v0.9.2

What's Changed

🚀 Features

• [Feature] metrics support by @CUHKSZzxy in #3534
• Relax FP8 TP requirement by @lzhangzz in #3697
• FA3 by @zhaochaoxing in #3623
• support qwen2/2.5-vl in turbomind by @irexyc in #3744
• feat: add pytorch_engine_qwen2_5vl_sm120 by @kolmogorov-quyet in #3750
• Internvl pt by @RunningLeon in #3765
• Improve internvl for turbomind engine by @lvhan028 in #3769

💥 Improvements

• Refactor linear by @grimoire in #3653
• remove python3.8 support and add python3.13 support by @lvhan028 in #3638
• refactor vl inputs split by @grimoire in #3699
• [Fix]: Replace mutable default with default_factory for scheduler_stats by @ConvolutedDog in #3730
• Fix the logic of calculating max_new_tokens and determining finish_reason by @lvhan028 in #3727
• Override HF config.json via CLI by @CUHKSZzxy in #3722
• feat(build): Integrate and build turbomind backend directly in setup.py by @windreamer in #3726
• Generate the benchmark output filename with given arguments by @lvhan028 in #3740
• Make loading llm without vlm as an option by @grimoire in #3745

🐞 Bug fixes

• add ray to ascend requirements by @sigma-plus in #3713
• fix accessing undefined attribute seq_aux of deepseek-r1-0528 by @lvhan028 in #3728
• [Fix]: Avoid quantize qk norm for qwen3 dense models by @taishan1994 in #3733
• fix py313 env creation failed when building lmdeploy-builder image by @lvhan028 in #3739
• [Fix]: kernel meta retrieval for SM7X does not work by @xiaoajie738 in #3746
• limit max_session_len by @grimoire in #3751
• fix internvl norm by @grimoire in #3756
• support qwen3 moe yarn and vlm hf_overrides by @grimoire in #3757
• [PD Disaggregation] fix double unshelf by @JimyMa in #3762
• fix(build): fix version parse regex to support post-release versions by @windreamer in #3764
• adapt transformers>=v4.52.0 to loading qwen2.5-vl with turbomind by @irexyc in #3771
• fix chat template with tool call by @RunningLeon in #3773
• fix vl nothink mode by @RunningLeon in #3776

📚 Documentations

• update reward model docs by @CUHKSZzxy in #3721

🌐 Other

• update twomicrobatch by @SHshenhao in #3651
• [CI]: Upgrade to py310 for ut by @RunningLeon in #3718
• [ci] update dailytest environment and scripts by @zhulinJulia24 in #3716
• Preliminary Blackwell (sm_120a, RTX 50 series) support by @lzhangzz in #3701
• [ci] add fp8 evaluation workflow by @zhulinJulia24 in #3729
• Add VRAM bandwidth utilization stat to attention test by @lzhangzz in #3731
• doc: fix dead links to MindX DL to recover CI. by @windreamer in #3741
• fix free cache in MPEngine branch by @JimyMa in #3670
• fix: make RelWithDebInfo default cmake build type by @windreamer in #3774
• bump version to v0.9.2 by @lvhan028 in #3770

New Contributors

• @sigma-plus made their first contribution in #3713
• @ConvolutedDog made their first contribution in #3730
• @windreamer made their first contribution in #3726
• @taishan1994 made their first contribution in #3733
• @xiaoajie738 made their first contribution in #3746
• @kolmogorov-quyet made their first contribution in #3750

Full Changelog: v0.9.1...v0.9.2