diff --git a/ggml/src/ggml-cpu/iqp.cpp b/ggml/src/ggml-cpu/iqp.cpp
index c2e07c0144..871a4f2955 100644
--- a/ggml/src/ggml-cpu/iqp.cpp
+++ b/ggml/src/ggml-cpu/iqp.cpp
@@ -16,6 +16,16 @@
 
 #define UNUSED GGML_UNUSED
 
+// smallest src1 batch for which the decode pays for itself
+#define GGML_IQP_MIN_BATCH 8
+
+// same, per expert, for MUL_MAT_ID
+#define GGML_IQP_MIN_BATCH_ID 8
+
+bool ggml_cpu_iqp_mul_mat_id_min_batch(int64_t cne1) {
+    return cne1 >= GGML_IQP_MIN_BATCH_ID;
+}
+
 // src0 rows interleaved per panel
 #define IQP_NB_ROWS 8
 
diff --git a/ggml/src/ggml-cpu/iqp.h b/ggml/src/ggml-cpu/iqp.h
index f33e091e32..f831c9b8eb 100644
--- a/ggml/src/ggml-cpu/iqp.h
+++ b/ggml/src/ggml-cpu/iqp.h
@@ -12,15 +12,8 @@
 extern "C" {
 #endif
 
-// smallest src1 batch for which the decode pays for itself
-#define GGML_IQP_MIN_BATCH 8
-
-// same, per expert, for MUL_MAT_ID
-#define GGML_IQP_MIN_BATCH_ID 8
-
-static inline bool ggml_cpu_iqp_mul_mat_id_min_batch(int64_t cne1) {
-    return cne1 >= GGML_IQP_MIN_BATCH_ID;
-}
+// whether cne1 rows of src1 are enough for the decode to pay for itself, per expert, for MUL_MAT_ID
+bool ggml_cpu_iqp_mul_mat_id_min_batch(int64_t cne1);
 
 bool ggml_cpu_iqp_supports_mul_mat(const struct ggml_tensor * dst);