move rp_bucket to relative_attention_bias' device

otherwise, `rp_bucket` will always be on cpu and fail if `self.relative_attention_bias` is on cuda

src/transformers/modeling_t5.py

@@ -286,6 +286,7 @@ class T5Attention(nn.Module):
             bidirectional=not self.is_decoder,
             num_buckets=self.relative_attention_num_buckets,
         )
+        rp_bucket = rp_bucket.to(self.relative_attention_bias.weight.device)
         values = self.relative_attention_bias(rp_bucket)  # shape (qlen, klen, num_heads)
         values = values.permute([2, 0, 1]).unsqueeze(0)  # shape (1, num_heads, qlen, klen)
         return values