Fix bug, update verl

Author: Reason-Wang

agents/agents/agents/chain/chain_base.py

@@ -214,11 +214,11 @@ def prepare_chain_messages(self, start_messages: Union[List[dict], np.ndarray]):
 
         return messages_list, other_info_list
 
-    def validate_run_args(self, max_steps: int, num_chains: int):
+    def validate_run_args(self, max_steps: int, num_chains: int, enable_streaming: bool):
         assert max_steps >= 1, "max_steps must be at least 1."
         assert num_chains >= 1, "num_chains must be at least 1."
         for observer in self.streaming_manager.observers:
-            if isinstance(observer, ConsoleStreamObserver):
+            if isinstance(observer, ConsoleStreamObserver) and enable_streaming:
                 assert num_chains == 1, "num_chains must be 1 when ConsoleStreamObserver is used."
 
 
@@ -241,7 +241,7 @@ async def run_async(self,
             enable_streaming: Whether to enable streaming mode.
             streaming_callback: Optional callback for streaming events.
         """
-        self.validate_run_args(max_steps, num_chains)
+        self.validate_run_args(max_steps, num_chains, enable_streaming)
         Monitor.ensure_started()
         self.reset()
         messages_list, other_info_list = self.prepare_chain_messages(start_messages)

agents/agents/agents/templates/templates.py

@@ -215,6 +215,13 @@ def encode(self, messages: List[Dict], tokenizer: PreTrainedTokenizer, return_te
         attention_mask = []
         labels = []
         action_mask = []
+
+        if tokenizer.bos_token and tokenizer.add_bos_token:
+            input_ids.append(tokenizer.bos_token_id)
+            attention_mask.append(1)
+            labels.append(-100)
+            action_mask.append(0)
+        
         for element, mask_flag in zip(elements, mask_flags):
             cur_input_ids = tokenizer.encode(element, add_special_tokens=False)
             input_ids.extend(cur_input_ids)

agents/agents/agents/templates/utils.py

@@ -139,7 +139,14 @@ def tokenize_conversation(
     :return: input_ids, attention_mask, labels, action_mask
     """
     chat = Chat(template=template, messages=messages, tokenizer=tokenizer)
-    return chat.tokenize(tokenizer,  add_generation_prompt=add_generation_prompt, tools=tools)
+    inputs = chat.tokenize(tokenizer,  add_generation_prompt=add_generation_prompt, tools=tools)
+    if max_length is not None:
+        inputs['input_ids'] = inputs['input_ids'][:, :max_length]
+        inputs['attention_mask'] = inputs['attention_mask'][:, :max_length]
+        inputs['labels'] = inputs['labels'][:, :max_length]
+        inputs['action_mask'] = inputs['action_mask'][:, :max_length]
+
+    return inputs
 
 def convert_inputs_to_vision_inputs(template: str,
                                     inputs: dict,