refact attn metadata build

Signed-off-by: weiguihua2 <[email protected]>

Author: weiguihua2

vllm_ascend/attention/utils.py

@@ -1,5 +1,5 @@
 from dataclasses import dataclass
-from typing import Any
+from typing import Any, Optional
 
 import torch
 
@@ -36,7 +36,7 @@ class AscendCommonAttentionMetadata:
 
     slot_mapping_cpu: torch.Tensor
 
-    actual_seq_lengths_q: list[int] = None
+    actual_seq_lengths_q: Optional[list[int]] = None
 
     positions: torch.Tensor = None
 
@@ -70,7 +70,7 @@ class TorchairCommonAttentionMetadata:
 
     decode_token_per_req: int
 
-    actual_seq_lengths_q: list[int] = None
+    actual_seq_lengths_q: Optional[list[int]] = None
 
     attn_mask: torch.Tensor = None
 

vllm_ascend/worker/model_runner_v1.py

@@ -812,7 +812,6 @@ def get_eagle_atten_dict(
                 attn_mask=self.attn_mask,
                 spec_attn_mask=self.spec_attn_mask,
                 attn_state=self.attn_state,
-                max_num_blocks_per_req=self.max_num_blocks_per_req,
                 decode_token_per_req=self.decode_token_per_req,
             )
             attn_metadata_i = self.attn_metadata_builder.build(