1e-5);
+        }
     }
+
+    llama_token_data_array cur_p;
+
+private:
+    const std::vector<float> probs_expected;
+
+    std::vector<llama_token_data> cur;
+};
+
+static void test_temp(const std::vector<float> & probs, const std::vector<float> & probs_expected, float temp) {
+    sampler_tester tester(probs, probs_expected);
+
+    DUMP(&tester.cur_p);
+    tester.apply(llama_sampler_init_temp(temp));
+    tester.apply(llama_sampler_init_dist(0));
+    DUMP(&tester.cur_p);
+
+    tester.check();
 }
 
-static void test_tfs(const std::vector<float> & probs, const std::vector<float> & expected_probs, float z) {
-    const size_t n_vocab = probs.size();
-    std::vector<llama_token_data> candidates;
-    candidates.reserve(n_vocab);
-    for (llama_token token_id = 0; token_id < (llama_token)n_vocab; token_id++) {
-        const float logit = logf(probs[token_id]);
-        candidates.emplace_back(llama_token_data{token_id, logit, 0.0f});
-    }
+static void test_temp_ext(const std::vector<float> & probs, const std::vector<float> & probs_expected, float temp, float delta, float exponent) {
+    sampler_tester tester(probs, probs_expected);
 
-    llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
-    DUMP(&candidates_p);
-    llama_sample_tail_free(nullptr, &candidates_p, z, 1);
-    DUMP(&candidates_p);
+    DUMP(&tester.cur_p);
+    tester.apply(llama_sampler_init_temp_ext(temp, delta, exponent));
+    tester.apply(llama_sampler_init_dist (0));
+    DUMP(&tester.cur_p);
 
-    GGML_ASSERT(candidates_p.size == expected_probs.size());
-    for (size_t i = 0; i < candidates_p.size; i++) {
-        GGML_ASSERT(fabs(candidates_p.data[i].p - expected_probs[i]) < 1e-3);
-    }
+    tester.check();
 }
 
-static void test_min_p(const std::vector<float> & probs, const std::vector<float> & expected_probs, float p) {
-    const size_t n_vocab = probs.size();
-    std::vector<llama_token_data> candidates;
-    candidates.reserve(n_vocab);
-    for (llama_token token_id = 0; token_id < (llama_token)n_vocab; token_id++) {
-        const float logit = logf(probs[token_id]);
-        candidates.emplace_back(llama_token_data{token_id, logit, 0.0f});
-    }
+static void test_top_k(const std::vector<float> & probs, const std::vector<float> & probs_expected, int k) {
+    sampler_tester tester(probs, probs_expected);
 
-    llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
-    DUMP(&candidates_p);
-    llama_sample_min_p(nullptr, &candidates_p, p, 1);
-    DUMP(&candidates_p);
-    llama_sample_softmax(nullptr, &candidates_p);
+    DUMP(&tester.cur_p);
+    tester.apply(llama_sampler_init_top_k(k));
+    tester.apply(llama_sampler_init_dist (0));
+    DUMP(&tester.cur_p);
 
-    GGML_ASSERT(candidates_p.size == expected_probs.size());
-    for (size_t i = 0; i < candidates_p.size; i++) {
-        GGML_ASSERT(fabs(candidates_p.data[i].p - expected_probs[i]) < 1e-3);
-    }
+    tester.check();
 }
 
-static void test_typical(const std::vector<float> & probs, const std::vector<float> & expected_probs, float p) {
-    const size_t n_vocab = probs.size();
-    std::vector<llama_token_data> candidates;
-    candidates.reserve(n_vocab);
-    for (llama_token token_id = 0; token_id < (llama_token)n_vocab; token_id++) {
-        const float logit = logf(probs[token_id]);
-        candidates.emplace_back(llama_token_data{token_id, logit, 0.0f});
-    }
+static void test_top_p(const std::vector<float> & probs, const std::vector<float> & probs_expected, float p) {
+    sampler_tester tester(probs, probs_expected);
 
-    llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
-    DUMP(&candidates_p);
-    llama_sample_typical(nullptr, &candidates_p, p, 1);
-    DUMP(&candidates_p);
+    DUMP(&tester.cur_p);
+    tester.apply(llama_sampler_init_top_p(p, 1));
+    tester.apply(llama_sampler_init_dist (0));
+    DUMP(&tester.cur_p);
 
