restore original RoPE logic

Author: JohannesGaessler

ggml-cuda.cu

@@ -6418,7 +6418,11 @@ void ggml_cuda_soft_max(const ggml_tensor * src0, const ggml_tensor * src1, ggml
 
 void ggml_cuda_rope(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     GGML_ASSERT(ggml_is_contiguous(src0)); // TODO: this restriction is temporary until non-cont support is implemented
-    ggml_cuda_op(src0, src1, dst, ggml_cuda_op_rope, true, true);
+
+    const int mode = ((int32_t *) dst->op_params)[2];
+    const bool is_glm = mode & 4;
+
+    ggml_cuda_op(src0, src1, dst, ggml_cuda_op_rope, true, !is_glm); // flatten support not implemented for glm
 }
 
 void ggml_cuda_alibi(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {