update common.py

vllm/v1/attention/backends/mla/common.py

@@ -860,7 +860,7 @@ class MLACommonMetadataBuilder(AttentionMetadataBuilder[M]):
                 prefill_metadata.cudnn_workspace = self.cudnn_workspace
 
         # TODO @ wangming
-        # decode_metadata = None
+        decode_metadata = None
         # if num_decodes > 0:
             # if self.use_spec_decode and not common_attn_metadata.spec_layer_decoding:
             #     query_lens = self.num_scheduled_tokens_np[:num_decodes]
@@ -920,14 +920,14 @@ class MLACommonMetadataBuilder(AttentionMetadataBuilder[M]):
             #             seq_lens=seq_lens[:self._num_decode_tokens],
             #         )
 
-            decode_metadata = self._build_decode(
-                block_table_tensor=block_table_tensor[:num_decodes, ...],
-                seq_lens_cpu=seq_lens_cpu[:num_decodes],
-                seq_lens_device=seq_lens[:num_decodes],
-                query_start_loc_cpu=query_start_loc_cpu[:num_decodes + 1],
-                query_start_loc_device=query_start_loc[:num_decodes + 1],
-                num_decode_tokens=num_decode_tokens,
-            )
+        decode_metadata = self._build_decode(
+            block_table_tensor=block_table_tensor[:num_decodes, ...],
+            seq_lens_cpu=seq_lens_cpu[:num_decodes],
+            seq_lens_device=seq_lens[:num_decodes],
+            query_start_loc_cpu=query_start_loc_cpu[:num_decodes + 1],
+            query_start_loc_device=query_start_loc[:num_decodes + 1],
+            num_decode_tokens=num_decode_tokens,
+        )
 
         attn_metadata = self.metadata_cls(
             num_reqs=common_attn_metadata.num_reqs,