whisper : bypass cross-attention scaling for OpenVINO backend (#3997)

* whisper : bypass cross-attention scaling for OpenVINO backend * whisper : use correct gating flag
2026-08-22 08:14:11 +03:00 · 2026-08-22 08:14:11 +03:00 · 233fe1fc9b
parent 51de5e8bb0
commit 233fe1fc9b
1 changed files with 2 additions and 0 deletions
--- 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,