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

0cc4m · web-flow · commit c9c64dee572c · 2025-05-20T10:11:56.000+02:00
diff --git 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) {

Original file line number	Diff line number	Diff line change
`@@ -1368,6 +1368,10 @@ ggml_tensor * llm_graph_context::build_attn(`
`1368`	`1368`
`1369`	`1369`	`if (wo) {`
`1370`	`1370`	`cur = build_lora_mm(wo, cur);`
	`1371`	`+ if (arch == LLM_ARCH_GLM4) {`
	`1372`	`+ // GLM4 seems to have numerical issues with half-precision accumulators`
	`1373`	`+ ggml_mul_mat_set_prec(cur, GGML_PREC_F32);`
	`1374`	`+ }`
`1371`	`1375`	`}`
`1372`	`1376`
`1373`	`1377`	`if (wo_b) {`