v0.9.1

What's Changed

🚀 Features

• feature: enable tool_call and reasoning_content parsing for qwen3 by @ywx217 in #3615
• Support Mooncake migration backend for PD disaggregation by @Risc-lt in #3620
• Support load fused moe weights by @RunningLeon in #3672
• Seperate api_server and pytorch engine into different processors by @grimoire in #3627
• add reward model api by @CUHKSZzxy in #3665

💥 Improvements

• [ascend]import patch at initiazing time by @JackWeiw in #3662
• [ascend]use custon transdata in python kernel by @JackWeiw in #3671
• move import transformers in patch by @grimoire in #3660
• set ray envs by @grimoire in #3643
• raise ImportError when enable ep and not install dlblas by @zhaochaoxing in #3636
• Reduce sampling memory usage by @lzhangzz in #3666

🐞 Bug fixes

• fix dockerfile by @lvhan028 in #3657
• Fix top-p only sampling with padded vocab size by @lzhangzz in #3661
• fix pt engine stop & cancel by @irexyc in #3681
• Fix convert bf16 to numpy by @RunningLeon in #3686
• disable torch.compile in cuda graph runner by @grimoire in #3691
• fix reward model api by @CUHKSZzxy in #3703

📚 Documentations

• add reward model documents by @CUHKSZzxy in #3706

🌐 Other

• upgrade torch and triton by @grimoire in #3677
• support do_preprocess=False for chat.completions by @irexyc in #3645
• [ci] change flash atten installation in pr test by @zhulinJulia24 in #3688
• fix profile_throughput.py by @irexyc in #3692
• fix profile_generation.py by @irexyc in #3707
• update dlblas version in dockerfile by @CUHKSZzxy in #3711
• bump version to v0.9.1 by @lvhan028 in #3685

New Contributors

• @ywx217 made their first contribution in #3615
• @Risc-lt made their first contribution in #3620

Full Changelog: v0.9.0...v0.9.1