[Misc] Remove CP Redundant Variables after FIA operator enables for CANN 8.5 (#6013)

### What this PR does / why we need it?
PCP/DCP splits the kv-cache onto different cards. After introducing the
parameter cp-kv-cache-interleave-size, the first size tokens will be
cached at Card 0, and so on.
However, if there are too few tokens, some cards will not store the
key-value pairs, resulting in values ​​of 0, corrupted values, and
precision issues. Currently, additional operations are introduced to
avoid this precision problem.

After we integrate FIA operator in mla_cp._forward_decode and CANN
updates to 8.5.0, we now can remove these additional operations.
### Does this PR introduce _any_ user-facing change?

### How was this patch tested?
passed all CI by CANN 8.5.0
- vLLM version: v0.13.0
- vLLM main:
2c24bc6996

Signed-off-by: dsxsteven <dsxsteven@sina.com>
Signed-off-by: dsxsteven <36877507+dsxsteven@users.noreply.github.com>

vllm_ascend/attention/context_parallel/attention_cp.py

@@ -73,9 +73,6 @@ class AscendAttentionCPMetadataBuilder(AscendAttentionMetadataBuilder):
         device: torch.device,
     ):
         super().__init__(kv_cache_spec, layer_names, vllm_config, device)
-        self.batch_seq_mask_buf = torch.empty(
-            vllm_config.scheduler_config.max_num_batched_tokens, dtype=torch.uint8, device=device
-        )
         self.pcp_size = get_pcp_group().world_size
         self.pcp_rank = get_pcp_group().rank_in_group if self.pcp_size > 1 else 0
         self.dcp_size = get_decode_context_model_parallel_world_size()
@@ -216,14 +213,9 @@ class AscendAttentionCPMetadataBuilder(AscendAttentionMetadataBuilder):
         if num_decodes > 0:
             num_computed_tokens_array = np.array(num_computed_tokens_of_pcp_dcp)
             num_computed_tokens_array = num_computed_tokens_array[:num_decodes]
-            batch_seq_mask = num_computed_tokens_array[:, self.pcp_rank, self.dcp_rank] == 0
             # TODO: numpy array mode of the shared memory is used to improve performance
-            self.batch_seq_mask_buf[: batch_seq_mask.shape[0]].copy_(
-                torch.from_numpy(batch_seq_mask), non_blocking=True
-            )
             decode_metadata = AscendMetadataForDecode(
                 num_computed_tokens_of_pcp_dcp=num_computed_tokens_array,
-                batch_seq_mask=self.batch_seq_mask_buf[: batch_seq_mask.shape[0]],
                 block_tables=block_table[:num_decodes],
             )
 
@@ -525,7 +517,7 @@ class AscendAttentionCPImpl(AscendAttentionBackendImpl):
             graph_params.handles[num_tokens].append(handle)
         else:
             attn_out, attn_lse = torch_npu.npu_fused_infer_attention_score(query, k_nope, value, **common_kwargs)
-        attn_out_lse = _process_attn_out_lse(attn_out, attn_lse, attn_metadata.decode_meta.batch_seq_mask)
+        attn_out_lse = _process_attn_out_lse(attn_out, attn_lse)
         attn_out = _npu_attention_update(self.head_size, attn_out_lse)
         return attn_out
 
@@ -633,9 +625,6 @@ class AscendAttentionCPImpl(AscendAttentionBackendImpl):
             actual_seq_lengths_kv=prefill_metadata.chunked_context.actual_seq_lengths_kv,
             actual_seq_lengths=attn_metadata.prefill.chunked_context.actual_chunk_seq_lengths,
         )
-        batch_chunk_seq_mask = attn_metadata.prefill.chunked_context.batch_chunk_seq_mask
-        lse_mask = batch_chunk_seq_mask[:, None, None].expand_as(prefix_chunk_lse)
-        prefix_chunk_lse = torch.where(lse_mask, -torch.inf, prefix_chunk_lse)
 
         return prefix_chunk_output, prefix_chunk_lse