From 233fe1fc9b48a09e361d3594520838ca266537fe Mon Sep 17 00:00:00 2001
From: Mostafa <mostafas.main.email@gmail.com>
Date: Sat, 22 Aug 2026 08:14:11 +0300
Subject: [PATCH] whisper : bypass cross-attention scaling for OpenVINO backend
 (#3997)

* whisper : bypass cross-attention scaling for OpenVINO backend

* whisper : use correct gating flag
---
 src/whisper.cpp | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/src/whisper.cpp b/src/whisper.cpp
index 89146e2e4..6a4f5c1b0 100644
--- a/src/whisper.cpp
+++ b/src/whisper.cpp
@@ -2310,7 +2310,9 @@ static struct ggml_cgraph * whisper_build_graph_cross(
                 layer.cross_attn_k_w,
                 cur);
 
+#ifndef GGML_USE_OPENVINO // scaling is handled internally in OpenVINO backend
         Kcross = ggml_scale(ctx0, Kcross, Kscale);
+#endif
 
         struct ggml_tensor * Vcross = ggml_mul_mat(ctx0,
                 layer.cross_attn_v_w,