v0.6.5

Highlights

• Significant progress on the V1 engine refactor and multimodal support: New model executable interfaces for text-only and multimodal models, multiprocessing, improved configuration handling, and profiling enhancements (#10374, #10570, #10699, #11074, #11076, #10382, #10665, #10564, #11125, #11185, #11242).
• Major improvements in torch.compile integration: Support for all attention backends, encoder-based models, dynamic FP8 fusion, shape specialization fixes, and performance optimizations (#10558, #10613, #10121, #10383, #10399, #10406, #10437, #10460, #10552, #10622, #10722, #10620, #10906, #11108, #11059, #11005, #10838, #11081, #11110).
• Expanded model support, including Aria, Cross Encoders, GLM-4, OLMo November 2024, Telechat2, LoRA improvements and multimodal Granite models (#10514, #10400, #10561, #10503, #10311, #10291, #9057, #10418, #5064).
• Use xgrammar as the default guided decoding backend (#10785)
• Improved hardware enablement for AMD ROCm, ARM AARCH64, TPU prefix caching, XPU AWQ/GPTQ, and various CPU/Gaudi/HPU/NVIDIA enhancements (#10254, #9228, #10307, #10107, #10667, #10565, #10239, #11016, #9735, #10355, #10700).
• Note: Changed default temperature for ChatCompletionRequest from 0.7 to 1.0 to align with OpenAI (#11219)

Model Support

• Added Aria (#10514), Cross Encoder (#10400), GLM-4 (#10561), OLMo (#10503), Telechat2 (#10311), Cohere R7B (#11203), GritLM embeddings (#10816)
• LoRA support for Internlm2, glm-4v, Pixtral-HF (#5064, #10418, #10795).
• Improved quantization (BNB, bitsandbytes) for multiple models (#10795, #10842, #10682, #10549)
• Expanded multimodal support (#10291, #11142).

Hardware Support

• AMD ROCm GGUF quantization (#10254), ARM AARCH64 enablement (#9228), TPU prefix caching (#10307), XPU AWQ/GPTQ (#10107), CPU/Gaudi/HPU enhancements (#10355, #10667, #10565, #10239, #11016, #9735, #10541, #10394, #10700).

Performance & Scheduling

• Prefix-cache aware scheduling (#10128), sliding window support (#10462), disaggregated prefill enhancements (#10502, #10884), evictor optimization (#7209).

Benchmark & Frontend

• Benchmark structured outputs and vision datasets (#10804, #10557, #10880, #10547).
• Frontend: Automatic chat format detection (#9919), input_audio support (#11027), CLI --version (#10369), extra fields in requests (#10463).

Documentation & Plugins

• Architecture overview (#10368), Helm chart (#9199), KubeAI integration (#10837), plugin system docs (#10372), disaggregated prefilling (#11197), structured outputs (#9943), usage section (#10827).

Bugfixes & Misc

• Updated defaults for chunked prefill (#10544)
• Add GH200 support (#11212, #11244)

