ci : integrate with ggml-org/ci (#2250)

* ci : run ctest

ggml-ci

* ci : add open llama 3B-v2 tests

ggml-ci

* ci : disable wget progress output

ggml-ci

* ci : add open llama 3B-v2 tg tests for q4 and q5 quantizations

ggml-ci

* tests : try to fix tail free sampling test

ggml-ci

* ci : add K-quants

ggml-ci

* ci : add short perplexity tests

ggml-ci

* ci : add README.md

* ppl : add --chunks argument to limit max number of chunks

ggml-ci

* ci : update README

1 files changed, 12 insertions, 3 deletions

diff --git a/llama.cpp b/llama.cpp
index 0f9d534..fa3b7c0 100644
--- a/llama.cpp
+++ b/llama.cpp
@@ -2024,9 +2024,18 @@ void llama_sample_tail_free(struct llama_context * ctx, llama_token_data_array *
     }
 
     // Normalize the second derivatives
-    float second_derivatives_sum = std::accumulate(second_derivatives.begin(), second_derivatives.end(), 0.0f);
-    for (float & value : second_derivatives) {
-        value /= second_derivatives_sum;
+    {
+        const float second_derivatives_sum = std::accumulate(second_derivatives.begin(), second_derivatives.end(), 0.0f);
+
+        if (second_derivatives_sum > 1e-6f) {
+            for (float & value : second_derivatives) {
+                value /= second_derivatives_sum;
+            }
+        } else {
+            for (float & value : second_derivatives) {
+                value = 1.0f / second_derivatives.size();
+            }
+        }
     }
 
     float cum_sum = 0.0f;