feat: move trajectory logging from backend to frontend (#518)

* feat: move trajectory logging from backend to frontend (Model class)

Implements RFC #511 - moves trajectory persistence and metrics reporting
from the backend to the Model class, allowing backends to focus solely on training.

Key changes:
- Model.log() now handles parquet writing, metrics calculation, history.jsonl
- Model gains base_path and report_metrics attributes for configuration
- TrainableModel.train() calls log() first, then backend._train_model()
- TrainableModel.delete_checkpoints() reads history.jsonl locally
- Backend._log() removed; _delete_checkpoints() renamed to _delete_checkpoint_files()
- Removed scale_learning_rate_by_reward_std_dev handling (users can implement if needed)

Breaking changes:
- Backend interface: _log() removed, _delete_checkpoints() renamed
- ServerlessBackend: trajectories now saved locally instead of sent to API

Closes #511

* fix: sort imports in vllm/server.py

---------

Co-authored-by: Cursor Bot <bot@cursor.com>

Author: corbt

src/art/backend.py

@@ -55,16 +55,14 @@ async def _get_step(self, model: "TrainableModel") -> int:
         response.raise_for_status()
         return response.json()
 
-    async def _delete_checkpoints(
+    async def _delete_checkpoint_files(
         self,
         model: "TrainableModel",
-        benchmark: str,
-        benchmark_smoothing: float,
+        steps_to_keep: list[int],
     ) -> None:
         response = await self._client.post(
-            "/_delete_checkpoints",
-            json=model.safe_model_dump(),
-            params={"benchmark": benchmark, "benchmark_smoothing": benchmark_smoothing},
+            "/_delete_checkpoint_files",
+            json={"model": model.safe_model_dump(), "steps_to_keep": steps_to_keep},
         )
         response.raise_for_status()
 
@@ -82,23 +80,6 @@ async def _prepare_backend_for_training(
         base_url, api_key = tuple(response.json())
         return base_url, api_key
 
-    async def _log(
-        self,
-        model: "Model",
-        trajectory_groups: list[TrajectoryGroup],
-        split: str = "val",
-    ) -> None:
-        response = await self._client.post(
-            "/_log",
-            json={
-                "model": model.safe_model_dump(),
-                "trajectory_groups": [tg.model_dump() for tg in trajectory_groups],
-                "split": split,
-            },
-            timeout=None,
-        )
-        response.raise_for_status()
-
     async def _train_model(
         self,
         model: "TrainableModel",

src/art/cli.py

@@ -173,7 +173,13 @@ async def art_error_handler(request: Request, exc: ARTError):
     app.post("/close")(backend.close)
     app.post("/register")(backend.register)
     app.post("/_get_step")(backend._get_step)
-    app.post("/_delete_checkpoints")(backend._delete_checkpoints)
+
+    @app.post("/_delete_checkpoint_files")
+    async def _delete_checkpoint_files(
+        model: TrainableModel = Body(...),
+        steps_to_keep: list[int] = Body(...),
+    ):
+        await backend._delete_checkpoint_files(model, steps_to_keep)
 
     @app.post("/_prepare_backend_for_training")
     async def _prepare_backend_for_training(
@@ -182,13 +188,7 @@ async def _prepare_backend_for_training(
     ):
         return await backend._prepare_backend_for_training(model, config)
 
-    @app.post("/_log")
-    async def _log(
-        model: Model,
-        trajectory_groups: list[TrajectoryGroup],
-        split: str = Body("val"),
-    ):
-        await backend._log(model, trajectory_groups, split)
+    # Note: /_log endpoint removed - logging now handled by frontend (Model.log())
 
     @app.post("/_train_model")
     async def _train_model(