What's Changed

• Add default value to avoid Falcon crash (#5363) by @wchen61 in #10347
• [Misc] Fix import error in tensorizer tests and cleanup some code by @DarkLight1337 in #10349
• [Doc] Remove float32 choice from --lora-dtype by @xyang16 in #10348
• [Bugfix] Fix fully sharded LoRA bug by @jeejeelee in #10352
• [Misc] Fix some help info of arg_utils to improve readability by @ShangmingCai in #10362
• [core][misc] keep compatibility for old-style classes by @youkaichao in #10356
• [Bugfix] Ensure special tokens are properly filtered out for guided structured output with MistralTokenizer by @gcalmettes in #10363
• [Misc] Bump up test_fused_moe tolerance by @ElizaWszola in #10364
• [Misc] bump mistral common version by @simon-mo in #10367
• [Docs] Add Nebius as sponsors by @simon-mo in #10371
• [Frontend] Add --version flag to CLI by @russellb in #10369
• [Doc] Move PR template content to docs by @russellb in #10159
• [Docs] Misc updates to TPU installation instructions by @mikegre-google in #10165
• [Frontend] Automatic detection of chat content format from AST by @DarkLight1337 in #9919
• [doc] add doc for the plugin system by @youkaichao in #10372
• [misc][plugin] improve log messages by @youkaichao in #10386
• [BugFix] [Kernel] Fix GPU SEGV occuring in fused_moe kernel by @rasmith in #10385
• [Misc] Update benchmark to support image_url file or http by @kakao-steve-ai in #10287
• [Misc] Medusa supports custom bias by @skylee-01 in #10361
• [Bugfix] Fix M-RoPE position calculation when chunked prefill is enabled by @imkero in #10388
• [V1] Add code owners for V1 by @WoosukKwon in #10397
• [2/N][torch.compile] make compilation cfg part of vllm cfg by @youkaichao in #10383
• [V1] Refactor model executable interface for all text-only language models by @ywang96 in #10374
• [CI/Build] Fix IDC hpu [Device not found] issue by @xuechendi in #10384
• [Bugfix][Hardware][CPU] Fix CPU embedding runner with tensor parallel by @Isotr0py in #10394
• [platforms] refactor cpu code by @youkaichao in #10402
• [Hardware] [HPU]add mark_step for hpu by @jikunshang in #10239
• [Bugfix] Fix mrope_position_delta in non-last prefill chunk by @imkero in #10403
• [Misc] Enhance offline_inference to support user-configurable paramet… by @wchen61 in #10392
• [Misc] Add uninitialized params tracking for AutoWeightsLoader by @Isotr0py in #10327
• [Bugfix] Ignore ray reinit error when current platform is ROCm or XPU by @HollowMan6 in #10375
• [4/N][torch.compile] clean up set_torch_compile_backend by @youkaichao in #10401
• [VLM] Report multi_modal_placeholders in output by @lk-chen in #10407
• [Model] Remove redundant softmax when using PoolingType.STEP by @Maybewuss in #10415
• [Model][LoRA]LoRA support added for glm-4v by @B-201 in #10418
• [Model] Remove transformers attention porting in VITs by @Isotr0py in #10414
• [Doc] Update doc for LoRA support in GLM-4V by @B-201 in #10425
• [5/N][torch.compile] torch.jit.script --> torch.compile by @youkaichao in #10406
• [Doc] Add documentation for Structured Outputs by @ismael-dm in #9943
• Fix open_collective value in FUNDING.yml by @andrew in #10426
• [Model][Bugfix] Support TP for PixtralHF ViT by @mgoin in #10405
• [Hardware][XPU] AWQ/GPTQ support for xpu backend by @yma11 in #10107
• [Kernel] Explicitly specify other value in tl.load calls by @angusYuhao in #9014
• [Kernel] Initial Machete W4A8 support + Refactors by @LucasWilkinson in #9855
• [3/N][torch.compile] consolidate custom op logging by @youkaichao in #10399
• [ci][bugfix] fix kernel tests by @youkaichao in #10431
• [misc] Allow partial prefix benchmarking & random input generation for prefix benchmarking by @rickyyx in #9929
• [ci/build] Have dependabot ignore all patch update by @khluu in #10436
• [Bugfix]Fix Phi-3 BNB online quantization by @jeejeelee in #10417
• [Platform][Refactor] Extract func get_default_attn_backend to Platform by @MengqingCao in #10358
• Add openai.beta.chat.completions.parse example to structured_outputs.rst by @mgoin in #10433
• [Bugfix] Guard for negative counter metrics to prevent crash by @tjohnson31415 in #10430
• [Misc] Avoid misleading warning messages by @jeejeelee in #10438
• [Doc] Add the start of an arch overview page by @russellb in #10368
• [misc][plugin] improve plugin loading by @youkaichao in #10443
• [CI][CPU] adding numa node number as container name suffix by @zhouyuan in #10441
• [BugFix] Fix hermes tool parser output error stream arguments in some cases (#10395) by @xiyuan-lee in #10398
• [Pixtral-Large] Pixtral actually has no bias in vision-lang adapter by @patrickvonplaten in #10449
• Fix: Build error seen on Power Architecture by @mikejuliet13 in #10421
• [Doc] fix link for page that was renamed by @russellb in #10455
• [6/N] to...