-    GGML_ASSERT(candidates_p.size == expected_probs.size());
-    for (size_t i = 0; i < candidates_p.size; i++) {
-        GGML_ASSERT(fabs(candidates_p.data[i].p - expected_probs[i]) < 1e-3);
-    }
+    tester.check();
+}
+
+static void test_min_p(const std::vector<float> & probs, const std::vector<float> & probs_expected, float p) {
+    sampler_tester tester(probs, probs_expected);
+
+    DUMP(&tester.cur_p);
+    tester.apply(llama_sampler_init_min_p(p, 1));
+    tester.apply(llama_sampler_init_dist (0));
+    DUMP(&tester.cur_p);
+
+    tester.check();
+}
+
+static void test_xtc(const std::vector<float> & probs, const std::vector<float> & probs_expected, float p, float t) {
+    sampler_tester tester(probs, probs_expected);
+
+    DUMP(&tester.cur_p);
+    tester.apply(llama_sampler_init_xtc(p, t, 0, 0));
+    DUMP(&tester.cur_p);
+
+    tester.check();
 }
 
-static void test_repetition_penalties(
+static void test_typical(const std::vector<float> & probs, const std::vector<float> & probs_expected, float p) {
+    sampler_tester tester(probs, probs_expected);
+
+    DUMP(&tester.cur_p);
+    tester.apply(llama_sampler_init_typical(p, 1));
+    DUMP(&tester.cur_p);
+
+    tester.check();
+}
+
+static void test_penalties(
     const std::vector<float> & probs, const std::vector<llama_token> & last_tokens,
-    const std::vector<float> & expected_probs, float repeat_penalty, float alpha_frequency, float alpha_presence
+    const std::vector<float> & probs_expected, float repeat_penalty, float alpha_frequency, float alpha_presence
 ) {
-    GGML_ASSERT(probs.size() == expected_probs.size());
+    GGML_ASSERT(probs.size() == probs_expected.size());
+
+    sampler_tester tester(probs, probs_expected);
 
     const size_t n_vocab = probs.size();
-    std::vector<llama_token_data> candidates;
-    candidates.reserve(n_vocab);
-    for (llama_token token_id = 0; token_id < (llama_token)n_vocab; token_id++) {
-        const float logit = logf(probs[token_id]);
-        candidates.emplace_back(llama_token_data{token_id, logit, 0.0f});
+    auto * sampler = llama_sampler_init_penalties(n_vocab, LLAMA_TOKEN_NULL, LLAMA_TOKEN_NULL, last_tokens.size(), repeat_penalty, alpha_frequency, alpha_presence, false, false);
+
+    for (size_t i = 0; i < last_tokens.size(); i++) {
+        llama_sampler_accept(sampler, last_tokens[i]);
     }
 
-    llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
-    llama_sample_softmax(nullptr, &candidates_p);
-    DUMP(&candidates_p);
-    llama_sample_repetition_penalties(nullptr, &candidates_p, (const llama_token *) last_tokens.data(), last_tokens.size(), repeat_penalty, alpha_frequency, alpha_presence);
-    llama_sample_softmax(nullptr, &candidates_p);
-    DUMP(&candidates_p);
+    DUMP(&tester.cur_p);
+    tester.apply(sampler);
+    tester.apply(llama_sampler_init_dist(0));
+    DUMP(&tester.cur_p);
 
-    GGML_ASSERT(candidates_p.size == expected_probs.size());
-    for (size_t i = 0; i < candidates_p.size; i++) {
-        GGML_ASSERT(fabs(candidates_p.data[i].p - expected_probs[i]) < 1e-3);
-    }
+    tester.check();
 }
 
-static 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
+static void test_dry(
+    const std::vector<float> & probs, const std::vector<llama_token> & last_tokens,
+    const std::vector<float> & expected_probs, float dry_multiplier, float dry_base,
+    int dry_allowed_length, int dry_penalty_last_n,
+    const std::vector<std::vector<llama_token>> & seq_breakers
 ) {
-    std::vector<llama_token_data> candidates;
-    candidates.reserve(n_vocab);
-    for (llama_token token_id = 0; token_id < (llama_token)n_vocab; token_id++) {
-        const float logit = logf(token_id);
-        candidates.emplace_back(llama_token_data{token_id, logit, 0.0f});
+    GGML_ASSERT(probs.size() == expected_probs.size());
+
+    sampler_tester tester(probs, expected_probs);
+
+    auto * sampler = llama_sampler_init_dry_testing(1024, dry_multiplier, dry_base, dry_allowed_length, dry_penalty_last_n, seq_breakers);
+
+    for (size_t i = 0; i < last_tokens.size(); i++) {
+        llama_sampler_accept(sampler, last_tokens[i]);
     }
 
