fix: fix the bug when return_numpy is false

modeling_jina_embeddings_v4.py

@@ -350,18 +350,19 @@ class JinaEmbeddingsV4Model(Qwen2_5_VLForConditionalGeneration):
                             embeddings = embeddings[:, :truncate_dim]
                     else:
                         embeddings = embeddings.multi_vec_emb
-                    if return_multivector:
+                    if return_multivector and not return_numpy:
                         # Get valid token mask from attention_mask
                         valid_tokens = attention_mask.bool()
                         # 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
-                        embeddings = torch.stack(embeddings)
-                    results.append(
-                        embeddings.cpu()
-                        if return_numpy
-                        else list(torch.unbind(embeddings))
-                    )
+                        results.extend(embeddings)
+                    else:
+                        results.append(
+                            embeddings.cpu()
+                            if return_numpy
+                            else list(torch.unbind(embeddings))
+                        )
         if return_numpy:
             return np.concatenate([result.numpy() for result in results], axis=0)
         return [item for sublist in results for item in sublist]