Describe the bug
The example from docs fails on AttributeError:
for metric in self.llm.get_metrics():
^^^^^^^^^^^^^^^^^^^^
AttributeError: 'AsyncLLM' object has no attribute 'get_metrics'
This is caused by the fact that self.llm is an instance of AsyncLLM, not of LLMEngine (which contains get_metrics method). The AsyncLLM is hardcoded and even after removing hardcoded override it fails on some downstream assert telling that I have to use async.
Steps/Code to reproduce bug
Follow steps from the setup guide and the single-node guide.
Expected behavior
The training should not crash!
Additional context
None