diff --git a/omni/layers/attention/backend/gdn_attn.py b/omni/layers/attention/backend/gdn_attn.py
index bb3d37ba19425d6ca64f33d13eb55c45b90181ad..b03d888d69f6efd34f034c749f863cdd8c89a127 100644
--- a/omni/layers/attention/backend/gdn_attn.py
+++ b/omni/layers/attention/backend/gdn_attn.py
@@ -26,20 +26,8 @@ class GDNAttentionBackend(AttentionBackend):
 
 @dataclass
 class GDNAttentionMetadata:
-    block_tables: torch.Tensor  # not used
-    query_lens: torch.Tensor  # not used
-    query_lens_list: List  # not used
-    seq_lens: torch.Tensor  # not used
-    seq_lens_list: List  # not used
-    max_query_len: Optional[int]  # not used
-    slot_mapping: torch.Tensor  # not used
-    slot_indices: torch.Tensor  # not used
-    is_only_prefill: bool  # not used
-    attn_state: AscendAttentionState  # not used
-    cos: Optional[torch.Tensor]  # not used
-    sin: Optional[torch.Tensor]  # not used
-    is_pd_seperate_d: bool  # not used
-    kv_index: Optional[torch.Tensor]  # not used
+    attn_state: AscendAttentionState
+    is_pd_seperate_d: bool
 
     num_prefills: int
     num_prefill_tokens: int
@@ -305,31 +293,13 @@ class GDNAttentionMetadataBuilder(AscendAttentionMetadataBuilder):
             non_spec_query_start_loc[num_decodes +
                                      1:].fill_(non_spec_num_query_tokens)
             
-        ascend_attn_metadata = super().build(
-            num_reqs=num_reqs,
-            num_actual_tokens=num_actual_tokens,
-            max_query_len=max_query_len,
-            common_prefix_len=None,
-            graph_pad_size=graph_pad_size,
-            **kwargs,
-        )
+        attn_state = self.runner.attn_state
+        is_pd_seperate_d = self.runner.vllm_config.kv_transfer_config is not None and \
+                           self.runner.vllm_config.kv_transfer_config.kv_role == 'kv_consumer'
 
         attn_metadata = GDNAttentionMetadata(
-            block_tables=ascend_attn_metadata.block_tables,
-            query_lens=ascend_attn_metadata.query_lens,
-            query_lens_list=ascend_attn_metadata.query_lens_list,
-            seq_lens=ascend_attn_metadata.seq_lens,
-            seq_lens_list=ascend_attn_metadata.seq_lens_list,
-            max_query_len=ascend_attn_metadata.max_query_len,
-            slot_mapping=ascend_attn_metadata.slot_mapping,
-            slot_indices=ascend_attn_metadata.slot_indices,
-            is_only_prefill=ascend_attn_metadata.is_only_prefill,
-            attn_state=ascend_attn_metadata.attn_state,
-            cos=ascend_attn_metadata.cos,
-            sin=ascend_attn_metadata.sin,
-            is_pd_seperate_d=ascend_attn_metadata.is_pd_seperate_d,
-            kv_index=ascend_attn_metadata.kv_index,
-            # up to here, just copy AscendMetadata and not used.
+            attn_state=attn_state,
+            is_pd_seperate_d=is_pd_seperate_d,
             num_prefills=num_prefills,
             num_prefill_tokens=num_prefill_tokens,
             num_decodes=num_decodes,
@@ -373,58 +343,19 @@ class GDNAttentionMetadataBuilder(AscendAttentionMetadataBuilder):
             dtype=self.runner.input_batch.block_table[0].get_device_tensor(max_pad_size).dtype
         )
 
-        query_lens = torch.ones(max_pad_size, dtype=torch.long, device=self.runner.device, pin_memory=True)
-        seq_lens = query_lens * 2
-
-        slot_indices = torch.stack([slot_mapping // self.block_size, slot_mapping % self.block_size], dim=1)
-
-        fake_positions = torch.zeros(max_pad_size, dtype=torch.int64, device=self.device)
-
-        cos, sin = None, None
-
         is_pd_seperate_d = self.runner.vllm_config.kv_transfer_config is not None and \
                            self.runner.vllm_config.kv_transfer_config.kv_role == 'kv_consumer'
 
         non_spec_query_start_loc = torch.tensor([0, 1], device=self.runner.device)
         non_spec_state_indices_tensor = self.block_table.block_table[:, 0]
 
-        ascend_attn_metadata = AscendMetadata(
-            num_actual_tokens=num_tokens,
-            block_tables=block_table,
-            query_lens=query_lens,
-            query_lens_list=query_lens.tolist(),
-            seq_lens=seq_lens,
-            seq_lens_list=seq_lens.tolist(),
-            slot_mapping=slot_mapping,
-            slot_indices=slot_indices,
-            is_only_prefill=False,
-            attn_state=self.runner.attn_state,
-            cos=cos,
-            sin=sin,
-            is_pd_seperate_d=is_pd_seperate_d
-        )
-
-          
         attn_metadata = GDNAttentionMetadata(
-            block_tables=ascend_attn_metadata.block_tables,
-            query_lens=ascend_attn_metadata.query_lens,
-            query_lens_list=ascend_attn_metadata.query_lens_list,
-            seq_lens=ascend_attn_metadata.seq_lens,
-            seq_lens_list=ascend_attn_metadata.seq_lens_list,
-            max_query_len=ascend_attn_metadata.max_query_len,
-            slot_mapping=ascend_attn_metadata.slot_mapping,
-            slot_indices=ascend_attn_metadata.slot_indices,
-            is_only_prefill=ascend_attn_metadata.is_only_prefill,
-            attn_state=ascend_attn_metadata.attn_state,
-            cos=ascend_attn_metadata.cos,
-            sin=ascend_attn_metadata.sin,
-            is_pd_seperate_d=ascend_attn_metadata.is_pd_seperate_d,
-            kv_index=ascend_attn_metadata.kv_index,
-            # up to here, just copy AscendMetadata
+            attn_state=self.runner.attn_state,
+            is_pd_seperate_d=is_pd_seperate_d,
             num_prefills=0,
             num_prefill_tokens=0,
             num_decodes=num_tokens,
-            is_decode=(num_decodes > 0),
+            is_decode=(num_tokens > 0),
             num_decode_tokens=num_tokens,
             num_spec_decodes=0,
             num_spec_decode_tokens=0,