improve communication protocol

Author: binary-husky

ajet/context_tracker/multiagent_tracking.py

@@ -84,19 +84,6 @@ def extract_text_content_from_content_dict(self, msg):
         #        },
         #    ],
         # }
-        # or tool_result format?? not observed yet:
-        # msg = {
-        #    "role": "tool",
-        #    "content": [
-        #        {
-        #           "type": "tool_result",
-        #           "id": "call_xxx",
-        #           "output": "tool output content",
-        #           "name": "tool_name"
-        #        },
-        #    ],
-        # }
-
 
         str_content = ""
         for item in msg["content"]:

ajet/default_config/ajet_default.yaml

@@ -2,7 +2,7 @@
 ajet:
   project_name: "ajet_default_project"
   experiment_name: "read_yaml_name"
-  experiment_dir: "auto"  # {exp-dir}/{experiment_name}
+  experiment_dir: "{exp-dir}/{experiment_name}"  # {exp-dir}/{experiment_name}
   backbone: verl # `debug` or `trinity` or `verl`
 
 
@@ -85,6 +85,7 @@ ajet:
       num_repeat: 1
 
 
+
   task_reader:
     # how to read dataset / environment
     type: huggingface_dat_repo # `env_service` or `jsonl_dataset_file` or `huggingface_dat_repo` or `data_generation` or `random_dummy`
@@ -306,8 +307,6 @@ ajet:
   swarm_mode_sample_collection_max_cached_episodes: 9999
 
   task_runner:
-    # submit llm infer submit method
-    llm_infer_submit_method: "async" # options: "sync", "async"
 
     # how to wrap the user-defined workflow
     wrapper_type: "asyncio-with-gc"

ajet/default_config/ajet_ts_default.yaml

@@ -2,7 +2,7 @@
 ajet:
   project_name: "ajet_default_project"
   experiment_name: "read_yaml_name"
-  experiment_dir: "auto"  # {exp-dir}/{experiment_name}
+  experiment_dir: "{exp-dir}/{experiment_name}"  # {exp-dir}/{experiment_name}
   backbone: verl
 
   model:
@@ -12,6 +12,10 @@ ajet:
   rollout:
     # the path to the workflow class
     user_workflow: null
+    # maximum number of parallel environments / simulate workers
+    max_env_worker: 128
+    # how many times a task should be repeated
+    num_repeat: 4
 
   task_reader:
     type: random_dummy # `env_service` or `jsonl_dataset_file` or `huggingface_dat_repo` or `data_generation` or `random_dummy`
@@ -53,12 +57,6 @@ ajet:
     train_batch_size: 32
     # [Hint]: The final number of samples per update will be: N_{sample} = (data.train_batch_size * rollout.num_repeat * rollout.multi_turn.expected_steps)
 
-  rollout:
-    # maximum number of parallel environments / simulate workers
-    max_env_worker: 128
-    # how many times a task should be repeated
-    num_repeat: 4
-
   trainer_common:
     logger: tensorboard
     n_gpus_per_node: 8

ajet/launcher.py

@@ -168,7 +168,7 @@ def main():
         from ajet.utils.swarm_overwatch import start_overwatch
 
         logger.info(f"Starting Swarm Overwatch for server: {args.swarm_overwatch}")
-        start_overwatch(args.swarm_overwatch, refresh_interval=1.0)
+        start_overwatch(args.swarm_overwatch, refresh_interval=2.0)
         return
 
     # Enforce GPU availability and free memory threshold before proceeding
@@ -204,7 +204,6 @@ def main():
 
     # read configuration from yaml
     exp_config = None
-    exp_dir = args.exp_dir or DEFAULT_DIR
     if args.swarm_server and (not args.conf):
         args.conf = os.path.abspath(
             os.path.join(
@@ -215,14 +214,18 @@ def main():
             "Please provide a valid config file for swarm server mode."
         )
     if args.conf:
+        exp_dir = args.exp_dir or DEFAULT_DIR
         yaml_path = args.conf
         (
             main_yaml_fp,
             exe_exp_base,
             exp_name,
             exp_config,
         ) = prepare_experiment_config(
-            yaml_path, exp_dir, args.backbone, storage=(not args.swarm_server)
+            yaml_path=yaml_path,
+            exp_base_dir=exp_dir,
+            backbone=args.backbone,
+            storage=(not args.swarm_server)
         )
 
     # setup environment variables

ajet/swarm_cli.py

@@ -61,7 +61,10 @@ def cmd_start(args):
         exp_name,
         exp_config,
     ) = prepare_experiment_config(
-        yaml_path, exp_dir, "verl", storage=False
+        yaml_path=yaml_path,
+        exp_base_dir=exp_dir,
+        backbone="verl",
+        storage=False
     )
 
     # Setup environment variables
@@ -73,7 +76,6 @@ def __init__(self, conf, backbone, exp_dir):
             self.swarm_server = True
             self.swarm_overwatch = ""
             self.debug = ""
-
     swarm_args = SwarmArgs(args.conf, "verl", args.exp_dir)
     env, exp_config = setup_environment_vars(swarm_args, exp_config, main_yaml_fp)
 
@@ -131,9 +133,9 @@ def main():
     parser_overwatch.add_argument(
         "--refresh-interval",
         type=float,
-        default=1.0,
+        default=2.0,
         required=False,
-        help="Refresh interval in seconds (default: 1.0)",
+        help="Refresh interval in seconds (default: 2.0)",
     )
     parser_overwatch.set_defaults(func=cmd_overwatch)