Fix vllm error cleanly by creating fake python module structure

ultimate_sota_training.py

@@ -69,7 +69,6 @@ def bootstrap_deps() -> None:
             "trl>=0.18.2,<0.24.0",
             "mergekit",
             "llm-blender",
-            "vllm",
             "weave",
             "wandb",
             "matplotlib",
@@ -105,6 +104,20 @@ import transformers.utils.hub
 if not hasattr(transformers.utils.hub, "TRANSFORMERS_CACHE"):
     transformers.utils.hub.TRANSFORMERS_CACHE = "/tmp"
 
+# CRITICAL FIX for vllm crash:
+# Create a valid Python module on disk to satisfy both importlib and TRL's hard imports.
+import os
+import sys
+vllm_dir = "/tmp/fake_vllm"
+os.makedirs(os.path.join(vllm_dir, "vllm", "distributed", "device_communicators", "pynccl"), exist_ok=True)
+open(os.path.join(vllm_dir, "vllm", "__init__.py"), "w").close()
+open(os.path.join(vllm_dir, "vllm", "distributed", "__init__.py"), "w").close()
+open(os.path.join(vllm_dir, "vllm", "distributed", "device_communicators", "__init__.py"), "w").close()
+with open(os.path.join(vllm_dir, "vllm", "distributed", "device_communicators", "pynccl", "__init__.py"), "w") as f:
+    f.write("class PyNcclCommunicator: pass\n")
+if vllm_dir not in sys.path:
+    sys.path.insert(0, vllm_dir)
+
 from trl import GRPOConfig, GRPOTrainer
 from unsloth import FastLanguageModel