v0.6.4.post1

This patch release covers bug fixes (#10347, #10349, #10348, #10352, #10363), keep compatibility for vLLMConfig usage in out of tree models (#10356)

What's Changed

• Add default value to avoid Falcon crash (#5363) by @wchen61 in #10347
• [Misc] Fix import error in tensorizer tests and cleanup some code by @DarkLight1337 in #10349
• [Doc] Remove float32 choice from --lora-dtype by @xyang16 in #10348
• [Bugfix] Fix fully sharded LoRA bug by @jeejeelee in #10352
• [Misc] Fix some help info of arg_utils to improve readability by @ShangmingCai in #10362
• [core][misc] keep compatibility for old-style classes by @youkaichao in #10356
• [Bugfix] Ensure special tokens are properly filtered out for guided structured output with MistralTokenizer by @gcalmettes in #10363
• [Misc] Bump up test_fused_moe tolerance by @ElizaWszola in #10364
• [Misc] bump mistral common version by @simon-mo in #10367

New Contributors

• @wchen61 made their first contribution in #10347

Full Changelog: v0.6.4...v0.6.4.post1

v0.6.4

Highlights

• Significant progress in V1 engine core refactor (#9826, #10135, #10288, #10211, #10225, #10228, #10268, #9954, #10272, #9971, #10224, #10166, #9289, #10058, #9888, #9972, #10059, #9945, #9679, #9871, #10227, #10245, #9629, #10097, #10203, #10148). You can checkout more details regarding the design and plan ahead in our recent meetup slides
• Signficant progress in torch.compile support. Many models now support torch compile with TorchInductor. You can checkout our meetup slides for more details. (#9775, #9614, #9639, #9641, #9876, #9946, #9589, #9896, #9637, #9300, #9947, #9138, #9715, #9866, #9632, #9858, #9889)

Model Support

• New LLMs and VLMs: Idefics3 (#9767), H2OVL-Mississippi (#9747), Qwen2-Audio (#9248), Pixtral models in the HF Transformers format (#9036), FalconMamba (#9325), Florence-2 language backbone (#9555)
• New encoder-decoder embedding models: BERT (#9056), RoBERTa & XLM-RoBERTa (#9387)
• Expanded task support: Llama embeddings (#9806), Math-Shepherd (Mistral reward modeling) (#9697), Qwen2 classification (#9704), Qwen2 embeddings (#10184), VLM2Vec (Phi-3-Vision embeddings) (#9303), E5-V (LLaVA-NeXT embeddings) (#9576), Qwen2-VL embeddings (#9944)
  • Add user-configurable --task parameter for models that support both generation and embedding (#9424)
  • Chat-based Embeddings API (#9759)
• Tool calling parser for Granite 3.0 (#9027), Jamba (#9154), granite-20b-functioncalling (#8339)
• LoRA support for Granite 3.0 MoE (#9673), Idefics3 (#10281), Llama embeddings (#10071), Qwen (#9622), Qwen2-VL (#10022)
• BNB quantization support for Idefics3 (#10310), Mllama (#9720), Qwen2 (#9467, #9574), MiniCPMV (#9891)
• Unified multi-modal processor for VLM (#10040, #10044)
• Simplify model interface (#9933, #10237, #9938, #9958, #10007, #9978, #9983, #10205)