-    llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
+    DUMP(&tester.cur_p);
+    tester.apply(sampler);
+    tester.apply(llama_sampler_init_dist(0));
+    DUMP(&tester.cur_p);
+    tester.check();
+}
+
+static 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
+) {
+    sampler_tester tester(n_vocab);
 
           llama_token min_token_id = 0;
     const llama_token max_token_id = n_vocab-1;
 
     for (auto s : samplers_sequence) {
         switch (s){
-            case 'k': llama_sample_top_k    (nullptr, &candidates_p, top_k, 1); break;
-            case 'f': GGML_ABORT("tail_free test not implemented");   break;
-            case 'y': GGML_ABORT("typical test not implemented");     break;
-            case 'p': llama_sample_top_p    (nullptr, &candidates_p, top_p, 1); break;
-            case 'm': llama_sample_min_p    (nullptr, &candidates_p, min_p, 1); break;
-            case 't': GGML_ABORT("temperature test not implemented"); break;
-            default : GGML_ABORT("Unknown sampler");                  break;
+            case 'k': tester.apply(llama_sampler_init_top_k(top_k)); break;
+            case 'y': GGML_ABORT("typical test not implemented");
+            case 'p': tester.apply(llama_sampler_init_top_p(top_p, 1)); break;
+            case 'm': tester.apply(llama_sampler_init_min_p(min_p, 1)); break;
+            case 't': GGML_ABORT("temperature test not implemented");
+            default : GGML_ABORT("Unknown sampler");
         }
 
-        llama_sample_softmax(nullptr, &candidates_p); // make sure tokens are sorted for tests
+        tester.apply(llama_sampler_init_dist(0));
+
+        auto & cur_p = tester.cur_p;
 
-        const int size = candidates_p.size;
+        const int size = cur_p.size;
 
         if (s == 'k') {
             const int expected_size = std::min(size, top_k);
             min_token_id = std::max(min_token_id, (llama_token)(n_vocab - top_k));
 
             GGML_ASSERT(size == expected_size);
-            GGML_ASSERT(candidates_p.data[0].id == max_token_id);
-            GGML_ASSERT(candidates_p.data[expected_size-1].id == min_token_id);
+            GGML_ASSERT(cur_p.data[0].id == max_token_id);
+            GGML_ASSERT(cur_p.data[expected_size-1].id == min_token_id);
         } else if (s == 'p') {
             const int softmax_divisor = n_vocab * (n_vocab-1) / 2 - min_token_id * (min_token_id-1) / 2;
             const int softmax_numerator_target = ceilf(top_p * softmax_divisor);
@@ -206,8 +233,8 @@ static void test_sampler_queue(
             }
 
             GGML_ASSERT(size == expected_size);
-            GGML_ASSERT(candidates_p.data[0].id == max_token_id);
-            GGML_ASSERT(candidates_p.data[expected_size-1].id == min_token_id);
+            GGML_ASSERT(cur_p.data[0].id == max_token_id);
+            GGML_ASSERT(cur_p.data[expected_size-1].id == min_token_id);
         } else if (s == 'm') {
             int expected_size = ceilf((1.0f-min_p) * n_vocab);
             expected_size = std::max(expected_size, 1);
@@ -219,29 +246,73 @@ static void test_sampler_queue(
             min_token_id = std::min(min_token_id, (llama_token)(n_vocab - 1));
 
             GGML_ASSERT(size == expected_size);
-            GGML_ASSERT(candidates_p.data[0].id == max_token_id);
-            GGML_ASSERT(candidates_p.data[expected_size-1].id == min_token_id);
+            GGML_ASSERT(cur_p.data[0].id == max_token_id);
+            GGML_ASSERT(cur_p.data[expected_size-1].id == min_token_id);
         } else {
             GGML_ABORT("fatal error");
         }
     }
 
