fix cache logic

modelling_RW.py

@@ -72,7 +72,7 @@ class RotaryEmbedding(torch.nn.Module):
         dtype=torch.bfloat16,
         start_idx: int = 0,
     ) -> torch.Tensor:
-        if seq_len != self.seq_len_cached and self.start_idx != start_idx:
+        if seq_len != self.seq_len_cached or self.start_idx != start_idx:
             self.seq_len_cached = seq_len
             t = torch.arange(start_idx, start_idx+seq_len, device=device).type_as(self.inv_freq)
             freqs = torch.einsum("i,j->ij", t, self.inv_freq)