From 7ad56c7382290514dc4ca02550e007817fda478b Mon Sep 17 00:00:00 2001
From: 0cc4m <picard12@live.de>
Date: Sat, 24 May 2025 13:27:53 +0000
Subject: [PATCH] Move GLM4 f32 attention fix to the correct function

---
 src/llama-graph.cpp | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/src/llama-graph.cpp b/src/llama-graph.cpp
index 13e36d161c614..cdd5887de961c 100644
--- a/src/llama-graph.cpp
+++ b/src/llama-graph.cpp
@@ -1287,6 +1287,10 @@ ggml_tensor * llm_graph_context::build_attn(
 
     if (wo) {
         cur = build_lora_mm(wo, cur);
+        if (arch == LLM_ARCH_GLM4) {
+            // GLM4 seems to have numerical issues with half-precision accumulators
+            ggml_mul_mat_set_prec(cur, GGML_PREC_F32);
+        }
     }
 
     if (wo_b) {
@@ -1367,10 +1371,6 @@ ggml_tensor * llm_graph_context::build_attn(
 
     if (wo) {
         cur = build_lora_mm(wo, cur);
-        if (arch == LLM_ARCH_GLM4) {
-            // GLM4 seems to have numerical issues with half-precision accumulators
-            ggml_mul_mat_set_prec(cur, GGML_PREC_F32);
-        }
     }
 
     if (wo_b) {