Simplify wandb log and update default trainer config (#22)

Author: pan-x-c

README.md

@@ -246,7 +246,7 @@ trinity run --config <config_path>
 
 
 
-For example, below is the command for fine-tuning Qwen-2.5-1B-Instruct on GSM8k dataset using GRPO algorithm:
+For example, below is the command for fine-tuning Qwen-2.5-1.5B-Instruct on GSM8k dataset using GRPO algorithm:
 
 ```shell
 trinity run --config examples/grpo_gsm8k/gsm8k.yaml

docs/sphinx_doc/source/main.md

@@ -226,7 +226,7 @@ trinity run --config <config_path>
 
 
 
-For example, below is the command for fine-tuning Qwen-2.5-1B-Instruct on GSM8k dataset using GRPO algorithm:
+For example, below is the command for fine-tuning Qwen-2.5-1.5B-Instruct on GSM8k dataset using GRPO algorithm:
 
 ```shell
 trinity run --config examples/grpo_gsm8k/gsm8k.yaml

docs/sphinx_doc/source/tutorial/example_dpo.md

@@ -6,7 +6,7 @@ This example describes DPO based on the Qwen-2.5-1.5B-Instruct model and [Human-
 
 ### Model Preparation
 
-Download the Qwen-2.5-1B-Instruct model to the local directory `$MODEL_PATH/Qwen2.5-1.5B-Instruct`:
+Download the Qwen-2.5-1.5B-Instruct model to the local directory `$MODEL_PATH/Qwen2.5-1.5B-Instruct`:
 
 ```shell
 # Using Modelscope

docs/sphinx_doc/source/tutorial/example_reasoning_basic.md

@@ -7,7 +7,7 @@ This example shows how to run RFT with the Qwen-2.5-1.5B-Instruct model and GSM8
 
 **Model Preparation.**
 
-Download the Qwen-2.5-1B-Instruct model to the local directory `$MODEL_PATH/Qwen2.5-1.5B-Instruct`:
+Download the Qwen-2.5-1.5B-Instruct model to the local directory `$MODEL_PATH/Qwen2.5-1.5B-Instruct`:
 
 ```bash
 # Using Modelscope

pyproject.toml

@@ -34,6 +34,7 @@ dependencies = [
     "fire",
     "flask",
     "requests",
+    "tensorboard",
 ]
 
 [project.scripts]

trinity/buffer/reader/file_reader.py

@@ -17,7 +17,7 @@
 
 
 class FileReader(BufferReader):
-    """Reader of the Queue buffer."""
+    """Reader of the File buffer."""
 
     def __init__(self, meta: DatasetConfig, config: BufferConfig) -> None:
         assert meta.storage_type == StorageType.FILE

trinity/cli/launcher.py

@@ -102,9 +102,8 @@ def both(config: Config) -> None:
                 logger.error(e)
                 logger.error("Evaluation failed.")
                 raise e
-
-        ray.get(explorer.log_finalize.remote(step=explore_iter_num))
-        ray.get(trainer.log_finalize.remote(step=train_iter_num))
+        ray.get(explorer.flush_log.remote(step=explore_iter_num))
+        ray.get(trainer.flush_log.remote(step=train_iter_num))
 
 
 def activate_data_module(data_workflow_url: str, config_path: str):

trinity/common/config.py

@@ -173,8 +173,7 @@ class ExplorerConfig:
 @dataclass
 class TrainerConfig:
     trainer_type: str = "verl"
-    trainer_data_type: str = "RFT"
-    trainer_config_path: str = "examples/ppo_countdown/train_countdown.yaml"
+    trainer_config_path: str = ""
     eval_interval: int = 100
     enable_preview: bool = True  # enable rollout preview in wandb
     trainer_config: Any = None
@@ -185,16 +184,6 @@ class TrainerConfig:
     # warmup config
     sft_warmup_iteration: int = 0
 
-    def __post_init__(self):
-        if self.trainer_type == "verl":
-            from trinity.common.verl_config import load_config
-
-            if not os.path.isfile(self.trainer_config_path):
-                raise ValueError(f"Invalid trainer config path: {self.trainer_config_path}")
-            self.trainer_config = load_config(self.trainer_config_path)
-        else:
-            raise ValueError(f"Invalid trainer type: {self.trainer_type}")
-
 
 @dataclass
 class MonitorConfig:
@@ -285,6 +274,15 @@ def _check_buffer(self) -> None:
 
     def check_and_update(self) -> None:
         """Check and update the config."""
+        if self.trainer.trainer_type == "verl":
+            from trinity.common.verl_config import load_config
+
+            if not os.path.isfile(self.trainer.trainer_config_path):
+                raise ValueError(f"Invalid trainer config path: {self.trainer.trainer_config_path}")
+            self.trainer.trainer_config = load_config(self.trainer.trainer_config_path)
+        else:
+            raise ValueError(f"Invalid trainer type: {self.trainer_type}")
+
         # check mode
         if self.mode not in ["explore", "train", "both"]:
             raise ValueError(f"Invalid mode: {self.mode}")

trinity/explorer/explorer.py

@@ -216,6 +216,9 @@ def explore_step(self) -> Tuple[bool, int]:
 
     def eval(self) -> bool:
         """Evaluation on all evaluation data samples."""
+        if self.eval_taskset is None:
+            self.logger.warning("No evaluation data samples. Skip evaluation.")
+            return True
         self.logger.info("Evaluation started.")
         st = time.time()
         all_metrics = defaultdict(list)
@@ -248,6 +251,6 @@ def sync_weight(self) -> None:
         else:  # online weights update
             self._online_weights_update()
 
-    def log_finalize(self, step: int) -> None:
-        """Commit the logging results to wandb"""
-        self.monitor.log({"dummy_log_explorer": step}, step=step, commit=True)
+    def flush_log(self, step: int) -> None:
+        """Flush the log of the current step."""
+        self.monitor.log({}, step=step, commit=True)

trinity/trainer/trainer.py

@@ -105,9 +105,9 @@ def sync_weight(self) -> None:
         if self.config.synchronizer.sync_method == "online":
             self.engine.sync_weight()
 
-    def log_finalize(self, step: int) -> None:
-        """Commit the logging results to wandb"""
-        self.engine.logger.log({"dummy_log_trainer": step}, step=step, commit=True)
+    def flush_log(self, step: int) -> None:
+        """Flush the log of the current step."""
+        self.engine.logger.log({}, step=step, commit=True)
 
 
 class TrainEngineWrapper(ABC):