diff --git a/pytorch_forecasting/metrics.py b/pytorch_forecasting/metrics.py
index 72ac211e..a0799503 100644
--- a/pytorch_forecasting/metrics.py
+++ b/pytorch_forecasting/metrics.py
@@ -1103,10 +1103,14 @@ def to_quantiles(self, y_pred: torch.Tensor, quantiles: List[float] = None, n_sa
             quantiles = self.quantiles
         try:
             distribution = self.map_x_to_distribution(y_pred)
-            quantiles = distribution.icdf(torch.tensor(quantiles, device=y_pred.device)[:, None, None]).permute(1, 2, 0)
+            quantiles = distribution.icdf(
+                torch.tensor(quantiles, device=y_pred.device, dtype=torch.float64)[:, None, None]
+            ).permute(1, 2, 0)
         except NotImplementedError:  # resort to derive quantiles empirically
             samples = torch.sort(self.sample(y_pred, n_samples), -1).values
-            quantiles = torch.quantile(samples, torch.tensor(quantiles, device=samples.device), dim=2).permute(1, 2, 0)
+            quantiles = torch.quantile(
+                samples, torch.tensor(quantiles, device=samples.device, dtype=torch.float64), dim=2
+            ).permute(1, 2, 0)
         return quantiles
 
 
diff --git a/pytorch_forecasting/models/deepar/__init__.py b/pytorch_forecasting/models/deepar/__init__.py
index 285a9a8d..5123f1f5 100644
--- a/pytorch_forecasting/models/deepar/__init__.py
+++ b/pytorch_forecasting/models/deepar/__init__.py
@@ -280,7 +280,7 @@ def decode_one(
                 hidden_state,
             ):
                 x = input_vector[:, [idx]]
-                x[:, 0, target_pos] = lagged_targets[-1]
+                x[:, 0, target_pos] = lagged_targets[-1].float()
                 for lag, lag_positions in lagged_target_positions.items():
                     if idx > lag:
                         x[:, 0, lag_positions] = lagged_targets[-lag]