llama : fix defrag logic

ggml-ci

Author: ggerganov

src/llama.cpp

@@ -8802,11 +8802,13 @@ static int llama_decode_impl(
 
     // decide if we need to defrag the kv cache
     if (cparams.causal_attn && cparams.defrag_thold >= 0.0f) {
-        const float fragmentation = kv_self.n >= 128 ? 1.0f - float(kv_self.used)/float(kv_self.n) : 0.0f;
+        // - do not defrag small contexts (i.e. < 2048 tokens)
+        // - do not defrag if the padding is bigger than the defrag threshold
+        const float fragmentation = (kv_self.n >= 2048 && kv_self.n*cparams.defrag_thold >= llama_kv_cache_get_padding(cparams)) ? 1.0f - float(kv_self.used)/float(kv_self.n) : 0.0f;
 
         // queue defragmentation for next llama_kv_cache_update
         if (fragmentation > cparams.defrag_thold) {
-            //LLAMA_LOG_INFO("fragmentation: %.2f\n", fragmentation);
+            LLAMA_LOG_DEBUG("%s: fragmentation: %.2f - requesting defrag\n", __func__, fragmentation);
 
             llama_kv_cache_defrag(kv_self);
         }