v0.5.5

Highlights

Performance Update

• We introduced a new mode that schedule multiple GPU steps in advance, reducing CPU overhead (#7000, #7387, #7452, #7703). Initial result shows 20% improvements in QPS for a single GPU running 8B and 30B models. You can set --num-scheduler-steps 8 as a parameter to the API server (via vllm serve) or AsyncLLMEngine. We are working on expanding the coverage to LLM class and aiming to turning it on by default
• Various enhancements:
  • Use flashinfer sampling kernel when avaiable, leading to 7% decoding throughput speedup (#7137)
  • Reduce Python allocations, leading to 24% throughput speedup (#7162, 7364)
  • Improvements to the zeromq based decoupled frontend (#7570, #7716, #7484)

Model Support

• Support Jamba 1.5 (#7415, #7601, #6739)
• Support for the first audio model UltravoxModel (#7615, #7446)
• Improvements to vision models:
  • Support image embeddings as input (#6613)
  • Support SigLIP encoder and alternative decoders for LLaVA models (#7153)
• Support loading GGUF model (#5191) with tensor parallelism (#7520)
• Progress in encoder decoder models: support for serving encoder/decoder models (#7258), and architecture for cross-attention (#4942)

Hardware Support

• AMD: Add fp8 Linear Layer for rocm (#7210)
• Enhancements to TPU support: load time W8A16 quantization (#7005), optimized rope (#7635), and support multi-host inference (#7457).
• Intel: various refactoring for worker, executor, and model runner (#7686, #7712)

Others

• Optimize prefix caching performance (#7193)
• Speculative decoding
  • Use target model max length as default for draft model (#7706)
  • EAGLE Implementation with Top-1 proposer (#6830)
• Entrypoints
  • A new chat method in the LLM class (#5049)
  • Support embeddings in the run_batch API (#7132)
  • Support prompt_logprobs in Chat Completion (#7453)
• Quantizations
  • Expand MoE weight loading + Add Fused Marlin MoE Kernel (#7527)
  • Machete - Hopper Optimized Mixed Precision Linear Kernel (#7174)
• torch.compile: register custom ops for kernels (#7591, #7594, #7536)

