Fix a bug in the rope calculation

convert-pth-to-ggml.py

@@ -73,7 +73,7 @@ fout.write(struct.pack("i", hparams["dim"]))
 fout.write(struct.pack("i", hparams["multiple_of"]))
 fout.write(struct.pack("i", hparams["n_heads"]))
 fout.write(struct.pack("i", hparams["n_layers"]))
-fout.write(struct.pack("i", 64)) # rot
+fout.write(struct.pack("i", hparams["dim"] // hparams["n_heads"])) # rot (obsolete)
 fout.write(struct.pack("i", ftype))
 
 # Is this correct??

main.cpp

@@ -400,7 +400,7 @@ bool llama_eval(
     const int n_ctx   = hparams.n_ctx;
     const int n_head  = hparams.n_head;
     const int n_vocab = hparams.n_vocab;
-    const int n_rot   = hparams.n_rot;
+    const int n_rot   = hparams.n_embd/hparams.n_head;
 
     const int d_key = n_embd/n_head;
 
@@ -628,6 +628,9 @@ int main(int argc, char ** argv) {
         params.prompt = gpt_random_prompt(rng);
     }
 
+//    params.prompt = R"(// this function checks if the number n is prime
+//bool is_prime(int n) {)";
+
     int64_t t_load_us = 0;
 
     gpt_vocab vocab;
@@ -691,7 +694,6 @@ int main(int argc, char ** argv) {
 
         if (i >= embd_inp.size()) {
             // sample next token
-            const int   top_k = params.top_k;
             const float top_p = params.top_p;
             const float temp  = params.temp;
 
@@ -702,7 +704,7 @@ int main(int argc, char ** argv) {
             {
                 const int64_t t_start_sample_us = ggml_time_us();
 
-                id = gpt_sample_top_k_top_p(vocab, logits.data() + (logits.size() - n_vocab), top_k, top_p, temp, rng);
+                id = llama_sample_top_p(vocab, logits.data() + (logits.size() - n_vocab), top_p, temp, rng);
 
                 t_sample_us += ggml_time_us() - t_start_sample_us;
             }

utils.cpp

@@ -257,7 +257,7 @@ std::vector<gpt_vocab::id> llama_tokenize(const gpt_vocab & vocab, const std::st
             }
         }
 
-        if (l == 0 && t != 13) {
+        if (l == 0) {
             break;
         }
 
@@ -367,6 +367,83 @@ gpt_vocab::id gpt_sample_top_k_top_p(
     return logits_id[idx].second;
 }
 
+gpt_vocab::id llama_sample_top_p(
+        const gpt_vocab & vocab,
+        const float * logits,
+        double top_p,
+        double temp,
+        std::mt19937 & rng) {
+    int n_logits = vocab.id_to_token.size();
+
+    std::vector<std::pair<double, gpt_vocab::id>> logits_id;
+    logits_id.reserve(n_logits);
+
+    {
+        const double scale = 1.0/temp;
+        for (int i = 0; i < n_logits; ++i) {
+            logits_id.push_back(std::make_pair(logits[i]*scale, i));
+        }
+    }
+
+    std::sort(
+            logits_id.begin(),
+            logits_id.end(),
+            [](const std::pair<double, gpt_vocab::id> & a, const std::pair<double, gpt_vocab::id> & b) {
+        return a.first > b.first;
+    });
+
+    double maxl = -INFINITY;
+    for (const auto & kv : logits_id) {
+        maxl = std::max(maxl, kv.first);
+    }
+
+    // compute probs for the top K tokens
+    std::vector<double> probs;
+    probs.reserve(logits_id.size());
+
+    double sum = 0.0;
+    for (const auto & kv : logits_id) {
+        double p = exp(kv.first - maxl);
+        probs.push_back(p);
+        sum += p;
+    }
+
+    // normalize the probs
+    for (auto & p : probs) {
+        p /= sum;
+    }
+
+    if (top_p < 1.0f) {
+        double cumsum = 0.0f;
+        for (int i = 0; i < (int) probs.size(); i++) {
+            cumsum += probs[i];
+            if (cumsum >= top_p) {
+                probs.resize(i + 1);
+                logits_id.resize(i + 1);
+                break;
+            }
+        }
+
+        cumsum = 1.0/cumsum;
+        for (int i = 0; i < (int) probs.size(); i++) {
+            probs[i] *= cumsum;
+        }
+    }
+
+    //printf("\n");
+    //for (int i = 0; i < (int) 10; i++) {
+    //    printf("%d: '%s' %f\n", i, vocab.id_to_token.at(logits_id[i].second).c_str(), probs[i]);
+    //}
+    //printf("\n\n");
+    //exit(0);
+
+    std::discrete_distribution<> dist(probs.begin(), probs.end());
+    int idx = dist(rng);
+
+    return logits_id[idx].second;
+}
+
+
 size_t ggml_quantize_q4_0(float * src, void * dst, int n, int k, int qk, int64_t * hist) {
     const int nb = k / qk;
     const size_t row_size = nb*(sizeof(float) + sizeof(uint8_t)*qk/2);

utils.h

@@ -18,7 +18,7 @@ struct gpt_params {
     int32_t n_predict = 128; // new tokens to predict
 
     // sampling parameters
-    int32_t top_k = 40;
+    int32_t top_k = 40; // unused
     float   top_p = 0.95f;
     float   temp  = 0.80f;
 
@@ -86,6 +86,13 @@ gpt_vocab::id gpt_sample_top_k_top_p(
         double temp,
         std::mt19937 & rng);
 
+gpt_vocab::id llama_sample_top_p(
+        const gpt_vocab & vocab,
+        const float * logits,
+        double top_p,
+        double temp,
+        std::mt19937 & rng);
+
 //
 // Quantization
 //