From 7519e2f11ac36f85819e305ded1f0dcca260e015 Mon Sep 17 00:00:00 2001
From: leejet <leejet714@gmail.com>
Date: Sun, 12 Oct 2025 17:27:51 +0800
Subject: [PATCH] to_add_out precision fix

---
 qwen_image.hpp | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)

diff --git a/qwen_image.hpp b/qwen_image.hpp
index ab16b82..90357af 100644
--- a/qwen_image.hpp
+++ b/qwen_image.hpp
@@ -97,7 +97,10 @@ namespace Qwen {
             blocks["to_out.0"] = std::shared_ptr<GGMLBlock>(new Linear(inner_dim, out_dim, out_bias));
             // to_out.1 is nn.Dropout
 
-            blocks["to_add_out"] = std::shared_ptr<GGMLBlock>(new Linear(inner_dim, out_context_dim, out_bias));
+            float scale = 1.f / 32.f;
+            // The purpose of the scale here is to prevent NaN issues in certain situations.
+            // For example when using CUDA but the weights are k-quants (not all prompts).
+            blocks["to_add_out"] = std::shared_ptr<GGMLBlock>(new Linear(inner_dim, out_context_dim, out_bias, false, false, scale));
         }
 
         std::pair<ggml_tensor*, ggml_tensor*> forward(struct ggml_context* ctx,