Mock vllm to bypass TRL missing module error

ultimate_sota_training.py

@@ -104,6 +104,13 @@ import transformers.utils.hub
 if not hasattr(transformers.utils.hub, "TRANSFORMERS_CACHE"):
     transformers.utils.hub.TRANSFORMERS_CACHE = "/tmp"
 
+# CRITICAL FIX for vllm crash:
+# TRL's latest versions unconditionally import vllm in their extras module.
+# Since Unsloth uses its own fast inference backend, we mock vllm to prevent the crash.
+import sys
+from unittest.mock import MagicMock
+sys.modules['vllm'] = MagicMock()
+
 from trl import GRPOConfig, GRPOTrainer
 from unsloth import FastLanguageModel