-    printf("Sampler queue %3s OK with n_vocab=%05ld top_k=%05d top_p=%f min_p=%f\n",
+    printf("Sampler queue %3s OK with n_vocab=%05zu top_k=%05d top_p=%f min_p=%f\n",
            samplers_sequence.c_str(), n_vocab, top_k, top_p, min_p);
 }
 
+static void bench(llama_sampler * cnstr, const char * cnstr_name, const std::vector<llama_token_data> & data, int n_iter) {
+    std::vector<llama_token_data> cur(data.size());
+    std::copy(data.begin(), data.end(), cur.begin());
+    llama_token_data_array cur_p = { cur.data(), cur.size(), -1, false };
+    llama_sampler_apply(cnstr, &cur_p);
+    llama_sampler_reset(cnstr);
+    const int64_t t_start = ggml_time_us();
+    for (int i = 0; i < n_iter; i++) {
+        std::copy(data.begin(), data.end(), cur.begin());
+        llama_token_data_array cur_p = { cur.data(), cur.size(), -1, false };
+        llama_sampler_apply(cnstr, &cur_p);
+        llama_sampler_reset(cnstr);
+    }
+    const int64_t t_end = ggml_time_us();
+    llama_sampler_free(cnstr);
+    printf("%-43s: %8.3f us/iter\n", cnstr_name, (t_end - t_start) / (float)n_iter);
+}
+
+#define BENCH(__cnstr, __data, __n_iter) bench((__cnstr), #__cnstr, (__data), (__n_iter))
+
+static void test_perf() {
+    const int n_vocab = 1 << 17;
+
+    std::vector<llama_token_data> data;
+
+    data.reserve(n_vocab);
+    for (int i = 0; i < n_vocab; i++) {
+        const float logit = 2.0f*((float)(rand())/RAND_MAX - 0.5f);
+        data.emplace_back(llama_token_data{i, logit, 0.0f});
+    }
+
+    BENCH(llama_sampler_init_top_k  (40),                     data, 32);
+    BENCH(llama_sampler_init_top_p  (0.8f, 1),                data, 32);
+    BENCH(llama_sampler_init_min_p  (0.2f, 1),                data, 32);
+    BENCH(llama_sampler_init_typical(0.5f, 1),                data, 32);
+    BENCH(llama_sampler_init_xtc    (1.0f, 0.1f, 1, 1),       data, 32);
+}
+
 int main(void) {
     ggml_time_init();
 
-    test_top_k({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f}, 1);
-    test_top_k({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f, 0.3f, 0.2f}, 3);
+    test_temp({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f, 0.3f, 0.2f, 0.1f}, 1.0f);
+    test_temp({0.1f, 0.2f, 0.3f, 0.4f}, {1.0f, 0.0f, 0.0f, 0.0f}, 0.0f);
+
+    test_temp_ext({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f, 0.3f, 0.2f, 0.1f}, 1.0f, 0.0f, 1.0f);
+    test_temp_ext({0.1f, 0.2f, 0.3f, 0.4f}, {1.0f, 0.0f, 0.0f, 0.0f}, 0.0f, 0.0f, 1.0f);
+
+    test_top_k({0.1f, 0.2f, 0.3f, 0.4f}, {1.0f}, 1);
+    test_top_k({0.1f, 0.2f, 0.3f, 0.4f}, {0.44444f, 0.33333f, 0.22222f}, 3);
     test_top_k({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f, 0.3f, 0.2f, 0.1f}, 4);
     test_top_k({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f, 0.3f, 0.2f, 0.1f}, 0);
 
-    test_top_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f}, 0);
-    test_top_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f, 0.3f}, 0.7f);
-    test_top_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f, 0.3f, 0.2f}, 0.8f);
-    test_top_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f, 0.3f, 0.2f, 0.1f}, 1);
+    test_top_p({0.1f, 0.2f, 0.3f, 0.4f}, {1.0f}, 0);
+    test_top_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.571429f, 0.428571f}, 0.7f);
+    test_top_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.44444f, 0.33333f, 0.22222f}, 0.8f);
+    test_top_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f, 0.3f, 0.2f, 0.1f}, 1.0f);
 
     test_min_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f/1.0f, 0.3f/1.0f, 0.2f/1.0f, 0.1f/1.0f}, 0.00f);
     test_min_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f/1.0f, 0.3f/1.0f, 0.2f/1.0f, 0.1f/1.0f}, 0.24f);
@@ -252,20 +323,31 @@ int main(void) {
     test_min_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f/0.4f},                                  0.76f);
     test_min_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f/0.4f},                                  1.00f);
 
