Cherry-pick #1439 to fix 'topk' on different devices for onnxruntime-gpu inference (#1603)

Co-authored-by: grimoire <yaoqian@sensetime.com>

mmdeploy/pytorch/functions/topk.py

@@ -17,7 +17,7 @@ def topk__dynamic(input: torch.Tensor,
                   sorted: bool = True):
     """Rewrite `topk` for default backend.
 
-    Cast k to tensor and makesure k is smaller than input.shape[dim].
+    Cast k to tensor and make sure k is smaller than input.shape[dim].
     """
     ctx = FUNCTION_REWRITER.get_context()
 
@@ -28,7 +28,8 @@ def topk__dynamic(input: torch.Tensor,
         k = torch.tensor(k, device=input.device, dtype=torch.long)
     # Always keep topk op for dynamic input
     if isinstance(size, torch.Tensor):
-        size = size.to(input.device)
+        # size would be treated as cpu tensor, trick to avoid that.
+        size = k.new_zeros(()) + size
     k = torch.where(k < size, k, size)
     return ctx.origin_func(input, k, dim=dim, largest=largest, sorted=sorted)