feat: Implement TinkerBackend and integrate with existing architecture

- Added TinkerBackend class to support Tinker functionality.
- Updated __init__.py files to include TinkerBackend in the module exports.
- Enhanced TinkerService to require tinker_args and improved renderer name handling.
- Introduced backend.py for TinkerBackend implementation, including service management and renderer configuration.

Author: bradhilton

dev/yes-no-maybe.py

@@ -40,27 +40,27 @@ def with_quotes(w: str) -> str:
 async def main():
     load_dotenv()
 
-    backend = LocalBackend(in_process=True)
+    backend = art.TinkerBackend()
     global model
     base_model = os.environ.get("BASE_MODEL", "Qwen/Qwen3-30B-A3B-Instruct-2507")
     model = art.TrainableModel(
         name=os.environ.get("MODEL_NAME", "012"),
         project="yes-no-maybe",
         base_model=base_model,
-        _internal_config=art.dev.InternalModelConfig(
-            # engine_args=art.dev.EngineArgs(
-            #     max_lora_rank=1,
-            # ),
-            # peft_args=art.dev.PeftArgs(
-            #     r=1,
-            # ),
-            tinker_args=art.dev.TinkerArgs(
-                renderer_name="qwen3_instruct",
-                training_client_args=art.dev.TinkerTrainingClientArgs(
-                    rank=1,
-                ),
-            ),
-        ),
+        # _internal_config=art.dev.InternalModelConfig(
+        #     # engine_args=art.dev.EngineArgs(
+        #     #     max_lora_rank=1,
+        #     # ),
+        #     # peft_args=art.dev.PeftArgs(
+        #     #     r=1,
+        #     # ),
+        #     tinker_args=art.dev.TinkerArgs(
+        #         renderer_name="qwen3_instruct",
+        #         training_client_args=art.dev.TinkerTrainingClientArgs(
+        #             rank=1,
+        #         ),
+        #     ),
+        # ),
     )
     await model.register(backend)
 

src/art/__init__.py

@@ -45,6 +45,7 @@ def __init__(self, **kwargs):
 from .gather import gather_trajectories, gather_trajectory_groups
 from .model import Model, TrainableModel
 from .serverless import ServerlessBackend
+from .tinker import TinkerBackend
 from .trajectories import Trajectory, TrajectoryGroup
 from .types import Messages, MessagesAndChoices, Tools, TrainConfig
 from .utils import retry
@@ -66,6 +67,7 @@ def __init__(self, **kwargs):
     "TrainableModel",
     "retry",
     "TrainConfig",
+    "TinkerBackend",
     "Trajectory",
     "TrajectoryGroup",
     "capture_yielded_trajectory",

src/art/loss.py

@@ -3,9 +3,10 @@
 from pydantic import BaseModel, ConfigDict
 import torch
 
-from art import dev
 from art.utils.group_aggregate import group_aggregate
 
+from . import dev
+
 if TYPE_CHECKING:
     from art.unsloth.service import TrainInputs
 

src/art/tinker/__init__.py

@@ -0,0 +1,3 @@
+from .backend import TinkerBackend
+
+__all__ = ["TinkerBackend"]

src/art/tinker/backend.py

@@ -0,0 +1,101 @@
+import os
+
+from mp_actors import move_to_child_process
+
+from ..local.backend import LocalBackend
+from ..local.service import ModelService
+from ..model import TrainableModel
+from ..utils.output_dirs import get_model_dir
+
+
+class TinkerBackend(LocalBackend):
+    def __init__(
+        self,
+        *,
+        tinker_api_key: str | None = None,
+        in_process: bool = False,
+        path: str | None = None,
+    ) -> None:
+        if not "TINKER_API_KEY" in os.environ or tinker_api_key is not None:
+            assert tinker_api_key is not None, (
+                "TINKER_API_KEY is not set and no tinker_api_key was provided"
+            )
+            print("Setting TINKER_API_KEY to", tinker_api_key, "in environment")
+            os.environ["TINKER_API_KEY"] = tinker_api_key
+        super().__init__(in_process=in_process, path=path)
+
+    async def _get_service(self, model: TrainableModel) -> ModelService:
+        from ..dev.get_model_config import get_model_config
+        from ..dev.model import TinkerArgs
+        from .service import TinkerService
+
+        if model.name not in self._services:
+            config = get_model_config(
+                base_model=model.base_model,
+                output_dir=get_model_dir(model=model, art_path=self._path),
+                config=model._internal_config,
+            )
+            config["tinker_args"] = config.get("tinker_args") or TinkerArgs(
+                renderer_name=get_renderer_name(model.base_model)
+            )
+            self._services[model.name] = TinkerService(
+                model_name=model.name,
+                base_model=model.base_model,
+                config=config,
+                output_dir=get_model_dir(model=model, art_path=self._path),
+            )
+            if not self._in_process:
+                self._services[model.name] = move_to_child_process(
+                    self._services[model.name],
+                    process_name="tinker-service",
+                )
+        return self._services[model.name]
+
+
+renderer_name_message = """
+To manually specify a renderer (and silence this message), you can set the "renderer_name" field like so:
+
+model = art.TrainableModel(
+    name="my-model",
+    project="my-project",
+    base_model="Qwen/Qwen3-8B",
+    _internal_config=art.dev.InternalModelConfig(
+        tinker_args=art.dev.TinkerArgs(renderer_name="qwen3_disable_thinking"),
+    ),
+)
+
+Valid renderer names are:
+
+- llama3
+- qwen3
+- qwen3_disable_thinking
+- qwen3_instruct
+- deepseekv3
+- deepseekv3_disable_thinking
+- gpt_oss_no_sysprompt
+- gpt_oss_low_reasoning
+- gpt_oss_medium_reasoning
+- gpt_oss_high_reasoning
+""".strip()
+
+
+def get_renderer_name(base_model: str) -> str:
+    if base_model.startswith("meta-llama/"):
+        return "llama3"
+    elif base_model.startswith("Qwen/Qwen3-"):
+        if "Instruct" in base_model:
+            return "qwen3_instruct"
+        else:
+            print("Defaulting to Qwen3 renderer without thinking for", base_model)
+            print(renderer_name_message)
+            return "qwen3_disable_thinking"
+    elif base_model.startswith("deepseek-ai/DeepSeek-V3"):
+        print("Defaulting to DeepSeekV3 renderer without thinking for", base_model)
+        print(renderer_name_message)
+        return "deepseekv3_disable_thinking"
+    elif base_model.startswith("openai/gpt-oss"):
+        print("Defaulting to GPT-OSS renderer without system prompt for", base_model)
+        print(renderer_name_message)
+        return "gpt_oss_no_sysprompt"
+    else:
+        raise ValueError(f"Unknown base model: {base_model}")

src/art/tinker/service.py

@@ -195,7 +195,8 @@ def _state_task(self) -> asyncio.Task["TinkerState"]:
         return asyncio.create_task(self._get_state())
 
     async def _get_state(self) -> "TinkerState":
-        config = self.config.get("tinker_args") or {"renderer_name": "qwen3_instruct"}
+        config = self.config.get("tinker_args")
+        assert config is not None, "Tinker args are required"
         service_client = tinker.ServiceClient()
         rest_client = service_client.create_rest_client()
         checkpoint_dir = self._get_last_checkpoint_dir()
@@ -204,9 +205,7 @@ async def _get_state(self) -> "TinkerState":
             with log_timing("Creating Tinker training client from checkpoint"):
                 training_client = await service_client.create_training_client_from_state_with_optimizer_async(
                     path=info["state_with_optimizer_path"],
-                    user_metadata=(self.config.get("tinker_args") or {}).get(
-                        "user_metadata", None
-                    ),
+                    user_metadata=config.get("user_metadata", None),
                 )
             with log_timing("Creating Tinker sampling client from checkpoint"):
                 sampler_client = await training_client.create_sampling_client_async(
@@ -229,7 +228,7 @@ async def _get_state(self) -> "TinkerState":
             training_client=training_client,
             sampler_client=sampler_client,
             renderer=renderers.get_renderer(
-                name=config.get("renderer_name"),
+                name=config["renderer_name"],
                 tokenizer=tokenizer_utils.get_tokenizer(self.base_model),
             ),
         )

src/art/utils/trajectory_logging.py

@@ -15,7 +15,7 @@
 from openai.types.chat.chat_completion import Choice
 from openai.types.chat.chat_completion_message_param import ChatCompletionMessageParam
 
-from art import Trajectory, TrajectoryGroup
+from art.trajectories import Trajectory, TrajectoryGroup
 
 
 def _flatten_message(msg: dict) -> dict:

src/art/utils/trajectory_migration.py

@@ -17,8 +17,7 @@
 
 import yaml
 
-from art import Trajectory, TrajectoryGroup
-from art.trajectories import History
+from art.trajectories import History, Trajectory, TrajectoryGroup
 from art.types import Choice, Message, MessageOrChoice
 
 # ============================================================================