fix: preserve _internal_config during Pydantic deserialization

The _internal_config field was being lost when TrainableModel was
deserialized from JSON (e.g., when sent from client to SkyPilot backend).
This is because Pydantic ignores fields starting with underscore during
model_validate().

Added a model_validator(mode="wrap") that extracts _internal_config from
the input data before validation and sets it after the model is created.

This fixes the "Cannot request more than 0 logprobs" error when using
_internal_config.engine_args with remote backends.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <[email protected]>

Author: JRMeyer

src/art/model.py

@@ -1,8 +1,17 @@
-from typing import TYPE_CHECKING, Generic, Iterable, Optional, TypeVar, cast, overload
+from typing import (
+    TYPE_CHECKING,
+    Any,
+    Generic,
+    Iterable,
+    Optional,
+    TypeVar,
+    cast,
+    overload,
+)
 
 import httpx
 from openai import AsyncOpenAI, DefaultAsyncHttpxClient
-from pydantic import BaseModel
+from pydantic import BaseModel, model_validator
 from typing_extensions import Never
 
 from . import dev
@@ -279,6 +288,19 @@ def __init__(
             # Bypass BaseModel __setattr__ to allow setting private attr
             object.__setattr__(self, "_internal_config", _internal_config)
 
+    @model_validator(mode="wrap")
+    @classmethod
+    def _preserve_internal_config(
+        cls, data: Any, handler: Any
+    ) -> "TrainableModel[ModelConfig]":
+        internal_config = None
+        if isinstance(data, dict) and "_internal_config" in data:
+            internal_config = data.pop("_internal_config")
+        model = handler(data)
+        if internal_config is not None:
+            object.__setattr__(model, "_internal_config", internal_config)
+        return model
+
     @overload
     def __new__(
         cls,

src/art/serverless/backend.py

@@ -133,7 +133,9 @@ async def _train_model(
                 epsilon_high=dev_config.get("epsilon_high"),
                 importance_sampling_level=dev_config.get("importance_sampling_level"),
                 learning_rate=config.learning_rate,
-                max_negative_advantage_importance_sampling_weight=dev_config.get("max_negative_advantage_importance_sampling_weight"),
+                max_negative_advantage_importance_sampling_weight=dev_config.get(
+                    "max_negative_advantage_importance_sampling_weight"
+                ),
                 ppo=dev_config.get("ppo"),
                 precalculate_logprobs=dev_config.get("precalculate_logprobs"),
                 scale_rewards=dev_config.get("scale_rewards"),