-    test_tfs({0.1f, 0.15f, 0.2f, 0.25f, 0.3f}, {0.3f}, 0.25f);
-    test_tfs({0.1f, 0.15f, 0.2f, 0.25f, 0.3f}, {0.3f, 0.25f}, 0.75f);
-    test_tfs({0.1f, 0.15f, 0.2f, 0.25f, 0.3f}, {0.3f, 0.25f}, 0.99f);
+    printf("XTC should:\n");
+    test_xtc({0.4f, 0.3f, 0.2f, 0.1f},   {0.1f},                                0.99f, 0.09f);
+    test_xtc({0.4f, 0.3f, 0.2f, 0.1f},   {0.2f, 0.1f},                          0.99f, 0.19f);
+    test_xtc({0.4f, 0.3f, 0.2f, 0.1f},   {0.3f, 0.2f, 0.1f},                    0.99f, 0.29f);
+
+    printf("XTC should not:\n");
+    test_xtc({0.4f, 0.3f, 0.2f, 0.1f},   {0.4f, 0.3f, 0.2f, 0.1f},              0.99f, 0.39f);
 
     test_typical({0.97f, 0.01f, 0.01f, 0.01f}, {0.97f}, 0.5f);
     test_typical({0.4f, 0.2f, 0.2f, 0.2f}, {0.2f, 0.2f, 0.2f}, 0.5f);
 
-    test_repetition_penalties({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0}, {0.25f, 0.25f, 0.25f, 0.25f, 0},   50.0f, 0.0f, 0.0f);
-    test_repetition_penalties({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 2}, {0.5f, 0.5f, 0, 0, 0},       50.0f, 0.0f, 0.0f);
-    test_repetition_penalties({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 2, 0, 0}, {0.5f, 0.5f, 0, 0, 0}, 50.0f, 0.0f, 0.0f);
+    test_penalties({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0}, {0.25f, 0.25f, 0.25f, 0.25f, 0},   50.0f, 0.0f, 0.0f);
+    test_penalties({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 2}, {0.5f, 0.5f, 0, 0, 0},       50.0f, 0.0f, 0.0f);
+    test_penalties({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 2, 0, 0}, {0.5f, 0.5f, 0, 0, 0}, 50.0f, 0.0f, 0.0f);
+
+    test_penalties({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0},             {0.249997f, 0.249997f, 0.249997f, 0.249997f, 0.000011f}, 1.0f, 5.0f, 5.0f);
+    test_penalties({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 2},       {0.499966f, 0.499966f, 0.000023f, 0.000023f, 0.000023f}, 1.0f, 5.0f, 5.0f);
+    test_penalties({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 2, 0, 0}, {0.499977f, 0.499977f, 0.000023f, 0.000023f, 0.000000f}, 1.0f, 5.0f, 5.0f);
 
-    test_repetition_penalties({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0},             {0.249997f, 0.249997f, 0.249997f, 0.249997f, 0.000011f}, 1.0f, 5.0f, 5.0f);
-    test_repetition_penalties({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 2},       {0.499966f, 0.499966f, 0.000023f, 0.000023f, 0.000023f}, 1.0f, 5.0f, 5.0f);
-    test_repetition_penalties({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 2, 0, 0}, {0.499977f, 0.499977f, 0.000023f, 0.000023f, 0.000000f}, 1.0f, 5.0f, 5.0f);
+
+    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, {});
+    test_dry({0.25f, 0.25f, 0.25f, 0.25f}, {0, 1, 2, 0, 1}, {0.296923f, 0.296923f, 0.296923f, 0.109232f}, 1.0f, 1.1f, 2, 5, {});
+    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}});
+    test_dry({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 2, 0, 1}, {0.241818f, 0.241818f, 0.241818f, 0.241818f, 0.032727f}, 2.0f, 1.1f, 2, 5, {});
+    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, {});
 
     test_sampler_queue(10000, "k", 10000, 1.0f, 1.0f);
     test_sampler_queue(10000, "k",     1, 1.0f, 1.0f);
@@ -297,5 +379,7 @@ int main(void) {
 
     printf("OK\n");
 
