feat: migrate to vllm AsyncLLMEngine with async generation support (#77)

fix: incorrect double use of timing variable (#71)

chore: minor updates on logging and resource allocation (#73)

feat: add fp16 training (#70)

fix: micro batch calculation in offline DPO (#68)

fix: incorrect state indexing in PPOMultiTurnLearner critic training (#67)

chore: update lora and add metrics (#66)

feat: support LoRA RL training (#64)

feat: support turn-level ppo for general agentic rl (#63)

fix: truncated importance sampling to handle precision mismatch (#62)

chore: updates for online preference learning (#61)

fix: use semantic version comparison for vLLM version check (#60)

fix examples

update readme

feat: refactor SFT to support multi turn chat data (#59)