diff --git a/scripts/inference.py b/scripts/inference.py
index 22d604d..cc6f90d 100644
--- a/scripts/inference.py
+++ b/scripts/inference.py
@@ -32,10 +32,14 @@ def main(configs : DictConfig) -> None:
     print_configs(configs)
     os.environ["DATA_ROOT_DIR"] = configs.data_root_dir
     os.environ["CKPT_ROOT_DIR"] = configs.ckpt_root_dir
+    hydra_output_dir = HydraConfig.get().runtime.output_dir
     configs = ConfigDict(
         OmegaConf.to_container(configs.exp, resolve=True)
     )
-    dump_dir = HydraConfig.get().runtime.output_dir
+    # exp.dump_dir allows the caller to separate data output from hydra's
+    # working directory, which is needed when parallel seed workers each
+    # require a distinct hydra.run.dir but share one output directory.
+    dump_dir = configs.get("dump_dir", None) or hydra_output_dir
     configs.dump_dir = dump_dir
     error_dir = Path(dump_dir) / "errors"
     if DIST_WRAPPER.rank == 0:
@@ -45,7 +49,7 @@ def main(configs : DictConfig) -> None:
     logger.info(
         f"Distributed environment: world size: {DIST_WRAPPER.world_size}, "
         + f"global rank: {DIST_WRAPPER.rank}, local rank: {DIST_WRAPPER.local_rank}"
-    )    
+    )
     device = torch.device("cuda:{}".format(DIST_WRAPPER.local_rank))
     torch.cuda.set_device(device)
     if DIST_WRAPPER.world_size > 1:
@@ -72,7 +76,7 @@ def main(configs : DictConfig) -> None:
     )
 
     infer_runner.run()
-    
+
 
 if __name__ == "__main__":
     main()
diff --git a/src/utils/inference/dumper.py b/src/utils/inference/dumper.py
index e543c41..5948ac7 100644
--- a/src/utils/inference/dumper.py
+++ b/src/utils/inference/dumper.py
@@ -127,25 +127,25 @@ def _save_structure_sequence(
                         new_atom_array,
                         per_chain_edits=per_chain_edits,
                         design_modality=design_modality,
-                    )    
-
-                    output_fpath = os.path.join(
-                        prediction_save_dir,
-                        f"{sample_name}_seed_{seed}_bb_{rank}_seq_{seq_var_idx}.cif",
-                    )                    
-
-                    if b_factor is not None:
-                        # b_factor.shape == [N_sample, N_atom]
-                        new_atom_array.set_annotation("b_factor", np.round(b_factor[idx], 2))
-
-                    save_structure_cif(
-                        atom_array=new_atom_array,
-                        pred_coordinate=pred_coordinates[idx],
-                        output_fpath=output_fpath,
-                        entity_poly_type=entity_poly_type,
-                        pdb_id=sample_name,
                     )
 
+                output_fpath = os.path.join(
+                    prediction_save_dir,
+                    f"{sample_name}_seed_{seed}_bb_{rank}_seq_{seq_var_idx}.cif",
+                )
+
+                if b_factor is not None:
+                    # b_factor.shape == [N_sample, N_atom]
+                    new_atom_array.set_annotation("b_factor", np.round(b_factor[idx], 2))
+
+                save_structure_cif(
+                    atom_array=new_atom_array,
+                    pred_coordinate=pred_coordinates[idx],
+                    output_fpath=output_fpath,
+                    entity_poly_type=entity_poly_type,
+                    pdb_id=sample_name,
+                )
+
     def _apply_sequence_variant_to_atom_array(
         self,
         atom_array: AtomArray,