diff --git a/llama.cpp b/llama.cpp
index 65830e54a..0bab95563 100644
--- a/llama.cpp
+++ b/llama.cpp
@@ -4559,14 +4559,14 @@ struct llm_build_context {
 
                 Qcur = ggml_rope_custom(
                     ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens), inp_pos,
-                    n_embd_head, 0, 0, n_orig_ctx, freq_base, freq_scale,
+                    hparams.n_rot, 0, 0, n_orig_ctx, freq_base, freq_scale,
                     ext_factor, attn_factor, beta_fast, beta_slow
                 );
                 cb(Qcur, "Qcur", il);
 
                 Kcur = ggml_rope_custom(
                     ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos,
-                    n_embd_head, 0, 0, n_orig_ctx, freq_base, freq_scale,
+                    hparams.n_rot, 0, 0, n_orig_ctx, freq_base, freq_scale,
                     ext_factor, attn_factor, beta_fast, beta_slow
                 );
                 cb(Kcur, "Kcur", il);
@@ -4689,6 +4689,7 @@ struct llm_build_context {
 
         const int64_t n_embd_head = hparams.n_embd_head_v;
         GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+        GGML_ASSERT(n_embd_head == hparams.n_rot);
 
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
@@ -4732,12 +4733,12 @@ struct llm_build_context {
                     case MODEL_7B:
                         Qcur = ggml_rope_custom(
                             ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos,
-                            n_embd_head, 0, 0, n_orig_ctx, freq_base, freq_scale,
+                            hparams.n_rot, 0, 0, n_orig_ctx, freq_base, freq_scale,
                             ext_factor, attn_factor, beta_fast, beta_slow
                         );
                         Kcur = ggml_rope_custom(
                             ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos,
-                            n_embd_head, 0, 0, n_orig_ctx, freq_base, freq_scale,
+                            hparams.n_rot, 0, 0, n_orig_ctx, freq_base, freq_scale,
                             ext_factor, attn_factor, beta_fast, beta_slow
                         );
                         break;
@@ -4810,6 +4811,7 @@ struct llm_build_context {
         const int64_t n_embd_head = hparams.n_embd_head_v;
         const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
         GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+        GGML_ASSERT(n_embd_head == hparams.n_rot);
 
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
@@ -4868,13 +4870,13 @@ struct llm_build_context {
 
                 // using mode = 2 for neox mode
                 Qcur = ggml_rope_custom(
-                    ctx0, Qcur, inp_pos, n_embd_head, 2, 0, n_orig_ctx,
+                    ctx0, Qcur, inp_pos, hparams.n_rot, 2, 0, n_orig_ctx,
                     freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow
                 );
                 cb(Qcur, "Qcur", il);
 
                 Kcur = ggml_rope_custom(
-                    ctx0, Kcur, inp_pos, n_embd_head, 2, 0, n_orig_ctx,
+                    ctx0, Kcur, inp_pos, hparams.n_rot, 2, 0, n_orig_ctx,
                     freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow
                 );
                 cb(Kcur, "Kcur", il);
@@ -5031,9 +5033,8 @@ struct llm_build_context {
         struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
 
         const int64_t n_embd_head = hparams.n_embd_head_v;
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-
-        const int64_t n_rot = n_embd_head_k / 2;
+        GGML_ASSERT(n_embd_head   == hparams.n_embd_head_k);
+        GGML_ASSERT(n_embd_head/2 == hparams.n_rot);
 
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
@@ -5110,7 +5111,7 @@ struct llm_build_context {
 
                 // RoPE the first n_rot of q/k, pass the other half, and concat.
                 struct ggml_tensor * qrot = ggml_view_3d(
-                        ctx0, tmpq, n_rot, n_head, n_tokens,
+                        ctx0, tmpq, hparams.n_rot, n_head, n_tokens,
                         ggml_element_size(tmpq) * n_embd_head,
                         ggml_element_size(tmpq) * n_embd_head * n_head,
                         0
@@ -5118,7 +5119,7 @@ struct llm_build_context {
                 cb(qrot, "qrot", il);
 
                 struct ggml_tensor * krot = ggml_view_3d(
-                        ctx0, tmpk, n_rot, n_head, n_tokens,
+                        ctx0, tmpk, hparams.n_rot, n_head, n_tokens,
                         ggml_element_size(tmpk) * n_embd_head,
                         ggml_element_size(tmpk) * n_embd_head * n_head,
                         0
@@ -5127,29 +5128,29 @@ struct llm_build_context {
 
