Can not disable fp16.

[Cc-Hy]

Hello, I' building a model which contains a pre-trained teacher model as

def __init__(xxx, teacher_model_cfg=None):
   ...
   if teacher_model_cfg:  
       self.teacher_model = builder.build_model(teacher_model_cfg)
       if self.teacher_model.init_cfg: 
           initialize(self.teacher_model, self.teacher_model.init_cfg)
       for p in self.teacher_model.parameters():
           p.requires_grad = False

And I'd like to use fp16 for training and add this in the cfg file

fp16 = dict(loss_scale='dynamic')

However, when I call the teacher model as

with torch.no_grad():
    teacher_x = self.teacher_model.forward_teacher(points)

I meet the error in mmdet3d/ops/spconv/ops.py: Where the dtype of the filters is float32 but the input is float16. And I think the teacher model (CenterPoint) might not support fp16, so I add force_fp32 to the teacher model's method as

 @force_fp32() 
 def forward_teacher(self,  
        points=None, 
        img=None, 
        img_metas=None): 
        voxels, num_points, coors = self.voxelize(points) 
        voxel_features = self.pts_voxel_encoder(voxels, num_points, coors) 
        batch_size = coors[-1, 0] + 1 
        x = self.pts_middle_encoder(voxel_features, coors, batch_size) 
        return x 

But I still meet the the error. And I tried several things but just can not disable the fp16 in the teacher model. So I'd like to ask how can I disable the fp 16 in the teacher model, or how can I solve this error and use the fp 16 in the teacher model. Thank you.