Hardware Support

• Gaudi: Add Intel Gaudi (HPU) inference backend (#6143)
• CPU: Add embedding models support for CPU backend (#10193)
• TPU: Correctly profile peak memory usage & Upgrade PyTorch XLA (#9438)
• Triton: Add Triton implementation for scaled_mm_triton to support fp8 and int8 SmoothQuant, symmetric case (#9857)

Performance

• Combine chunked prefill with speculative decoding (#9291)
• fused_moe Performance Improvement (#9384)

Engine Core

• Override HF config.json via CLI (#5836)
• Add goodput metric support (#9338)
• Move parallel sampling out from vllm core, paving way for V1 engine (#9302)
• Add stateless process group for easier integration with RLHF and disaggregated prefill (#10216, #10072)

Others

• Improvements to the pull request experience with DCO, mergify, stale bot, etc. (#9436, #9512, #9513, #9259, #10082, #10285, #9803)
• Dropped support for Python 3.8 (#10038, #8464)
• Basic Integration Test For TPU (#9968)
• Document the class hierarchy in vLLM (#10240), explain the integration with Hugging Face (#10173).
• Benchmark throughput now supports image input (#9851)

What's Changed

• [TPU] Fix TPU SMEM OOM by Pallas paged attention kernel by @WoosukKwon in #9350
• [Frontend] merge beam search implementations by @LunrEclipse in #9296
• [Model] Make llama3.2 support multiple and interleaved images by @xiangxu-google in #9095
• [Bugfix] Clean up some cruft in mamba.py by @tlrmchlsmth in #9343
• [Frontend] Clarify model_type error messages by @stevegrubb in #9345
• [Doc] Fix code formatting in spec_decode.rst by @mgoin in #9348
• [Bugfix] Update InternVL input mapper to support image embeds by @hhzhang16 in #9351
• [BugFix] Fix chat API continuous usage stats by @njhill in #9357
• pass ignore_eos parameter to all benchmark_serving calls by @gracehonv in #9349
• [Misc] Directly use compressed-tensors for checkpoint definitions by @mgoin in #8909
• [Bugfix] Fix vLLM UsageInfo and logprobs None AssertionError with empty token_ids by @CatherineSue in #9034
• [Bugfix][CI/Build] Fix CUDA 11.8 Build by @LucasWilkinson in #9386
• [Bugfix] Molmo text-only input bug fix by @mrsalehi in #9397
• [Misc] Standardize RoPE handling for Qwen2-VL by @DarkLight1337 in #9250
• [Model] VLM2Vec, the first multimodal embedding model in vLLM by @DarkLight1337 in #9303
• [CI/Build] Test VLM embeddings by @DarkLight1337 in #9406
• [Core] Rename input data types by @DarkLight1337 in #8688
• [Misc] Consolidate example usage of OpenAI client for multimodal models by @ywang96 in #9412
• [Model] Support SDPA attention for Molmo vision backbone by @Isotr0py in #9410
• Support mistral interleaved attn by @patrickvonplaten in #9414
• [Kernel][Model] Improve continuous batching for Jamba and Mamba by @mzusman in #9189
