Much faster rng sampling (#1187)

src/llama-sampling.cpp

@@ -702,23 +702,39 @@ llama_token llama_sample_token_with_rng_impl(struct llama_sampling * smpl, llama
     GGML_ASSERT(smpl);
 
     const int64_t t_start_sample_us = ggml_time_us();
-    llama_sample_softmax_impl((struct llama_sampling *) nullptr, candidates);
 
-    std::vector<float> probs;
-    probs.reserve(candidates->size);
-    for (size_t i = 0; i < candidates->size; ++i) {
-        probs.push_back(candidates->data[i].p);
+    if (candidates->size < 2) {
+        smpl->t_sample_us += ggml_time_us() - t_start_sample_us;
+        smpl->n_sample++;
+        return candidates->data[0].id;
     }
 
-    std::discrete_distribution<> dist(probs.begin(), probs.end());
-    int idx = dist(rng);
+    std::vector<float> probs(candidates->size);
+    probs[0] = candidates->data[0].logit;
+    float max = probs[0];
+    for (int j = 1; j < candidates->size; ++j) {
+        probs[j] = candidates->data[j].logit;
+        max = std::max(max, probs[j]);
+    }
 
-    llama_token result = candidates->data[idx].id;
+    float sump = 0;
+    for (int j = 0; j < candidates->size; ++j) {
+        float p = expf(probs[j] - max);
+        sump += p;
+        probs[j] = sump;
+    }
+    probs.back() += sump;
+
+    auto p = sump * rng() / rng.max();
+    auto iter = std::upper_bound(probs.begin(), probs.end(), p);
+    GGML_ASSERT(iter != probs.end());
+    auto idx = std::distance(probs.begin(), iter);
+    auto id  = candidates->data[idx].id;
 
     smpl->t_sample_us += ggml_time_us() - t_start_sample_us;
     smpl->n_sample++;
 
-    return result;
+    return id;
 }
 
 llama_token llama_sample_token_impl(struct llama_sampling * smpl, llama_token_data_array * candidates) {