sampling : when top-k <= 0 -> noop (#13173)

ggml-ci

diff --git a/include/llama.h b/include/llama.h
index a13350e15be6aaf3029fb9cce9b74caf01c31b18..06c56395c139fb8cc936542cae2a124c16cfba71 100644 (file)
--- a/include/llama.h
+++ b/include/llama.h
@@ -1232,6 +1232,7 @@ extern "C" {
         "will be removed in the future (see https://github.com/ggml-org/llama.cpp/pull/9896#discussion_r1800920915)");
 
     /// @details Top-K sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751
+    /// Setting k <= 0 makes this a noop
     LLAMA_API struct llama_sampler * llama_sampler_init_top_k      (int32_t k);
 
     /// @details Nucleus sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751

diff --git a/src/llama-sampling.cpp b/src/llama-sampling.cpp
index d14979850285dbebe5320fb119d7e7736b16c11a..c0a5f9340d5851beade2deb00ef41bea2e38bc7e 100644 (file)
--- a/src/llama-sampling.cpp
+++ b/src/llama-sampling.cpp
@@ -232,7 +232,7 @@ static void llama_sampler_top_k_impl(llama_token_data_array * cur_p, int32_t k)
     // }
 
     if (k <= 0) {
-        k = cur_p->size;
+        return;
     }
 
     k = std::min(k, (int) cur_p->size);
@@ -298,6 +298,7 @@ static void llama_sampler_top_k_impl(llama_token_data_array * cur_p, int32_t k)
         }
         cur_p->sorted = true;
     }
+
     cur_p->size = k;
 }