Merge branch 'master' into batch_perplexity

ggml-org · ggerganov · Apr 13, 2023 · Mar 22, 2023 · Mar 24, 2023 · Mar 24, 2023
commit a17e745b6e2169daa394613bf2dbad1cee79a2bf
diff --git a/examples/perplexity/perplexity.cpp b/examples/perplexity/perplexity.cpp
@@ -29,6 +29,7 @@ void perplexity(llama_context * ctx, const gpt_params & params) {
     int seq_count = tokens.size() / params.n_ctx;
     int n_vocab = llama_n_vocab(ctx);
 
+    double nll = 0.0;
     fprintf(stderr, "%s : calculating perplexity over %d chunks, batch_size=%d\n", __func__, seq_count, params.n_batch);
 
     for (int i = 0; i < seq_count; ++i) {
@@ -71,7 +72,7 @@ void perplexity(llama_context * ctx, const gpt_params & params) {
             std::vector<float> tok_logits(
                 logits.begin() + j * n_vocab,
                 logits.begin() + (j + 1) * n_vocab);
-            double prob = softmax(tok_logits)[tokens[start + j + 1]];
+            float prob = softmax(tok_logits)[tokens[start + j + 1]];
             nll += -std::log(prob);
             ++count;
         }