+    test_perf();
+
     return 0;
 }

package/src/llama.cpp/tests/test-tokenizer-0.cpp

@@ -7,6 +7,7 @@
 #include <map>
 #include <vector>
 #include <fstream>
+#include <thread>
 
 //static const std::map<std::string, std::vector<llama_token>> & k_tests() {
 //    static std::map<std::string, std::vector<llama_token>> _k_tests = {
@@ -194,45 +195,64 @@ int main(int argc, char **argv) {
 
     const bool add_special = false;
 
-    for (const auto & test_kv : k_tests) {
-        const std::vector<llama_token> res = llama_tokenize(ctx, test_kv.first, add_special, false);
-
-        printf("\n");
-        printf("src: '%s'\n", test_kv.first.c_str());
-        printf("res: '%s'\n", llama_detokenize(ctx, res).c_str());
-        printf("tok: ");
-        for (const auto & tok : res) {
-            printf("%d ", tok);
-        }
-        printf("\n");
-
-        bool correct = res.size() == test_kv.second.size();
-        for (int i = 0; i < (int) res.size() && correct; ++i) {
-            if (test_kv.second[i] != res[i]) {
-                correct = false;
+    // multi-threaded tokenization
+    const int nthread = std::thread::hardware_concurrency();
+    std::vector<std::thread> threads(nthread);
+
+    for (int i = 0; i < nthread; i++) {
+        threads[i] = std::thread([&, i]() {
+            for (const auto & test_kv : k_tests) {
+                const std::vector<llama_token> res = common_tokenize(ctx, test_kv.first, add_special, false);
+
+                // here only print the result of the first thread
+                // because the other threads are running the same tests
+                if (i != 0) {
+                    continue;
+                }
+
+                printf("\n");
+                printf("src: '%s'\n", test_kv.first.c_str());
+                printf("res: '%s'\n", common_detokenize(ctx, res).c_str());
+                printf("tok: ");
+                for (const auto & tok : res) {
+                    printf("%d ", tok);
+                }
+                printf("\n");
+
+                bool correct = res.size() == test_kv.second.size();
+                for (int i = 0; i < (int) res.size() && correct; ++i) {
+                    if (test_kv.second[i] != res[i]) {
+                        correct = false;
+                    }
+                }
+
+                if (!correct) {
+                    fprintf(stderr, "%s : failed test:    '%s'\n", __func__, test_kv.first.c_str());
+                    fprintf(stderr, "%s : detokenized to: '%s' instead of '%s'\n", __func__,
+                        common_detokenize(ctx, res).c_str(),
+                        common_detokenize(ctx, test_kv.second).c_str());
+                    fprintf(stderr, "%s : expected tokens: ", __func__);
+                    for (const auto & t : test_kv.second) {
+                        fprintf(stderr, "%6d '%s', ", t, common_token_to_piece(ctx, t).c_str());
+                    }
+                    fprintf(stderr, "\n");
+                    fprintf(stderr, "%s : got tokens:      ", __func__);
+                    for (const auto & t : res) {
+                        fprintf(stderr, "%6d '%s', ", t, common_token_to_piece(ctx, t).c_str());
+                    }
+                    fprintf(stderr, "\n");
+
+                    success = false;
+                }
             }
-        }
-
-        if (!correct) {
-            fprintf(stderr, "%s : failed test:    '%s'\n", __func__, test_kv.first.c_str());
-            fprintf(stderr, "%s : detokenized to: '%s' instead of '%s'\n", __func__,
-                llama_detokenize(ctx, res).c_str(),
-                llama_detokenize(ctx, test_kv.second).c_str());
-            fprintf(stderr, "%s : expected tokens: ", __func__);
-            for (const auto & t : test_kv.second) {
-                fprintf(stderr, "%6d '%s', ", t, llama_token_to_piece(ctx, t).c_str());
-            }
-            fprintf(stderr, "\n");
-            fprintf(stderr, "%s : got tokens:      ", __func__);
-            for (const auto & t : res) {
-                fprintf(stderr, "%6d '%s', ", t, llama_token_to_piece(ctx, t).c_str());
-            }
-            fprintf(stderr, "\n");
+        });
+    }
 