What's Changed

• [ci][frontend] deduplicate tests by @youkaichao in #7101
• [Doc] [SpecDecode] Update MLPSpeculator documentation by @tdoublep in #7100
• [Bugfix] Specify device when loading LoRA and embedding tensors by @jischein in #7129
• [MISC] Use non-blocking transfer in prepare_input by @comaniac in #7172
• [Core] Support loading GGUF model by @Isotr0py in #5191
• [Build] Add initial conditional testing spec by @simon-mo in #6841
• [LoRA] Relax LoRA condition by @jeejeelee in #7146
• [Model] Support SigLIP encoder and alternative decoders for LLaVA models by @DarkLight1337 in #7153
• [BugFix] Fix DeepSeek remote code by @dsikka in #7178
• [ BugFix ] Fix ZMQ when VLLM_PORT is set by @robertgshaw2-neuralmagic in #7205
• [Bugfix] add gguf dependency by @kpapis in #7198
• [SpecDecode] [Minor] Fix spec decode sampler tests by @LiuXiaoxuanPKU in #7183
• [Kernel] Add per-tensor and per-token AZP epilogues by @ProExpertProg in #5941
• [Core] Optimize evictor-v2 performance by @xiaobochen123 in #7193
• [Core] Subclass ModelRunner to support cross-attention & encoder sequences (towards eventual encoder/decoder model support) by @afeldman-nm in #4942
• [Bugfix] Fix GPTQ and GPTQ Marlin CPU Offloading by @mgoin in #7225
• [BugFix] Overhaul async request cancellation by @njhill in #7111
• [Doc] Mock new dependencies for documentation by @ywang96 in #7245
• [BUGFIX]: top_k is expected to be an integer. by @Atllkks10 in #7227
• [Frontend] Gracefully handle missing chat template and fix CI failure by @DarkLight1337 in #7238
• [distributed][misc] add specialized method for cuda platform by @youkaichao in #7249
• [Misc] Refactor linear layer weight loading; introduce BasevLLMParameter and weight_loader_v2 by @dsikka in #5874
• [ BugFix ] Move zmq frontend to IPC instead of TCP by @robertgshaw2-neuralmagic in #7222
• Fixes typo in function name by @rafvasq in #7275
• [Bugfix] Fix input processor for InternVL2 model by @Isotr0py in #7164
• [OpenVINO] migrate to latest dependencies versions by @ilya-lavrenov in #7251
• [Doc] add online speculative decoding example by @stas00 in #7243
• [BugFix] Fix frontend multiprocessing hang by @maxdebayser in #7217
• [Bugfix][FP8] Fix dynamic FP8 Marlin quantization by @mgoin in #7219
• [ci] Make building wheels per commit optional by @khluu in #7278
• [Bugfix] Fix gptq failure on T4s by @LucasWilkinson in #7264
• [FrontEnd] Make merge_async_iterators is_cancelled arg optional by @njhill in #7282
• [Doc] Update supported_hardware.rst by @mgoin in #7276
• [Kernel] Fix Flashinfer Correctness by @LiuXiaoxuanPKU in #7284
• [Misc] Fix typos in scheduler.py by @ruisearch42 in #7285
• [Frontend] remove max_num_batched_tokens limit for lora by @NiuBlibing in #7288
• [Bugfix] Fix LoRA with PP by @andoorve in #7292
• [Model] Rename MiniCPMVQwen2 to MiniCPMV2.6 by @jeejeelee in #7273
• [Bugfix][Kernel] Increased atol to fix failing tests by @ProExpertProg in #7305
• [Frontend] Kill the server on engine death by @joerunde in #6594
• [Bugfix][fast] Fix the get_num_blocks_touched logic by @zachzzc in #6849
• [Doc] Put collect_env issue output in a block by @mgoin in #7310
• [CI/Build] Dockerfile.cpu improvements by @dtrifiro in #7298
• [Bugfix] Fix new Llama3.1 GGUF model loading by @Isotr0py in #7269
• [Misc] Temporarily resolve the error of BitAndBytes by @jeejeelee in #7308
• Add Skywork AI as Sponsor by @simon-mo in #7314
• [TPU] Add Load-time W8A16 quantization for TPU Backend by @lsy323 in #7005
• [Core] Support serving encoder/decoder models by @DarkLight1337 in #7258
• [TPU] Fix dockerfile.tpu by @WoosukKwon in #7331
• [Performance] Optimize e2e overheads: Reduce python allocations by @alexm-neuralmagic in #7162
• [Bugfix] Fix speculative decoding with MLPSpeculator with padded vocabulary by @tjohnson31415 in #7218
• [Speculative decoding] [Multi-Step] decouple should_modify_greedy_probs_inplace by @SolitaryThinker in #6971
• [Core] Streamline stream termination in AsyncLLMEngine by @njhill in #7336
• [Model][Jamba] Mamba cache single buffer by @mzusman in #6739
• [VLM][Doc] Add stop_token_ids to InternVL example by @Isotr0py in #7354
• [Performance] e2e overheads reduction: Small followup diff by @alexm-neuralmagic in #7364
• [Bugfix] Fix reinit procedure in ModelInputForGPUBuilder by @alexm-neuralmagic in #7360
• [Frontend] Support embeddings in the run_batch API by @pooyadavoodi in #7132
• [Bugfix] Fix ITL recording in serving benchmark by @ywang96 in #7372
• [Core] Add span metrics for model_forward, scheduler and sampler time by @sfc-gh-mkeralapura in #7089
