v0.9.2rc2

What's Changed

• [Kernel] Enable fp8 support for pplx and BatchedTritonExperts. by @bnellnm in #18864
• [Misc] Fix Unable to detect current VLLM config. Defaulting to NHD kv cache layout warning by @NickLucche in #20400
• [Bugfix] Register reducer even if transformers_modules not available by @eicherseiji in #19510
• Change warn_for_unimplemented_methods to debug by @mgoin in #20455
• [Platform] Add custom default max tokens by @gmarinho2 in #18557
• Add ignore consolidated file in mistral example code by @princepride in #20420
• [Misc] small update by @reidliu41 in #20462
• [Structured Outputs][V1] Skipping with models doesn't contain tokenizers by @aarnphm in #20365
• [Perf] Optimize Vectorization Utils for Int 8 Quantization Kernels by @yewentao256 in #20331
• [Misc] Add SPDX-FileCopyrightText by @jeejeelee in #20428
• Support Llama 4 for fused_marlin_moe by @mgoin in #20457
• [Bug][Frontend] Fix structure of transcription's decoder_prompt by @sangbumlikeagod in #18809
• [Model][3/N] Automatic conversion of CrossEncoding model by @noooop in #20168
• [Doc] Fix classification table in list of supported models by @DarkLight1337 in #20489
• [CI] add kvcache-connector dependency definition and add into CI build by @panpan0000 in #18193
• [Misc] Small: Remove global media connector. Each test should have its own test connector object. by @huachenheli in #20395
• Enable V1 for Hybrid SSM/Attention Models by @tdoublep in #20016
• [feat]: CUTLASS block scaled group gemm for SM100 by @djmmoss in #19757
• [CI Bugfix] Fix pre-commit failures on main by @mgoin in #20502
• [Doc] fix mutltimodal_inputs.md gh examples link by @GuyStone in #20497
• [Misc] Add security warning for development mode endpoints by @reidliu41 in #20508
• [doc] small fix by @reidliu41 in #20506
• [Misc] Remove the unused LoRA test code by @jeejeelee in #20494
• Fix unknown attribute of topk_indices_dtype in CompressedTensorsW8A8Fp8MoECutlassMethod by @luccafong in #20507
• [v1] Re-add fp32 support to v1 engine through FlexAttention by @Isotr0py in #19754
• [Misc] Add logger.exception for TPU information collection failures by @reidliu41 in #20510
• [Misc] remove unused import by @reidliu41 in #20517
• test_attention compat with coming xformers change by @bottler in #20487
• [BUG] Fix #20484. Support empty sequence in cuda penalty kernel by @vadiklyutiy in #20491
• [Bugfix] Fix missing per_act_token parameter in compressed_tensors_moe by @luccafong in #20509
• [BugFix] Fix: ImportError when building on hopper systems by @LucasWilkinson in #20513
• [TPU][Bugfix] fix the MoE OOM issue by @yaochengji in #20339
• [Frontend] Support image object in llm.chat by @sfeng33 in #19635
• [Benchmark] Add support for multiple batch size benchmark through CLI in benchmark_moe.py + Add Triton Fused MoE kernel config for FP8 E=16 on B200 by @b8zhong in #20516
• [Misc] call the pre-defined func by @reidliu41 in #20518
• [V0 deprecation] Remove V0 CPU/XPU/TPU backends by @WoosukKwon in #20412
• [V1] Support any head size for FlexAttention backend by @DarkLight1337 in #20467
• [BugFix][Spec Decode] Fix spec token ids in model runner by @WoosukKwon in #20530
• [Bugfix] Add use_cross_encoder flag to use correct activation in ClassifierPooler by @DarkLight1337 in #20527

New Contributors

• @sangbumlikeagod made their first contribution in #18809
• @djmmoss made their first contribution in #19757
• @GuyStone made their first contribution in #20497
• @bottler made their first contribution in #20487

Full Changelog: v0.9.2rc1...v0.9.2rc2