volcengine
qwen2.5-vl chat_template Does Not Support Tool Call
It appears that the qwen2.5-vl chat_template currently does not support tool calls. Because of this, the following script from VERL cannot work as expected:
Script: https://github.com/volcengine/verl/blob/main/examples/sglang_multiturn/geo3k/run_qwen2.5-3b_geo3k_multiturn.sh
This script relies on tool-call capability during chat generation, but the qwen2.5-vl template does not include the required tool schema formatting or tool-call tokens, resulting in failure during execution.
class AgentLoopWorkerBase: """Agent loop worker takes a batch of messages and run each message in an agent loop."""
def __init__(
self,
config: DictConfig,
server_handles: list[ray.actor.ActorHandle],
reward_router_address: str = None,
):
"""Initialize agent loop manager.
Args:
config (DictConfig): YAML config.
server_handles (List[ray.actor.ActorHandle]): OpenAI compatible LLM server actor handles.
"""
self.config = config
# for recipe to change
if not hasattr(self, "server_manager"):
self.server_manager = AsyncLLMServerManager(config, server_handles)
self.reward_router_address = reward_router_address
model_path = config.actor_rollout_ref.model.path
self.model_name = "/".join(model_path.split("/")[-2:])
local_path = copy_to_local(config.actor_rollout_ref.model.path)
self.tokenizer = hf_tokenizer(local_path, trust_remote_code=True)
self.processor = hf_processor(local_path, trust_remote_code=True)
the agent loop uses the default chat template of the LLM. Could you please explain how the chat template is overrided. ps: the sglang multiturn rollout can not support the rollout of qwen2.5vl, so we use agent loop instead.
Thanks a lot!
CONFIG_PATH="$PROJECT_DIR/examples/sglang_multiturn/config"
python3 -m verl.trainer.main_ppo \
--config-path="$CONFIG_PATH" \
--config-name='geo3k_multiturn_grpo' \
It loads the geo3k_multiturn_grpo config, where the template are overrided