diff --git a/ggml/include/ggml.h b/ggml/include/ggml.h
index 3e0a852a38..32462d79a7 100644
--- a/ggml/include/ggml.h
+++ b/ggml/include/ggml.h
@@ -1983,6 +1983,7 @@ extern "C" {
 
     // set the offset dims for RoPE
     // a must be GGML_OP_ROPE or GGML_OP_ROPE_BACK
+    // vision RoPE is not supported
     // example: (marking: x = rotated, 0 = unrotated)
     //     n_embd = 10, n_dims = 4, offset = 2 --> [00xxxx0000]
     GGML_API struct ggml_tensor * ggml_rope_set_offset(
diff --git a/ggml/src/ggml-cpu/ops.cpp b/ggml/src/ggml-cpu/ops.cpp
index 001e1ae85a..2b5f68444b 100644
--- a/ggml/src/ggml-cpu/ops.cpp
+++ b/ggml/src/ggml-cpu/ops.cpp
@@ -5979,6 +5979,8 @@ static void ggml_compute_forward_rope_flt(
     memcpy(&beta_slow,   (int32_t *) dst->op_params + 10, sizeof(float));
     memcpy(&sections,    (int32_t *) dst->op_params + 11, sizeof(int)*4);
 
+    const int n_offs = ((int32_t *) dst->op_params)[15];
+
     GGML_TENSOR_UNARY_OP_LOCALS
 
     //printf("ne0: %d, ne1: %d, ne2: %d, ne3: %d\n", ne0, ne1, ne2, ne3);
@@ -5995,6 +5997,10 @@ static void ggml_compute_forward_rope_flt(
     GGML_ASSERT(n_dims <= ne0);
     GGML_ASSERT(n_dims % 2 == 0);
 
+    GGML_ASSERT(n_offs >= 0);
+    GGML_ASSERT(n_offs % 2 == 0);
+    GGML_ASSERT(n_offs + n_dims <= ne0);
+
     // rows per thread
     const int dr = (nr + nth - 1)/nth;
 
@@ -6020,6 +6026,7 @@ static void ggml_compute_forward_rope_flt(
 
     if (is_vision) {
         GGML_ASSERT(n_dims == ne0/2);
+        GGML_ASSERT(n_offs == 0);
     }
 
     const float * freq_factors = NULL;
@@ -6068,12 +6075,12 @@ static void ggml_compute_forward_rope_flt(
 
                 switch (mode) {
                     case GGML_ROPE_TYPE_NORMAL:
-                        rotate_pairs<T>(n_dims, 1, cache, src, dst_data, 1);
+                        rotate_pairs<T>(n_dims, 1, cache, src + n_offs, dst_data + n_offs, 1);
                         break;
                     case GGML_ROPE_TYPE_NEOX:
                     case GGML_ROPE_TYPE_MROPE:
                     case GGML_ROPE_TYPE_IMROPE:
-                        rotate_pairs<T>(n_dims, n_dims/2, cache, src, dst_data);
+                        rotate_pairs<T>(n_dims, n_dims/2, cache, src + n_offs, dst_data + n_offs);
                         break;
                     case GGML_ROPE_TYPE_VISION:
                         rotate_pairs<T>(ne0, n_dims, cache, src, dst_data);
@@ -6084,7 +6091,11 @@ static void ggml_compute_forward_rope_flt(
 
                 if (!is_vision) {
                     // fill the remain channels with data from src tensor
-                    for (int64_t i0 = n_dims; i0 < ne0; i0 += 2) {
+                    for (int64_t i0 = 0; i0 < ne0; i0 += 2) {
+                        if (i0 == n_offs) {
+                            i0 += n_dims - 2; // skip the rotated channels
+                            continue;
+                        }
                         const T * const src = (T *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
                         T * dst_data  = (T *)((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
 
diff --git a/ggml/src/ggml.c b/ggml/src/ggml.c
index 2355a1cc37..1a60fec791 100644
--- a/ggml/src/ggml.c
+++ b/ggml/src/ggml.c
@@ -4429,6 +4429,11 @@ struct ggml_tensor * ggml_rope_set_offset(
         struct ggml_tensor  * a,
         int                   n_offs) {
     GGML_ASSERT(a->op == GGML_OP_ROPE || a->op == GGML_OP_ROPE_BACK);
+    GGML_ASSERT(n_offs >= 0);
+
+    const int32_t mode = ggml_get_op_params_i32(a, 2);
+    GGML_ASSERT(mode != GGML_ROPE_TYPE_VISION);
+
     ggml_set_op_params_i32(a, 15, n_offs);
     return a;
 }