Move GLM4 f32 attention fix to the correct function (#13750)

0cc4m · web-flow · commit 259469c4b57c · 2025-05-24T16:49:12.000+02:00
diff --git 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) {
