Merge branch 'optimize-block-tables' into 'v0.9.2-dev'

[fix]优化 block_tables 的生成逻辑，增加对混合情况的检测，确保在存在空和非空块时正确计算最大块长度。

See merge request dcutoolkit/deeplearing/vllm!191

vllm/attention/backends/utils.py

@@ -246,12 +246,33 @@ class CommonMetadataBuilder(AttentionMetadataBuilder[TAttentionMetadata]):
                 device, non_blocking=True)
 
         else:
-            block_tables = make_tensor_with_pad(
-                self.block_tables,
-                pad=0,
-                dtype=torch.int,
-                device=device,
-            )
+            has_empty: bool = any(len(bt) == 0 for bt in self.block_tables)
+            has_non_empty = any(len(bt) > 0 for bt in self.block_tables)
+            max_block_length = 0
+            if has_empty and has_non_empty:
+                for inter_data in self.input_builder.inter_data_list:
+                    block_tables = inter_data.block_tables
+                    if block_tables:
+                        for seq_id in inter_data.seq_ids:
+                            if seq_id in block_tables:
+                                block_table = block_tables[seq_id]
+                                max_block_length = max(max_block_length, len(block_table))
+            if max_block_length >0:
+                block_tables = make_tensor_with_pad(
+                    self.block_tables,
+                    pad=0,
+                    dtype=torch.int,
+                    device=device,
+                    max_len=max_block_length,
+                )
+            else:
+                 block_tables = make_tensor_with_pad(
+                    self.block_tables,
+                    pad=0,
+                    dtype=torch.int,
+                    device=device,
+                )
+
         assert max_query_len > 0, "query_lens: {}".format(query_lens)
 
         assert device is not None