Replace step-0 overwrite with _state cache invalidation

The real issue is that UnslothService._state (a cached_property)
may be initialized before the fork copies the checkpoint, caching
the base model instead of the forked weights. Invalidating the
cache after fork ensures the trainer picks up the forked checkpoint
on next access.

The step-0 overwrite was unnecessary — vLLM's start_openai_server
already calls get_last_checkpoint_dir() which finds the forked
checkpoint at its original step number.

Author: arcticfly

src/art/local/backend.py

@@ -1434,16 +1434,18 @@ async def _experimental_fork_checkpoint(
 
         shutil.copytree(source_checkpoint_dir, dest_checkpoint_dir)
 
-        # Also overwrite the initial empty checkpoint at step 0 so vLLM
-        # loads the forked weights on startup (it uses @0 by default)
-        step0_dir = get_step_checkpoint_dir(dest_model_dir, 0)
-        if os.path.exists(step0_dir) and step0_dir != dest_checkpoint_dir:
+        # Invalidate the UnslothService _state cache so the trainer
+        # re-initializes with the forked checkpoint instead of the base model.
+        # _state is a cached_property that reads get_last_checkpoint_dir() on
+        # first access; if it was accessed before the fork, it cached the base
+        # model and will never pick up the forked weights.
+        service = await self._get_service(model)
+        if hasattr(service, "_state") and "_state" in service.__dict__:
+            del service.__dict__["_state"]
             if verbose:
                 print(
-                    f"Overwriting initial checkpoint at {step0_dir} with forked weights"
+                    "Invalidated UnslothService _state cache to pick up forked checkpoint"
                 )
-            shutil.rmtree(step0_dir)
-            shutil.copytree(dest_checkpoint_dir, step0_dir)
 
         if verbose:
             print(