diff --git a/rvc/f0/fcpe.py b/rvc/f0/fcpe.py
index bf8897e..d60a1fc 100644
--- a/rvc/f0/fcpe.py
+++ b/rvc/f0/fcpe.py
@@ -1,8 +1,7 @@
-from typing import Any, Optional, Union
+from typing import Optional, Union
 
 import numpy as np
 import torch
-from torchfcpe import spawn_bundled_infer_model
 
 from .f0 import F0Predictor
 
@@ -24,6 +23,8 @@ class FCPE(F0Predictor):
             device,
         )
 
+        from torchfcpe import spawn_bundled_infer_model  # must be imported at here, or it will cause fairseq crash on training
+
         self.model = spawn_bundled_infer_model(self.device)
 
     def compute_f0(
diff --git a/rvc/layers/encoders.py b/rvc/layers/encoders.py
index 2c3f00a..1fc2478 100644
--- a/rvc/layers/encoders.py
+++ b/rvc/layers/encoders.py
@@ -192,7 +192,7 @@ class PosteriorEncoder(nn.Module):
     def __call__(
         self, x: torch.Tensor, x_lengths: torch.Tensor, g: Optional[torch.Tensor] = None
     ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
-        super().__call__(x, x_lengths, g=g)
+        return super().__call__(x, x_lengths, g=g)
 
     def forward(
         self, x: torch.Tensor, x_lengths: torch.Tensor, g: Optional[torch.Tensor] = None
diff --git a/rvc/layers/utils.py b/rvc/layers/utils.py
index d979087..418578b 100644
--- a/rvc/layers/utils.py
+++ b/rvc/layers/utils.py
@@ -20,9 +20,9 @@ def slice_on_last_dim(
     start_indices: List[int],
     segment_size=4,
 ) -> torch.Tensor:
-    new_shape = x.shape
+    new_shape = [*x.shape]
     new_shape[-1] = segment_size
-    ret = torch.empty(new_shape)
+    ret = torch.empty(new_shape, device=x.device)
     for i in range(x.size(0)):
         idx_str = start_indices[i]
         idx_end = idx_str + segment_size