• [Bugfix] Fix PerTensorScaleParameter weight loading for fused models by @dsikka in #7376
• [Misc] Add numpy implementation of compute_slot_mapping by @Yard1 in #7377
• [Core] Fix edge case in chunked prefill + block manager v2 by @cadedaniel in #7380
• [Bugfix] Fix phi3v batch inference when images have different aspect ratio by @Isotr0py in #7392
• [TPU] Use mark_dynamic to reduce compilation time by @WoosukKwon in #7340
• Updating LM Format Enforcer version to v0.10.6 by @noamgat in #7189
• [core] [2/N] refactor worker_base input preparation for multi-step by @SolitaryThinker in #7387
• [CI/Build] build on empty device for better dev experience by @tomeras91 in #4773
• [Doc] add instructions about building vLLM with VLLM_TARGET_DEVICE=empty by @tomeras91 in #7403
• [misc] add commit id in collect env by @youkaichao in #7405
• [Docs] Update readme by @simon-mo in #7316
• [CI/Build] Minor refactoring for vLLM assets by @ywang96 in #7407
• [Kernel] Flashinfer correctness fix for v0.1.3 by @LiuXiaoxuanPKU in #7319
• [Core][VLM] Support image embeddings as input by @ywang96 in #6613
• [Frontend] Disallow passing model as both argument and option by @DarkLight1337 in #7347
• [CI/Build] bump Dockerfile.neuron image base, use public ECR by @dtrifiro in #6832
• [Bugfix] Fix logit soft cap in flash-attn backend by @WoosukKwon in #7425
• [ci] Entrypoints run upon changes in vllm/ by @khluu in #7423
• [ci] Cancel fastcheck run when PR is marked ready by @khluu in #7427
• [ci] Cancel fastcheck when PR is ready by @khluu in #7433
• [Misc] Use scalar type to dispatch to different gptq_marlin kernels by @LucasWilkinson in #7323
• [Core] Consolidate GB constant and enable float GB arguments by @DarkLight1337 in #7416
• [Core/Bugfix] Add FP8 K/V Scale and dtype conversion for prefix/prefill Triton Kernel by @jon-chuang in #7208
• [Bugfix] Handle PackageNotFoundError when checking for xpu version by @sasha0552 in #7398
• [CI/Build] bump minimum cmake version by @dtrifiro in #6999
• [Core] Shut down aDAG workers with clean async llm engine exit by @ruisearch42 in #7224
• [mypy] Misc. typing improvements by @DarkLight1337 in #7417
• [Misc] improve logits processors logging message by @aw632 in #7435
• [ci] Remove fast check cancel workflow by @khluu in #7455
• [Bugfix] Fix weight loading for Chameleon when TP>1 by @DarkLight1337 in #7410
• [hardware] unify usage of is_tpu to current_platform.is_tpu() by @youkaichao in #7102
• [TPU] Suppress import custom_ops warning by @WoosukKwon in #7458
• Revert "[Doc] Update supported_hardware.rst (#7276)" by @WoosukKwon in #7467
• [Frontend][Core] Add plumbing to support audio language models by @petersalas in #7446
• [Misc] Update LM Eval Tolerance by @dsikka in #7473
• [Misc] Update gptq_marlin to use new vLLMParameters by @dsikka in #7281
• [Misc] Update Fused MoE weight loading by @dsikka in #7334
• [Misc] Update awq and awq_marlin to use vLLMParameters by @dsikka in #7422
• Announce NVIDIA Meetup by @simon-mo in #7483
• [frontend] spawn engine process from api server process by @youkaichao in #7484
• [Misc] compressed-tensors code reuse by @kylesayrs in #7277
• [misc][plugin] add plugin system implementation by @youkaichao in #7426
• [TPU] Support multi-host inference by @WoosukKwon in #7457
• [Bugfix][CI] Import ray under guard by @WoosukKwon in #7486
• [CI/Build]Reduce the time consumption for LoRA tests by @jeejeelee in #7396
• [misc][ci] fix cpu test with plugins by @youkaichao in #7489
• [Bugfix][Docs] Update list of mock imports by @DarkLight1337 in #7493
• [doc] update test script to include cudagraph by @youkaichao in #7501
• Fix empty output when temp is too low by @CatherineSue in #2937
• [ci] fix model tests by @youkaichao in #7507
• [Bugfix][Frontend] Disable embedding API for chat models by @QwertyJack in #7504
• [Misc] Deprecation Warning when setting --engine-use-ray by @wallashss in #7424
• [VLM][Core] Support profiling with multiple multi-modal inputs per prompt by @DarkLight1337 in #7126
• [core] [3/N] multi-step args and sequence.py by @SolitaryThinker in #7452
• [TPU] Set per-rank XLA cache by @WoosukKwon in #7533
• [Misc] Revert compressed-tensors code reuse by @kylesayrs in #7521
• llama_index serving integration documentation by @pavanjava in #6973
• [Bugfix][TPU] Correct env variable for XLA cache path by @WoosukKwon in #7544
• [Bugfix] update neuron for version > 0.5.0 by @omrishiv in #7175
• [Misc] Update dockerfile for CPU to cover protobuf installation by @PHILO-HE in #7182
• [Bugfix] Fix default weight loading for scalars by @mgoin in #7534
• [Bugfix][Harmless] Fix hardcoded float16 dtype for model_is_embedding by @mgoin in #7566
• [Misc] Add quantization config support for speculative model. by @ShangmingCai in #7343
• [Feature]: Add OpenAI server prompt_logprobs support #6508 by @gnpinkert in #7453
• [ci/test] rearrange tests and make adag test soft fail by @youkaichao in #7572
• Chat method for offline llm by @nunjunj in #5049
• [CI] Move quantization cpu offload tests out of fastcheck by @mgoin in #7574
• [Misc/Testing] Use torch.testing.assert_close by @jon-chuang in #7324
• register custom op for flash attn and use from torch.ops by @youkaichao in #7536
• [Core] Use uvloop with zmq-decoupled front-end by @njhill in #7570
• [CI] Fix crashes of performance benchmark by @KuntaiDu in #7500
• [Bugfix][Hardware][AMD][Frontend] add quantization param to embedding checking method by @gongdao123 in #7513
• support tqdm in notebooks by @fzyzcjy in #7510
• [Feature][Hardware][Amd] Add fp8 Linear Layer for Rocm by @charlifu in #7210
• [Kernel] W8A16 Int8 inside FusedMoE by @mzusman in #7415
• [Kernel] Add tuned triton configs for ExpertsInt8 by @mgoin in #7601
• [spec decode] [4/N] Move update_flash_attn_metadata to attn backend by @SolitaryThinker in #7571
• [Core] Fix tracking of model forward time to the span traces in case of PP>1 by @sfc-gh-mkeralapura in #7440
• [Doc] Add docs for llmcompressor INT8 and FP8 checkpoints by @mgoin in #7444
• [Doc] Update quantization supported hardware table by @mgoin in #7595
• [Kernel] register punica functions as torch ops by @bnellnm in #7591
• [Kernel][Misc] dynamo support for ScalarType by @bnellnm in #7594
• [Kernel] fix types used in aqlm and ggml kernels to support dynamo by @bnellnm in #7596
• [Model] Align nemotron config with final HF state and fix lm-eval-small by @mgoin in #7611
• [Bugfix] Fix custom_ar support check by @bnellnm in #7617
• .[Build/CI] Enabling passing AMD tests. by @Alexei-V-Ivanov-AMD in #7610
• [Bugfix] Clear engine reference in AsyncEngineRPCServer by @ruisearch42 in #7618
• [aDAG] Unflake aDAG + PP tests by @rkooo567 in #7600
• [Bugfix] add >= 1.0 constraint for openai dependency by @metasyn in #7612
• [misc] use nvml to get consistent device name by @youkaichao in #7582
• [ci][test] fix engine/logger test by @youkaichao in #7621
• [core][misc] update libcudart finding by @youkaichao in #7620
• [Model] Pipeline parallel support for JAIS by @mrbesher in #7603
• [ci][test] allow longer wait time for api server by @youkaichao in #7629
• [Misc]Fix BitAndBytes exception messages by @jeejeelee in #7626
• [VLM] Refactor MultiModalConfig initialization and profiling by @ywang96 in #7530
• [TPU] Skip creating empty tensor by @WoosukKwon in #7630
• [TPU] Use mark_dynamic only for dummy run by @WoosukKwon in #7634
• [TPU] Optimize RoPE forward_native2 by @WoosukKwon in #7636
• [ Bugfix ] Fix Prometheus Metrics With zeromq Frontend by @robertgshaw2-neuralmagic in #7279
• [CI/Build] Add text-only test for Qwen models by @alex-jw-brooks in #7475
• [Misc] Refactor Llama3 RoPE initialization by @WoosukKwon in #7637
• [Core] Optimize SPMD architecture with delta + serialization optimization by @rkooo567 in #7109
• [Core] Use flashinfer sampling kernel when available by @peng1999 in #7137
• fix xpu build by @jikunshang in #7644
• [Misc] Remove Gemma RoPE by @WoosukKwon in #7638
• [MISC] Add prefix cache hit rate to metrics by @comaniac in #7606
• [Bugfix] fix lora_dtype value type in arg_utils.py - part 2 by @c3-ali in #5428
• [core] Multi Step Scheduling by @SolitaryThinker in #7000
• [Core] Support tensor parallelism for GGUF quantization by @Isotr0py in #7520
• [Bugfix] Don't disable existing loggers by @a-ys in #7664
• [TPU] Fix redundant input tensor cloning by @WoosukKwon in #7660
• [Bugfix] use StoreBoolean instead of type=bool for --disable-logprobs-during-spec-decoding by @tjohnson31415 in #7665
• [doc] fix doc build error caused by msgspec by @youkaichao in #7659
• [Speculative Decoding] Fixing hidden states handling in batch expansion by @abhigoyal1997 in #7508
• [ci] Install Buildkite test suite analysis by @khluu in #7667
• [Bugfix] support tie_word_embeddings for all models by @zijian-hu in #5724
• [CI] Organizing performance benchmark files by @KuntaiDu in #7616
• [misc] add nvidia related library in collect env by @youkaichao in #7674
• [XPU] fallback to native implementation for xpu custom op by @jianyizh in #7670
• [misc][cuda] add warning for pynvml user by @youkaichao in #7675
• [Core] Refactor executor classes to make it easier to inherit GPUExecutor by @jikunshang in #7673
• [Kernel] (1/N) Machete - Hopper Optimized Mixed Precision Linear Kernel by @LucasWilkinson in #7174
• [OpenVINO] Updated documentation by @ilya-lavrenov in #7687
• [VLM][Model] Add test for InternViT vision encoder by @Isotr0py in #7409
• [Hardware] [Intel GPU] refactor xpu worker/executor by @jikunshang in #7686
• [CI/Build] Pin OpenTelemetry versions and make availability errors clearer by @ronensc in #7266
• [Misc] Add jinja2 as an explicit build requirement by @LucasWilkinson in #7695
• [Core] Add AttentionState abstraction by @Yard1 in #7663
• [Intel GPU] fix xpu not support punica kernel (which use torch.library.custom_op) by @jikunshang in #7685
• [ci][test] adjust max wait time for cpu offloading test by @youkaichao in #7709
• [Core] Pipe worker_class_fn argument in Executor by @Yard1 in #7707
• [ci] try to log process using the port to debug the port usage by @youkaichao in #7711
• [Model] Add AWQ quantization support for InternVL2 model by @Isotr0py in #7187
• [Doc] Section for Multimodal Language Models by @ywang96 in #7719
• [mypy] Enable following imports for entrypoints by @DarkLight1337 in #7248
• [Bugfix] Mirror jinja2 in pyproject.toml by @sasha0552 in #7723
• [BugFix] Avoid premature async generator exit and raise all exception variations by @njhill in #7698
• [BUG] fix crash on flashinfer backend with cudagraph disabled, when attention group_size not in [1,2,4,8] by @learninmou in #7509
• [Bugfix][Hardware][CPU] Fix mm_limits initialization for CPU backend by @Isotr0py in #7735
• [Spec Decoding] Use target model max length as default for draft model by @njhill in #7706
• [Bugfix] chat method add_generation_prompt param by @brian14708 in #7734
• [Bugfix][Frontend] Fix Issues Under High Load With zeromq Frontend by @robertgshaw2-neuralmagic in #7394
• [Bugfix] Pass PYTHONPATH from setup.py to CMake by @sasha0552 in #7730
• [multi-step] Raise error if not using async engine by @SolitaryThinker in #7703
• [Frontend] Improve Startup Failure UX by @robertgshaw2-neuralmagic in #7716
• [misc] Add Torch profiler support by @SolitaryThinker in #7451
• [Model] Add UltravoxModel and UltravoxConfig by @petersalas in #7615
• [ci] [multi-step] narrow multi-step test dependency paths by @SolitaryThinker in #7760
• [Kernel] Expand MoE weight loading + Add Fused Marlin MoE Kernel by @dsikka in #7527
• [distributed][misc] error on same VLLM_HOST_IP setting by @youkaichao in #7756
• [AMD][CI/Build] Disambiguation of the function call for ROCm 6.2 headers compatibility by @gshtras in #7477
• [Kernel] Replaced blockReduce[...] functions with cub::BlockReduce by @ProExpertProg in #7233
• [Model] Fix Phi-3.5-vision-instruct 'num_crops' issue by @zifeitong in #7710
• [Bug][Frontend] Improve ZMQ client robustness by @joerunde in #7443
• Revert "[Kernel] Expand MoE weight loading + Add Fused Marlin MoE Kernel (#7527)" by @mgoin in #7764
• [TPU] Avoid initializing TPU runtime in is_tpu by @WoosukKwon in #7763
• [ci] refine dependency for distributed tests by @youkaichao in #7776
• [Misc] Use torch.compile for GemmaRMSNorm by @WoosukKwon in #7642
• [Speculative Decoding] EAGLE Implementation with Top-1 proposer by @abhigoyal1997 in #6830
• Fix ShardedStateLoader for vllm fp8 quantization by @sfc-gh-zhwang in #7708
• [Bugfix] Don't build machete on cuda <12.0 by @LucasWilkinson in #7757
• [Misc] update fp8 to use vLLMParameter by @dsikka in #7437
• [Bugfix] spec decode handle None entries in topk args in create_sequence_group_output by @tjohnson31415 in #7232
• [Misc] Enhance prefix-caching benchmark tool by @Jeffwan in #6568
• [Doc] Fix incorrect docs from #7615 by @petersalas in #7788
• [Bugfix] Use LoadFormat values as choices for vllm serve --load-format by @mgoin in #7784
• [ci] Cleanup & refactor Dockerfile to pass different Python versions and sccache bucket via build args by @khluu in #7705
• [Misc] fix typo in triton import warning by @lsy323 in #7794
• [Frontend] error suppression cleanup by @joerunde in #7786
• [Ray backend] Better error when pg topology is bad. by @rkooo567 in #7584
• [Hardware][Intel GPU] refactor xpu_model_runner, fix xpu tensor parallel by @jikunshang in #7712
• [misc] Add Torch profiler support for CPU-only devices by @DamonFool in #7806
• [BugFix] Fix server crash on empty prompt by @maxdebayser in #7746
• [github][misc] promote asking llm first by @youkaichao in #7809
• [Misc] Update marlin to use vLLMParameters by @dsikka in #7803
• Bump version to v0.5.5 by @simon-mo in #7823

New Contributors

• @jischein made their first contribution in #7129
• @kpapis made their first contribution in #7198
• @xiaobochen123 made their first contribution in #7193
• @Atllkks10 made their first contribution in #7227
• @stas00 made their first contribution in #7243
• @maxdebayser made their first contribution in #7217
• @NiuBlibing made their first contribution in #7288
• @lsy323 made their first contribution in #7005
• @pooyadavoodi made their first contribution in #7132
• @sfc-gh-mkeralapura made their first contribution in #7089
• @jon-chuang made their first contribution in #7208
• @aw632 made their first contribution in #7435
• @petersalas made their first contribution in #7446
• @kylesayrs made their first contribution in #7277
• @QwertyJack made their first contribution in #7504
• @wallashss made their first contribution in #7424
• @pavanjava made their first contribution in #6973
• @PHILO-HE made their first contribution in #7182
• @gnpinkert made their first contribution in #7453
• @gongdao123 made their first contribution in #7513
• @charlifu made their first contribution in #7210
• @metasyn made their first contribution in #7612
• @mrbesher made their first contribution in #7603
• @alex-jw-brooks made their first contribution in #7475
• @a-ys made their first contribution in #7664
• @zijian-hu made their first contribution in #5724
• @jianyizh made their first contribution in #7670
• @learninmou made their first contribution in #7509
• @brian14708 made their first contribution in #7734
• @sfc-gh-zhwang made their first contribution in #7708

Full Changelog: v0.5.4...v0.5.5