To cuda

osrt/sam/transformer.py

@@ -199,11 +199,10 @@ class Attention(nn.Module):
         self.internal_dim = embedding_dim // downsample_rate
         self.num_heads = num_heads
         assert self.internal_dim % num_heads == 0, "num_heads must divide embedding_dim."
-        device = next(super().parameters()).device
-        self.q_proj = nn.Linear(embedding_dim, self.internal_dim).to(device)
-        self.k_proj = nn.Linear(embedding_dim, self.internal_dim).to(device)
-        self.v_proj = nn.Linear(embedding_dim, self.internal_dim).to(device)
-        self.out_proj = nn.Linear(self.internal_dim, embedding_dim).to(device)
+        self.q_proj = nn.Linear(embedding_dim, self.internal_dim).cuda()
+        self.k_proj = nn.Linear(embedding_dim, self.internal_dim).cuda()
+        self.v_proj = nn.Linear(embedding_dim, self.internal_dim).cuda()
+        self.out_proj = nn.Linear(self.internal_dim, embedding_dim).cuda()
 
     def _separate_heads(self, x: Tensor, num_heads: int) -> Tensor:
         b, n, c = x.shape