deprecate use_workflow

Author: thwu1

rllm/trainer/agent_trainer.py

@@ -40,8 +40,7 @@ def __init__(
             agent_args: Optional arguments to pass to the agent class
             env_args: Optional arguments to pass to the environment class
         """
-
-        if workflow_class is not None and config.rllm.workflow.use_workflow:
+        if workflow_class is not None:
             if agent_class is not None:
                 raise ValueError("agent_class is not supported when using workflow, instead use workflow_args['agent_cls']")
             if agent_args is not None:
@@ -81,7 +80,7 @@ def _train_tinker(self):
         from rllm.trainer.tinker.tinker_agent_trainer import TinkerAgentTrainer
         from rllm.trainer.tinker.tinker_workflow_trainer import TinkerWorkflowTrainer
 
-        if self.config.rllm.workflow.use_workflow:
+        if self.workflow_class is not None:
             trainer = TinkerWorkflowTrainer(
                 config=self.config,
                 workflow_class=self.workflow_class,

rllm/trainer/config/tinker_agent_trainer.yaml

@@ -63,10 +63,6 @@ trainer:
   val_before_train: true
   default_local_dir: '/tmp/rllm-tinker-checkpoints'
 
-rllm:
-  workflow:
-    use_workflow: false
-
 # Hydra configuration
 hydra:
   run:

rllm/trainer/config/tinker_workflow_trainer.yaml

@@ -60,10 +60,6 @@ trainer:
   val_before_train: true
   default_local_dir: '/tmp/rllm-tinker-checkpoints'
 
-rllm:
-  workflow:
-    use_workflow: true
-
 # Hydra configuration
 hydra:
   run:

rllm/trainer/verl/train_agent_ppo.py

@@ -10,7 +10,7 @@
 import ray
 from omegaconf import OmegaConf
 
-from rllm.trainer.env_agent_mappings import AGENT_CLASS_MAPPING, ENV_CLASS_MAPPING, WORKFLOW_CLASS_MAPPING
+from rllm.trainer.env_agent_mappings import AGENT_CLASS_MAPPING, ENV_CLASS_MAPPING
 from rllm.trainer.verl.agent_ppo_trainer import AgentPPOTrainer
 
 # Local application imports
@@ -155,9 +155,8 @@ def run(self, config, workflow_class=None, workflow_args=None, agent_class=None,
         val_reward_fn = load_reward_manager(config, tokenizer, num_examine=1, **config.reward_model.get("reward_kwargs", {}))
         resource_pool_manager = ResourcePoolManager(resource_pool_spec=resource_pool_spec, mapping=mapping)
 
-        if config.rllm.workflow.use_workflow:
-            if workflow_class is None:
-                workflow_class = WORKFLOW_CLASS_MAPPING[config.rllm.workflow.name]
+        if workflow_class is not None:
+            # Should provide workflow_class if want to use workflow trainer
             workflow_args = workflow_args or {}
             if config.rllm.workflow.get("workflow_args") is not None:
                 for key, value in config.rllm.workflow.get("workflow_args").items():