diff --git a/llama.cpp b/llama.cpp
index b752ddc6b401f..2b91eec86c64f 100644
--- a/llama.cpp
+++ b/llama.cpp
@@ -5188,7 +5188,14 @@ static bool llm_load_tensors(
                     {
                         model.output_norm   = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
                         model.output_norm_b = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "bias"),   {n_embd});
-                        model.output        = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab});
+                        model.output        = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, false);
+                        if (!model.output) {
+                            // needs to be on GPU
+                            model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
+                            ml.n_created--; // artificial tensor
+                            ml.size_data += ggml_nbytes(model.output);
+                        }
+
                     }
 
                     for (int i = 0; i < n_layer; ++i) {