RKNN模拟器推理及连板推理的结果不一致

这个rknn模型在pc的模拟器可以跑出和pytorch和onnx完全一致的结果，没有问题。
但是连板子跑，或者直接在板子上调用rknnlite跑，结果都是一样错误的，而且多次运行还会得出完全不同的输出结果，请问怎么解决？
rknn模型文件链接

https://wwtb.lanzoum.com/isdy31l6e66f
连板运行日志、驱动和runtime版本如下，日志是没有任何错误的

1. I Start adb...
   
2. I Connect to Device success!
   
3. I NPUTransfer: Starting NPU Transfer Client, Transfer version 2.1.0 (b5861e7@2020-11-23T11:50:36)
   
4. D NPUTransfer: Transfer spec = local:transfer_proxy
   
5. D NPUTransfer: Transfer interface successfully opened, fd = 3
   
6. D RKNNAPI: ==============================================
   
7. D RKNNAPI: RKNN VERSION:
   
8. D RKNNAPI:   API: 1.6.0 (535b468 build@2023-12-11T09:05:46)
   
9. D RKNNAPI:   DRV: rknn_server: 1.6.0 (535b468 build@2023-12-11T17:05:07)
   
10. D RKNNAPI:   DRV: rknnrt: 1.6.0 (585b3edcf@2023-12-11T15:30:19)
    
11. D RKNNAPI: ==============================================
    
12. D RKNNAPI: Input tensors:
    
13. D RKNNAPI:   index=0, name=input, n_dims=4, dims=[3, 384, 640, 3], n_elems=2211840, size=4423680, w_stride = 0, size_with_stride = 0, fmt=NHWC, type=FP16, qnt_type=NONE, zp=0, scale=1.000000
    
14. D RKNNAPI: Output tensors:
    
15. D RKNNAPI:   index=0, name=bbox_preds, n_dims=3, dims=[3, 4, 5040], n_elems=60480, size=120960, w_stride = 0, size_with_stride = 0, fmt=UNDEFINED, type=FP16, qnt_type=NONE, zp=0, scale=1.000000
    
16. D RKNNAPI:   index=1, name=cls_scores, n_dims=3, dims=[3, 40, 5040], n_elems=604800, size=1209600, w_stride = 0, size_with_stride = 0, fmt=UNDEFINED, type=FP16, qnt_type=NONE, zp=0, scale=1.000000

复制代码