降低max len#

Author: root

examples/werewolf/train.sh

@@ -24,7 +24,7 @@ python -m agentlightning.verl \
     actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=1 \
     actor_rollout_ref.rollout.multi_turn.format=hermes \
     actor_rollout_ref.model.path=${BASE_MODEL} \
-    data.max_prompt_length=16384 \
+    data.max_prompt_length=10240 \
     data.max_response_length=1024 \
     data.truncation='error' \
     trainer.val_before_train=True \

examples/werewolf/werewolf_agent.py

@@ -360,7 +360,7 @@ def _process_triplets_with_rewards(self, wolf_win_flag: bool, NAME_TO_ROLE: dict
             # 添加日志检查
             prompt_length = len(prompt_ids)
             print(f"Prompt length: {prompt_length} tokens")
-            # if prompt_length >= 16384:  # 你的 max_prompt_length TODO: 过长上下文发送处理.拆掉上下文中的think
+            # if prompt_length >= 10240:  # 你的 max_prompt_length TODO: 过长上下文发送处理.拆掉上下文中的think
             #     print(f"WARNING: Prompt truncated! Original length: {prompt_length}")
             prompt = self.tokenizer.decode(prompt_ids)
             response = self.tokenizer.decode(response_ids)