fix by review

Signed-off-by: tangtianyi <tangtianyi4@huawei.com>

Author: qyqc731

vllm_ascend/attention/attention_mask.py

@@ -41,14 +41,17 @@ def __init__(
         dtype: torch.dtype,
         device: torch.device = None,
     ):
+        # NOTE: The device argument specifies the target NPU 
+        # to be used for the newly added FIA operator.
+        # Only pass this parameter when using the new FIA operator.
+
         attn_mask = _generate_attn_mask(max_seq_len, dtype)
 
         self._seq_len_cached = attn_mask.shape[0]
         self.attn_mask_cache = attn_mask
         self.device = device
         if self.device:
-            #NOTE: New compressed mask needs to be sent to certain device, 
-            # so device needs to be passed here.
+
             assigned_mask_dim = 2048
             self.chunked_prefill_attn_mask = torch.triu(torch.ones(assigned_mask_dim, assigned_mask_dim), diagonal=1
             ).to(torch.int8).to(device)

vllm_ascend/attention/attention_v1.py

@@ -460,6 +460,8 @@ def _forward_v1_style(
                 attn_metadata.seq_lens.to(device=query.device)
 
         if self.compressed_mask:
+            # TODO:The npu_fused_infer_attention_score op is planned to
+            # be utilized in a wider range in upcoming versions.
             num_block, block_size, head_num, head_dim = self.key_cache.shape
             key = self.key_cache.view(num_block, block_size, -1)
             value = self.value_cache.view(num_block, block_size, -1)

vllm_ascend/worker/model_runner_v1.py

@@ -825,7 +825,7 @@ def _make_attention_mask(self, seq_lens, position,
                              attn_state) -> torch.Tensor:
         # Chunk Prefill situation.
         if attn_state == AscendAttentionState.ChunkedPrefill and not self.vllm_config.model_config.use_mla:
-            if selkf.compressed_mas:
+            if self.compressed_mask:
                 return self.attn_mask_builder.get_splitfuse_attn_mask()
             else:
                 return self.attn_mask_builder.get_splitfuse_attn_mask(