diff --git a/otherarch/gptj_v3.cpp b/otherarch/gptj_v3.cpp
index ff82ded76..68f998b36 100644
--- a/otherarch/gptj_v3.cpp
+++ b/otherarch/gptj_v3.cpp
@@ -494,8 +494,8 @@ bool gptj_eval(
                 }
             }
 
-            struct ggml_tensor *Qcur = ggml_rope_custom_inplace(ctx0, ggml_reshape_3d(ctx0, ggml_mul_mat(ctx0, model.layers[il].c_attn_q_proj_w, cur), n_embd / n_head, n_head, N), KQ_pos, n_rot, 0, n_ctx, 0, freq_base, freq_scale, NAN, 1, 32, 1);
-            struct ggml_tensor *Kcur = ggml_rope_custom_inplace(ctx0, ggml_reshape_3d(ctx0, ggml_mul_mat(ctx0, model.layers[il].c_attn_k_proj_w, cur), n_embd / n_head, n_head, N), KQ_pos, n_rot, 0, n_ctx, 0, freq_base, freq_scale, NAN, 1, 32, 1);
+            struct ggml_tensor *Qcur = ggml_rope_custom_inplace(ctx0, ggml_reshape_3d(ctx0, ggml_mul_mat(ctx0, model.layers[il].c_attn_q_proj_w, cur), n_embd / n_head, n_head, N), KQ_pos, n_rot, 0, n_ctx, 0, freq_base, freq_scale, 0, 1, 32, 1);
+            struct ggml_tensor *Kcur = ggml_rope_custom_inplace(ctx0, ggml_reshape_3d(ctx0, ggml_mul_mat(ctx0, model.layers[il].c_attn_k_proj_w, cur), n_embd / n_head, n_head, N), KQ_pos, n_rot, 0, n_ctx, 0, freq_base, freq_scale, 0, 1, 32, 1);
 
             // store key and value to memory
             {
diff --git a/otherarch/llama_v3.cpp b/otherarch/llama_v3.cpp
index 3a12e53cf..fbecde210 100644
--- a/otherarch/llama_v3.cpp
+++ b/otherarch/llama_v3.cpp
@@ -1614,11 +1614,11 @@ static struct ggml_cgraph * llama_v3_build_graph(
             }
 #endif
 
-            struct ggml_tensor *Kcur = ggml_rope_custom_inplace(ctx0, ggml_reshape_3d(ctx0, tmpk, n_embd_head, n_head_kv, N), KQ_pos, n_embd_head, 0, 0, 0, freq_base, freq_scale, NAN, 1, 32, 1);
+            struct ggml_tensor *Kcur = ggml_rope_custom_inplace(ctx0, ggml_reshape_3d(ctx0, tmpk, n_embd_head, n_head_kv, N), KQ_pos, n_embd_head, 0, 0, 0, freq_base, freq_scale, 0, 1, 32, 1);
             offload_func_kq(Kcur);
             ggml_set_name(Kcur, "Kcur");
 
-            struct ggml_tensor *Qcur = ggml_rope_custom_inplace(ctx0, ggml_reshape_3d(ctx0, tmpq, n_embd_head, n_head, N), KQ_pos, n_embd_head, 0, 0, 0, freq_base, freq_scale, NAN, 1, 32, 1);
+            struct ggml_tensor *Qcur = ggml_rope_custom_inplace(ctx0, ggml_reshape_3d(ctx0, tmpq, n_embd_head, n_head, N), KQ_pos, n_embd_head, 0, 0, 0, freq_base, freq_scale, 0, 1, 32, 1);
             offload_func_kq(Qcur);
             ggml_set_name(Qcur, "Qcur");
 
diff --git a/otherarch/neox_v3.cpp b/otherarch/neox_v3.cpp
index 0bdfece34..15afdbd71 100644
--- a/otherarch/neox_v3.cpp
+++ b/otherarch/neox_v3.cpp
@@ -522,8 +522,8 @@ bool gpt_neox_eval(
             }
 
             // using mode = 2 for GPT-NeoX mode
-            Qcur = ggml_rope_custom_inplace(ctx0, Qcur, KQ_pos, n_rot, 2, n_ctx, 0, freq_base, freq_scale, NAN, 1, 32, 1);
-            Kcur = ggml_rope_custom_inplace(ctx0, Kcur, KQ_pos, n_rot, 2, n_ctx, 0, freq_base, freq_scale, NAN, 1, 32, 1);
+            Qcur = ggml_rope_custom_inplace(ctx0, Qcur, KQ_pos, n_rot, 2, n_ctx, 0, freq_base, freq_scale, 0, 1, 32, 1);
+            Kcur = ggml_rope_custom_inplace(ctx0, Kcur, KQ_pos, n_rot, 2, n_ctx, 0, freq_base, freq_scale, 0, 1, 32, 1);
 
             // store key and value to memory
             {