From f356e277697efaa85642bf7d93d82c377c82ce2e Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Stanis=C5=82aw=20Szymczyk?= <sszymczy@gmail.com>
Date: Sat, 10 Aug 2024 10:32:15 +0200
Subject: [PATCH] llama : for clarity set is_encoding to true before building
 worst-case graph only if the model contains encoder

---
 src/llama.cpp | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/src/llama.cpp b/src/llama.cpp
index ce502b9c9..9c4f2aa72 100644
--- a/src/llama.cpp
+++ b/src/llama.cpp
@@ -16706,7 +16706,8 @@ struct llama_context * llama_new_context_with_model(
 
     ctx->sampling.rng = std::mt19937(params.seed);
     ctx->logits_all   = params.logits_all;
-    ctx->is_encoding  = true;
+    // build worst-case graph for encoder if a model contains encoder
+    ctx->is_encoding  = llama_model_has_encoder(model);
 
     uint32_t kv_size = cparams.n_ctx;
     ggml_type type_k = params.type_k;