From cb5072fa778101e1d66bfb64214abe8109a396cc Mon Sep 17 00:00:00 2001
From: bri-prism <288398250+bri-prism@users.noreply.github.com>
Date: Thu, 15 Oct 2026 14:09:03 -0700
Subject: [PATCH] llama: skip graph reservation for unchanged layer capture

---
 src/llama-context.cpp | 4 ++++
 1 file changed, 4 insertions(+)

diff --git a/src/llama-context.cpp b/src/llama-context.cpp
index 3a37572122df..0f0f4dc7e382 100644
--- a/src/llama-context.cpp
+++ b/src/llama-context.cpp
@@ -1356,6 +1356,10 @@ void llama_context::set_embeddings_layer_inp(uint32_t lid, bool enable) {
 
     GGML_ASSERT(lid <= model.hparams.n_layer());
 
+    if (cparams.embeddings_layer_inp[lid] == enable) {
+        return;
+    }
+
     cparams.embeddings_layer_inp[lid] = enable;
 
     // note: without this reserve, the draft acceptance drops to zero. not sure why - this is unexpected