Fix sync_all_reduce to consider update->compute->update case

ignite/metrics/metric.py

@@ -574,8 +574,10 @@ def another_wrapper(self: Metric, *args: Any, **kwargs: Any) -> Callable:
                             f"number or tensor but `{attr}` has type {type(t)}"
                         )
                     unreduced_attrs[attr] = t
-                    # Here `clone` is necessary since `idist.all_reduce` modifies `t` inplace in the case
+                    # Here `clone` is necessary since `idist.all_reduce` modifies `t` inp-lace in the case
                     # `t` is a tensor and its `device` is same as that of the process.
+                    # TODO: Remove this dual behavior of `all_reduce` to always either return a new tensor or
+                    #       modify it in-place.
                     t_reduced = idist.all_reduce(cast(float, t) if isinstance(t, Number) else t.clone(), **op_kwargs)
                     setattr(self, attr, t_reduced)
 

ignite/metrics/precision.py

@@ -143,7 +143,7 @@ def compute(self) -> Union[torch.Tensor, float]:
         fraction = self._numerator / (self._denominator + (self.eps if self._average != "samples" else 0))
 
         if self._average == "weighted":
-            _weight = idist.all_reduce(self._weight)
+            _weight = idist.all_reduce(self._weight.clone())
             sum_of_weights = cast(torch.Tensor, _weight).sum() + self.eps
             return ((fraction @ _weight) / sum_of_weights).item()  # type: ignore
         elif self._average == "micro" or self._average == "samples":