• [Model][Bugfix] Add FATReLU activation and support for openbmb/MiniCPM-S-1B-sft by @streaver91 in #9396
• [Performance][Spec Decode] Optimize ngram lookup performance by @LiuXiaoxuanPKU in #9333
• [CI/Build] mypy: Resolve some errors from checking vllm/engine by @russellb in #9267
• [Bugfix][Kernel] Prevent integer overflow in fp8 dynamic per-token quantize kernel by @tlrmchlsmth in #9425
• [BugFix] [Kernel] Fix GPU SEGV occurring in int8 kernels by @rasmith in #9391
• Add notes on the use of Slack by @terrytangyuan in #9442
• [Kernel] Add Exllama as a backend for compressed-tensors by @LucasWilkinson in #9395
• [Misc] Print stack trace using logger.exception by @DarkLight1337 in #9461
• [misc] CUDA Time Layerwise Profiler by @LucasWilkinson in #8337
• [Bugfix] Allow prefill of assistant response when using mistral_common by @sasha0552 in #9446
• [TPU] Call torch._sync(param) during weight loading by @WoosukKwon in #9437
• [Hardware][CPU] compressed-tensor INT8 W8A8 AZP support by @bigPYJ1151 in #9344
• [Core] Deprecating block manager v1 and make block manager v2 default by @KuntaiDu in #8704
• [CI/Build] remove .github from .dockerignore, add dirty repo check by @dtrifiro in #9375
• [Misc] Remove commit id file by @DarkLight1337 in #9470
• [torch.compile] Fine-grained CustomOp enabling mechanism by @ProExpertProg in #9300
• [Bugfix] Fix support for dimension like integers and ScalarType by @bnellnm in #9299
• [Bugfix] Add random_seed to sample_hf_requests in benchmark_serving script by @wukaixingxp in #9013
• [Bugfix] Print warnings related to mistral_common tokenizer only once by @sasha0552 in #9468
• [Hardwware][Neuron] Simplify model load for transformers-neuronx library by @sssrijan-amazon in #9380
• Support BERTModel (first encoder-only embedding model) by @robertgshaw2-neuralmagic in #9056
• [BugFix] Stop silent failures on compressed-tensors parsing by @dsikka in #9381
• [Bugfix][Core] Use torch.cuda.memory_stats() to profile peak memory usage by @joerunde in #9352
• [Qwen2.5] Support bnb quant for Qwen2.5 by @blueyo0 in #9467
• [CI/Build] Use commit hash references for github actions by @russellb in #9430
• [BugFix] Typing fixes to RequestOutput.prompt and beam search by @njhill in #9473
• [Frontend][Feature] Add jamba tool parser by @tomeras91 in #9154
• [BugFix] Fix and simplify completion API usage streaming by @njhill in #9475
• [CI/Build] Fix lint errors in mistral tokenizer by @DarkLight1337 in #9504
• [Bugfix] Fix offline_inference_with_prefix.py by @tlrmchlsmth in #9505
• [Misc] benchmark: Add option to set max concurrency by @russellb in #9390
• [Model] Add user-configurable task for models that support both generation and embedding by @DarkLight1337 in #9424
• [CI/Build] Add error matching config for mypy by @russellb in #9512
• [Model] Support Pixtral models ...

