From: 0cc4m <redacted>
Date: Tue, 20 May 2025 08:11:56 +0000 (+0200)
Subject: Set GLM4 blk.*.attn_output.weight, kqv_out-* matmul to GGML_PREC_F32 to fix infinity... 
X-Git-Url: https://git.djapps.eu/?a=commitdiff_plain;h=c9c64dee572c5eedf073a6c6eb5d92d8283f7639;p=pkg%2Fggml%2Fsources%2Fllama.cpp

Set GLM4 blk.*.attn_output.weight, kqv_out-* matmul to GGML_PREC_F32 to fix infinity values in output (#13639)
---

diff --git a/src/llama-graph.cpp b/src/llama-graph.cpp
index 410d2608..e745b41e 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) {