Fix condition when rolling cache (#150)

awq/modules/fused/attn.py

@@ -143,7 +143,7 @@ class QuantAttentionFused(nn.Module):
         will_cache_be_exceeded = self.start_pos + seqlen > self.max_seq_len
 
         # Reset and avoid retaining state when processing context
-        if will_cache_be_exceeded:
+        if will_cache_be_exceeded and seqlen > 1:
             self.start_pos = self.cache.roll_kv_n_steps(self.start_pos, n=self.start_pos)
         # Slowly roll out old tokens without performance hit if exceeded during decoding 
         elif will_cache_be_exceeded and seqlen == 1: