refactor: update backend imports and improve model registration in TinkerService

- Changed import of TinkerBackend from art.local to art.tinker.
- Refactored model registration logic in TinkerService to use a models dictionary instead of sampling clients.
- Enhanced checkpoint handling and model registration for better clarity and maintainability.
- Updated response handling in the Proxy class to use integer IDs instead of UUIDs for requests and responses.

Author: bradhilton

dev/yes-no-maybe.py

@@ -6,7 +6,7 @@
 import openai
 
 import art
-from art.local import LocalBackend
+from art.tinker import TinkerBackend
 
 
 async def rollout(client: openai.AsyncOpenAI, prompt: str) -> art.Trajectory:
@@ -40,7 +40,7 @@ def with_quotes(w: str) -> str:
 async def main():
     load_dotenv()
 
-    backend = art.TinkerBackend()
+    backend = TinkerBackend()
     global model
     base_model = os.environ.get("BASE_MODEL", "Qwen/Qwen3-30B-A3B-Instruct-2507")
     model = art.TrainableModel(

src/art/tinker/prefix_cache.py

@@ -28,10 +28,9 @@ def __init__(self, max_entries: int = 1024) -> None:
 
     @staticmethod
     def _encode_tokens(tokens: Sequence[int]) -> str:
-        packed = bytearray()
-        for token in tokens:
-            packed.extend(struct.pack(">I", token))
-        return packed.hex()
+        if not tokens:
+            return ""
+        return struct.pack(f">{len(tokens)}I", *tokens).hex()
 
     def lookup(self, rendered_tokens: Sequence[int]) -> PrefixEntry | None:
         key = self._encode_tokens(rendered_tokens)