-            success = false;
-        }
+    for (int i = 0; i < nthread; i++) {
+        threads[i].join();
     }
 
+    // single threaded tokenization
     if (!fname_text.empty()) {
         fprintf(stderr, "%s : tokenizing: '%s'\n", __func__, fname_text.c_str());
 
@@ -253,7 +273,7 @@ int main(int argc, char **argv) {
         {
             const auto t_start = ggml_time_us();
 
-            res = llama_tokenize(ctx, text, add_special, false);
+            res = common_tokenize(ctx, text, add_special, false);
 
             const auto t_end = ggml_time_us();
 

package/src/llama.cpp/tests/test-tokenizer-1-bpe.cpp

@@ -78,10 +78,10 @@ int main(int argc, char **argv) {
     const int n_vocab = llama_n_vocab(model);
 
     for (int i = 0; i < n_vocab; ++i) {
-        std::string str = llama_detokenize(ctx, std::vector<int>(1, i));
+        std::string str = common_detokenize(ctx, std::vector<int>(1, i));
         try {
             auto cps = unicode_cpts_from_utf8(str);
-            std::vector<llama_token> tokens = llama_tokenize(ctx, str, false, true);
+            std::vector<llama_token> tokens = common_tokenize(ctx, str, false, true);
             if (ignore_merges && tokens.size() > 1) {
                 fprintf(stderr,
                         "%s : error: token %d detokenizes to '%s'(%zu) but "
@@ -94,7 +94,7 @@ int main(int argc, char **argv) {
                 fprintf(stderr, "]\n");
                 return 2;
             }
-            std::string check = llama_detokenize(ctx, tokens);
+            std::string check = common_detokenize(ctx, tokens);
             if (check != str) {
                 fprintf(stderr, "%s : error: token %d detokenizes to '%s'(%zu) but tokenization of this detokenizes to '%s'(%zu)\n",
                     __func__, i, str.c_str(), str.length(), check.c_str(), check.length());
@@ -123,8 +123,8 @@ int main(int argc, char **argv) {
                     }
 
                     std::string str = unicode_cpt_to_utf8(cp);
-                    std::vector<llama_token> tokens = llama_tokenize(ctx, str, false);
-                    std::string check = llama_detokenize(ctx, tokens);
+                    std::vector<llama_token> tokens = common_tokenize(ctx, str, false);
+                    std::string check = common_detokenize(ctx, tokens);
                     if (cp != 9601 && str != check) {
                         fprintf(stderr, "error: codepoint 0x%x detokenizes to '%s'(%zu) instead of '%s'(%zu)\n",
                                 cp, check.c_str(), check.length(), str.c_str(), str.length());

package/src/llama.cpp/tests/test-tokenizer-1-spm.cpp

@@ -66,9 +66,9 @@ int main(int argc, char ** argv) {
     const int n_vocab = llama_n_vocab(model);
 
     for (int i = 0; i < n_vocab; ++i) {
-        std::string str = llama_detokenize(ctx, std::vector<int>(1, i), true);
-        std::vector<llama_token> tokens = llama_tokenize(ctx, str, false, true);
-        std::string check = llama_detokenize(ctx, tokens);
+        std::string str = common_detokenize(ctx, std::vector<int>(1, i), true);
+        std::vector<llama_token> tokens = common_tokenize(ctx, str, false, true);
+        std::string check = common_detokenize(ctx, tokens);
         if (check != str) {
             fprintf(stderr, "%s : error: token %d detokenizes to '%s'(%zu) but tokenization of this detokenizes to '%s'(%zu)\n",
                 __func__, i, str.c_str(), str.length(), check.c_str(), check.length());
@@ -93,8 +93,8 @@ int main(int argc, char ** argv) {
                     }
 
                     std::string str = unicode_cpt_to_utf8(cp);
-                    std::vector<llama_token> tokens = llama_tokenize(ctx, str, false, true);
-                    std::string check = llama_detokenize(ctx, tokens);
+                    std::vector<llama_token> tokens = common_tokenize(ctx, str, false, true);
+                    std::string check = common_detokenize(ctx, tokens);
                     if (cp != 9601 && str != check) {
                         fprintf(stderr, "error: codepoint 0x%x detokenizes to '%s'(%zu) instead of '%s'(%zu)\n",
                                 cp, check.c_str(), check.length(), str.c_str(), str.length());