refactor: Simplify training client arguments handling in TinkerBackend and TinkerService

- Removed TinkerTrainingClientArgs import and adjusted the configuration to use a default empty dictionary for training client arguments.
- Ensured that default values for 'rank' and 'train_unembed' are set within TinkerService when creating the training client.

Author: bradhilton

src/art/tinker/backend.py

@@ -26,7 +26,7 @@ def __init__(
 
     async def _get_service(self, model: TrainableModel) -> ModelService:
         from ..dev.get_model_config import get_model_config
-        from ..dev.model import TinkerArgs, TinkerTrainingClientArgs
+        from ..dev.model import TinkerArgs
         from .service import TinkerService
 
         if model.name not in self._services:
@@ -38,10 +38,8 @@ async def _get_service(self, model: TrainableModel) -> ModelService:
             config["tinker_args"] = config.get("tinker_args") or TinkerArgs(
                 renderer_name=get_renderer_name(model.base_model)
             )
-            config["tinker_args"]["training_client_args"] = config["tinker_args"].get(
-                "training_client_args"
-            ) or TinkerTrainingClientArgs(
-                rank=8,
+            config["tinker_args"]["training_client_args"] = (
+                config["tinker_args"].get("training_client_args") or {}
             )
             self._services[model.name] = TinkerService(
                 model_name=model.name,

src/art/tinker/service.py

@@ -213,10 +213,15 @@ async def _get_state(self) -> "TinkerState":
                 )
         else:
             with log_timing("Creating Tinker training client"):
+                training_client_args = config.get("training_client_args", {})
+                if "rank" not in training_client_args:
+                    training_client_args["rank"] = 8
+                if "train_unembed" not in training_client_args:
+                    training_client_args["train_unembed"] = False
                 training_client = (
                     await service_client.create_lora_training_client_async(
                         base_model=self.base_model,
-                        **config.get("training_client_args", {}),
+                        **training_client_args,
                     )
                 )
             sampler_client = await self._save_checkpoint(