直接运行 https://gitee.com/ascend/op-plugin/blob/master/test/test_custom_ops/test_npu_flash_attention.py (最近修改在commit `529b54cd45b49a56136351b717c6619bb14f33fc`)，会报错：

```

in supported_op_exec

qk = torch.matmul(query, key.transpose(2, 3)).mul(scale)

RuntimeError: "baddbmm_with_gemm" not implemented for 'Half'

```

由于CPU host不支持fp16类型

需要将

```python

output = self.supported_op_exec(query, key, value)

```

修改为

```python

output = self.supported_op_exec(query.npu(), key.npu(), value.npu())

```

(但是不经过`trans_BNSD2BSH`，和supported_op_exec计算逻辑一致)