vLLM V0 to V1: Correctness Before Corrections in RL

This post details the migration from vLLM V0 to V1, focusing on ensuring backend correctness for RL training. Key fixes include aligning logprob semantics (using `processed_logprobs`), managing runtime defaults (disabling prefix caching and async scheduling), handling inflight weight updates by matching V0's behavior of keeping cached state intact, and ensuring fp32 precision for the `lm_head` for final projection. These fixes were crucial to eliminate train-inference mismatches and achieve parity with the V0 reference.