                 // get the second half of tmpq, e.g tmpq[n_rot:, :, :]
                 struct ggml_tensor * qpass = ggml_view_3d(
-                        ctx0, tmpq, n_rot, n_head, n_tokens,
+                        ctx0, tmpq, hparams.n_rot, n_head, n_tokens,
                         ggml_element_size(tmpq) * n_embd_head,
                         ggml_element_size(tmpq) * n_embd_head * n_head,
-                        ggml_element_size(tmpq) * n_rot
+                        ggml_element_size(tmpq) * hparams.n_rot
                         );
                 cb(qpass, "qpass", il);
 
                 struct ggml_tensor * kpass = ggml_view_3d(
-                        ctx0, tmpk, n_rot, n_head, n_tokens,
+                        ctx0, tmpk, hparams.n_rot, n_head, n_tokens,
                         ggml_element_size(tmpk) * n_embd_head,
                         ggml_element_size(tmpk) * n_embd_head * n_head,
-                        ggml_element_size(tmpk) * n_rot
+                        ggml_element_size(tmpk) * hparams.n_rot
                         );
                 cb(kpass, "kpass", il);
 
                 struct ggml_tensor * qrotated = ggml_rope_custom(
-                    ctx0, qrot, inp_pos, n_rot, 2, 0, n_orig_ctx,
+                    ctx0, qrot, inp_pos, hparams.n_rot, 2, 0, n_orig_ctx,
                     freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow
                 );
                 cb(qrotated, "qrotated", il);
 
                 struct ggml_tensor * krotated = ggml_rope_custom(
-                    ctx0, krot, inp_pos, n_rot, 2, 0, n_orig_ctx,
+                    ctx0, krot, inp_pos, hparams.n_rot, 2, 0, n_orig_ctx,
                     freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow
                 );
                 cb(krotated, "krotated", il);
@@ -5529,6 +5530,7 @@ struct llm_build_context {
 
         const int64_t n_embd_head = hparams.n_embd_head_v;
         GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+        GGML_ASSERT(n_embd_head == hparams.n_rot);
 
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
@@ -5691,13 +5693,13 @@ struct llm_build_context {
 
                 // using mode = 2 for neox mode
                 Qcur = ggml_rope_custom(
-                    ctx0, Qcur, inp_pos, n_embd_head, 2, 0, n_orig_ctx,
+                    ctx0, Qcur, inp_pos, hparams.n_rot, 2, 0, n_orig_ctx,
                     freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow
                 );
                 cb(Qcur, "Qcur", il);
 
                 Kcur = ggml_rope_custom(
-                    ctx0, Kcur, inp_pos, n_embd_head, 2, 0, n_orig_ctx,
+                    ctx0, Kcur, inp_pos, hparams.n_rot, 2, 0, n_orig_ctx,
                     freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow
                 );
                 cb(Kcur, "Kcur", il);
@@ -5872,6 +5874,7 @@ struct llm_build_context {
 
         const int64_t n_embd_head = hparams.n_embd_head_v;
         GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+        GGML_ASSERT(n_embd_head == hparams.n_rot);
 
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
@@ -5915,13 +5918,13 @@ struct llm_build_context {
                 cb(Vcur, "Vcur", il);
 
                 Qcur = ggml_rope_custom(
-                        ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens), inp_pos,
+                        ctx0, ggml_reshape_3d(ctx0, Qcur, hparams.n_rot, n_head,    n_tokens), inp_pos,
                         n_embd_head, 2, 0, n_orig_ctx, freq_base, freq_scale,
                         ext_factor, attn_factor, beta_fast, beta_slow);
                 cb(Qcur, "Qcur", il);
 
                 Kcur = ggml_rope_custom(
-                        ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos,
+                        ctx0, ggml_reshape_3d(ctx0, Kcur, hparams.n_rot, n_head_kv, n_tokens), inp_pos,
                         n_embd_head, 2, 0, n_orig_ctx, freq_base, freq_scale,
                         ext_factor, attn_factor, beta_fast, beta_slow);
                 cb(Kcur, "Kcur", il);