fix: fix the bug when return_numpy is false

modeling_jina_embeddings_v4.py

@@ -356,9 +356,11 @@ class JinaEmbeddingsV4Model(Qwen2_5_VLForConditionalGeneration):
                         # Remove padding by selecting only valid tokens for each sequence
                         embeddings = [emb[mask] for emb, mask in zip(embeddings, valid_tokens)]
                         # Stack back into tensor with variable sequence lengths
-                        results.extend(embeddings)
+                        results.append(embeddings)
                     else:
                         results.append(
+                            # If return_numpy is True, move embeddings to CPU for numpy conversion
+                            # Otherwise, unbind the tensor into a list of individual tensors along dim=0
                             embeddings.cpu()
                             if return_numpy
                             else list(torch.unbind(embeddings))