Update utils_llama.py

research/long-context-llama/H2O/utils_llama.py

@@ -177,6 +177,8 @@ class H2OLlamaAttention(nn.Module):
                 # sin and cos are specific to RoPE models; cache_position needed for the static cache
                 key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx)
 
+            import pdb; pdb.set_trace()
+
             kv_seq_len = past_key_value.get_seq_length() if past_key_value is not None else key_states.shape[-2]
             cos, sin = self.rotary_emb(value_states, kv_seq_len)