diff --git a/users/schmitt/experiments/config/pipelines/global_vs_segmental_2022_23_rf/dependencies/returnn/network_builder_rf/global_/decoder.py b/users/schmitt/experiments/config/pipelines/global_vs_segmental_2022_23_rf/dependencies/returnn/network_builder_rf/global_/decoder.py
index 82d6e9279..4e06921f4 100644
--- a/users/schmitt/experiments/config/pipelines/global_vs_segmental_2022_23_rf/dependencies/returnn/network_builder_rf/global_/decoder.py
+++ b/users/schmitt/experiments/config/pipelines/global_vs_segmental_2022_23_rf/dependencies/returnn/network_builder_rf/global_/decoder.py
@@ -25,13 +25,13 @@ def decoder_default_initial_state(
     """Default initial state"""
     state = rf.State()
 
-    if self.trafo_att:
+    if self.trafo_att and not use_mini_att and not use_zero_att:
       state.trafo_att = self.trafo_att.default_initial_state(batch_dims=batch_dims)
-      att_dim = self.trafo_att.model_dim
-    else:
-      att_dim = self.att_num_heads * self.enc_out_dim
+      # att_dim = self.trafo_att.model_dim
+    # else:
+      # att_dim = self.att_num_heads * self.enc_out_dim
 
-    state.att = rf.zeros(list(batch_dims) + [att_dim])
+    state.att = rf.zeros(list(batch_dims) + [self.att_dim])
     state.att.feature_dim_axis = len(state.att.dims) - 1
 
     if "lstm" in self.decoder_state:
diff --git a/users/schmitt/experiments/config/pipelines/global_vs_segmental_2022_23_rf/dependencies/returnn/network_builder_rf/segmental/recog.py b/users/schmitt/experiments/config/pipelines/global_vs_segmental_2022_23_rf/dependencies/returnn/network_builder_rf/segmental/recog.py
index dc8836ca1..01312f608 100644
--- a/users/schmitt/experiments/config/pipelines/global_vs_segmental_2022_23_rf/dependencies/returnn/network_builder_rf/segmental/recog.py
+++ b/users/schmitt/experiments/config/pipelines/global_vs_segmental_2022_23_rf/dependencies/returnn/network_builder_rf/segmental/recog.py
@@ -438,17 +438,27 @@ def get_score(
 
   ilm_eos_log_prob = rf.zeros(batch_dims, dtype="float32")
   if ilm_state is not None:
-    ilm_step_out, ilm_state = model.label_decoder.loop_step(
-      **att_enc_args,
-      enc_spatial_dim=enc_spatial_dim,
-      input_embed=input_embed_label_model,
-      segment_lens=segment_lens,
-      segment_starts=segment_starts,
-      center_positions=center_positions,
-      state=ilm_state,
-      use_mini_att=ilm_type == "mini_att",
-      use_zero_att=ilm_type == "zero_att",
-    )
+    if model.center_window_size is None:
+      ilm_step_out, ilm_state = model.label_decoder.loop_step(
+        **att_enc_args,
+        enc_spatial_dim=enc_spatial_dim,
+        input_embed=input_embed_label_model,
+        state=ilm_state,
+        use_mini_att=ilm_type == "mini_att",
+        use_zero_att=ilm_type == "zero_att",
+      )
+    else:
+      ilm_step_out, ilm_state = model.label_decoder.loop_step(
+        **att_enc_args,
+        enc_spatial_dim=enc_spatial_dim,
+        input_embed=input_embed_label_model,
+        segment_lens=segment_lens,
+        segment_starts=segment_starts,
+        center_positions=center_positions,
+        state=ilm_state,
+        use_mini_att=ilm_type == "mini_att",
+        use_zero_att=ilm_type == "zero_att",
+      )
     ilm_logits, _ = model.label_decoder.decode_logits(
       input_embed=input_embed_label_model,
       s=ilm_step_out["s"],
@@ -773,7 +783,19 @@ def model_recog(
 
   # ILM
   if ilm_type is not None:
-    ilm_state = model.label_decoder.default_initial_state(batch_dims=batch_dims_, use_mini_att=ilm_type == "mini_att")
+    if isinstance(model.label_decoder, GlobalAttDecoder):
+      ilm_state = model.label_decoder.decoder_default_initial_state(
+        batch_dims=batch_dims_,
+        use_mini_att=ilm_type == "mini_att",
+        use_zero_att=ilm_type == "zero_att",
+        enc_spatial_dim=enc_spatial_dim
+      )
+    else:
+      ilm_state = model.label_decoder.default_initial_state(
+        batch_dims=batch_dims_,
+        use_mini_att=ilm_type == "mini_att",
+        use_zero_att=ilm_type == "zero_att",
+      )
   else:
     ilm_state = None