From c9c64dee572c5eedf073a6c6eb5d92d8283f7639 Mon Sep 17 00:00:00 2001
From: 0cc4m <picard12@live.de>
Date: Tue, 20 May 2025 10:11:56 +0200
Subject: [PATCH] Set GLM4 blk.*.attn_output.weight, kqv_out-* matmul to
 GGML_PREC_F32 to fix infinity values in output (#13639)

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

diff --git a/src/llama-graph.cpp b/src/llama-graph.cpp
index 410d26087..e745b41e3 100644
--- a/src/llama-graph.cpp
+++ b/src/llama-graph.cpp
@@ -1368,6 +1368,10 @@ 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) {