zero-mAP fix return `.detach()` to EMA (#9056)

utils/torch_utils.py

@@ -422,7 +422,7 @@ class ModelEMA:
         for k, v in self.ema.state_dict().items():
             if v.dtype.is_floating_point:  # true for FP16 and FP32
                 v *= d
-                v += (1 - d) * msd[k]
+                v += (1 - d) * msd[k].detach()
         assert v.dtype == msd[k].dtype == torch.float32, f'EMA {v.dtype} and model {msd[k]} must be updated in FP32'
 
     def update_attr(self, model, include=(), exclude=('process_group', 'reducer')):