[BugFix]fix ring_mla invalid param

Signed-off-by: weiguihua2 <weiguihua2@huawei.com>

Author: weiguihua2

vllm_ascend/worker/model_runner_v1.py

@@ -1326,7 +1326,7 @@ def _list_to_tensor(lst, device, dtype=torch.int32):
             attn_mask_seqlens = torch.tensor([chunk_seqlens, chunk_seqlens], dtype=torch.int32)
             head_attn_nomask_seqlens = torch.tensor([chunk_seqlens, kv_with_q_head_nomask_seqlens], dtype=torch.int32)
             tail_attn_nomask_seqlens = torch.tensor([chunk_seqlens, kv_with_q_tail_nomask_seqlens], dtype=torch.int32)
-            cp_prefill_mask = torch.triu(torch.ones(512, 512, device=self.device, dtype=torch.bfloat16), 1)
+            cp_prefill_mask = torch.triu(torch.ones(512, 512, device=self.device, dtype=self.dtype), 1)
 
             self.extra_long_seq_kwargs = {
                 'attn_mask_seqlens': attn_mask_seqlens,