refact attn metadata build

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

Author: weiguihua2

vllm_ascend/attention/mla_v1.py

@@ -287,8 +287,7 @@ def build_torchair_graph_dummy(
             self, common_attn_metadata: AscendCommonAttentionMetadata,) -> AscendMLAMetadata:
         device = self.device
         num_reqs = common_attn_metadata.num_reqs
-        _, max_blocks = self.max_blocks
-        block_table = torch.zeros((num_reqs, max_blocks),
+        block_table = torch.zeros((num_reqs, self.max_blocks),
                                   dtype=torch.int32,
                                   device=device)
         block_table = self._get_graph_runner_block_tables(

vllm_ascend/torchair/torchair_attention.py

@@ -120,8 +120,7 @@ def build_torchair_graph_dummy(
             self, common_attn_metadata: AscendCommonAttentionMetadata) -> AscendTorchairMetadata:
         device = self.device
         num_reqs = common_attn_metadata.num_reqs
-        _, max_blocks = self.max_blocks
-        block_table = torch.zeros((num_reqs, max_blocks),
+        block_table = torch.zeros((num_reqs, self.max_blocks),
                                   dtype=torch.int32,
                                   device=device)
         block_table = self._get_graph_runner_block_tables(

vllm_ascend/torchair/torchair_model_runner.py

@@ -78,7 +78,7 @@ def _build_attention_metadata(self, with_prefill, num_reqs, skip_attn):
                 actual_seq_lengths_q=self.actual_seq_lengths_q,
                 attn_mask=self.attn_mask,
                 spec_attn_mask=self.spec_attn_mask,
-                attn_state=self.attn_state,
+                decode_token_per_req=self.decode_token_per_req,
             )
             attn_metadata = self.attn_metadata_builder.build_torchair_graph_dummy(common_attn_metadata)
         else: