Better vllm deletion

unsloth/dataprep/synthetic.py

@@ -26,6 +26,7 @@ import time
 from unsloth_zoo.vllm_utils import (
     load_vllm,
     patch_vllm,
+    delete_vllm,
 )
 import numpy as np
 
@@ -189,6 +190,14 @@ class SyntheticDataKit:
         for _ in range(10):
             torch.cuda.empty_cache()
             gc.collect()
+
+        # Delete vLLM module as well
+        # We delete llm.llm_engine.model_executor, so first make it accessible
+        class Dummy0: model_executor = 1
+        class Dummy1: llm_engine = Dummy0()
+        class Dummy2: llm = Dummy1()
+        llm = Dummy2().llm.llm_engine.model_executor
+        delete_vllm(llm)
     pass
 
     def __enter__(self): return self

unsloth/models/_utils.py

@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-__version__ = "2025.4.7"
+__version__ = "2025.4.8"
 
 __all__ = [
     "SUPPORTS_BFLOAT16",