v0.6.3.post1

Highlights

New Models

• Support Ministral 3B and Ministral 8B via interleaved attention (#9414)
• Support multiple and interleaved images for Llama3.2 (#9095)
• Support VLM2Vec, the first multimodal embedding model in vLLM (#9303)

Important bug fix

• Fix chat API continuous usage stats (#9357)
• Fix vLLM UsageInfo and logprobs None AssertionError with empty token_ids (#9034)
• Fix Molmo text-only input bug (#9397)
• Fix CUDA 11.8 Build (#9386)
• Fix _version.py not found issue (#9375)

Other Enhancements

• Remove block manager v1 and make block manager v2 default (#8704)
• Spec Decode Optimize ngram lookup performance (#9333)

What's Changed

• [TPU] Fix TPU SMEM OOM by Pallas paged attention kernel by @WoosukKwon in #9350
• [Frontend] merge beam search implementations by @LunrEclipse in #9296
• [Model] Make llama3.2 support multiple and interleaved images by @xiangxu-google in #9095
• [Bugfix] Clean up some cruft in mamba.py by @tlrmchlsmth in #9343
• [Frontend] Clarify model_type error messages by @stevegrubb in #9345
• [Doc] Fix code formatting in spec_decode.rst by @mgoin in #9348
• [Bugfix] Update InternVL input mapper to support image embeds by @hhzhang16 in #9351
• [BugFix] Fix chat API continuous usage stats by @njhill in #9357
• pass ignore_eos parameter to all benchmark_serving calls by @gracehonv in #9349
• [Misc] Directly use compressed-tensors for checkpoint definitions by @mgoin in #8909
• [Bugfix] Fix vLLM UsageInfo and logprobs None AssertionError with empty token_ids by @CatherineSue in #9034
• [Bugfix][CI/Build] Fix CUDA 11.8 Build by @LucasWilkinson in #9386
• [Bugfix] Molmo text-only input bug fix by @mrsalehi in #9397
• [Misc] Standardize RoPE handling for Qwen2-VL by @DarkLight1337 in #9250
• [Model] VLM2Vec, the first multimodal embedding model in vLLM by @DarkLight1337 in #9303
• [CI/Build] Test VLM embeddings by @DarkLight1337 in #9406
• [Core] Rename input data types by @DarkLight1337 in #8688
• [Misc] Consolidate example usage of OpenAI client for multimodal models by @ywang96 in #9412
• [Model] Support SDPA attention for Molmo vision backbone by @Isotr0py in #9410
• Support mistral interleaved attn by @patrickvonplaten in #9414
• [Kernel][Model] Improve continuous batching for Jamba and Mamba by @mzusman in #9189
• [Model][Bugfix] Add FATReLU activation and support for openbmb/MiniCPM-S-1B-sft by @streaver91 in #9396
• [Performance][Spec Decode] Optimize ngram lookup performance by @LiuXiaoxuanPKU in #9333
• [CI/Build] mypy: Resolve some errors from checking vllm/engine by @russellb in #9267
• [Bugfix][Kernel] Prevent integer overflow in fp8 dynamic per-token quantize kernel by @tlrmchlsmth in #9425
• [BugFix] [Kernel] Fix GPU SEGV occurring in int8 kernels by @rasmith in #9391
• Add notes on the use of Slack by @terrytangyuan in #9442
• [Kernel] Add Exllama as a backend for compressed-tensors by @LucasWilkinson in #9395
• [Misc] Print stack trace using logger.exception by @DarkLight1337 in #9461
• [misc] CUDA Time Layerwise Profiler by @LucasWilkinson in #8337
• [Bugfix] Allow prefill of assistant response when using mistral_common by @sasha0552 in #9446
• [TPU] Call torch._sync(param) during weight loading by @WoosukKwon in #9437
• [Hardware][CPU] compressed-tensor INT8 W8A8 AZP support by @bigPYJ1151 in #9344
• [Core] Deprecating block manager v1 and make block manager v2 default by @KuntaiDu in #8704
• [CI/Build] remove .github from .dockerignore, add dirty repo check by @dtrifiro in #9375

New Contributors

• @gracehonv made their first contribution in #9349
• @streaver91 made their first contribution in #9396

Full Changelog: v0.6.3...v0.6.3.post1

v0.6.3

Highlights

Model Support

• New Models:
  • Text: Granite MoE (#8206), Mamba (#6484, #8533)
  • Vision: GLM-4V (#9242), Molmo (#9016), NVLM-D (#9045)
  • Reward model support: Qwen2.5-Math-RM-72B (#8896)
• Expansion in functionality:
  • Add Gemma2 embedding model (#9004)
  • Support input embeddings for qwen2vl (#8856), minicpmv (#9237)
  • LoRA:
    • LoRA support for MiniCPMV2.5 (#7199), MiniCPMV2.6 (#8943)
    • Expand lora modules for mixtral (#9008)
  • Pipeline parallelism support to remaining text and embedding models (#7168, #9090)
  • Expanded bitsandbytes quantization support for Falcon, OPT, Gemma, Gemma2, and Phi (#9148)
  • Tool use:
    • Add support for Llama 3.1 and 3.2 tool use (#8343)
    • Support tool calling for InternLM2.5 (#8405)
• Out of tree support enhancements: Explicit interface for vLLM models and support OOT embedding models (#9108)

Documentation

• New compatibility matrix for mutual exclusive features (#8512)
• Reorganized installation doc, note that we publish a per-commit docker image (#8931)

Hardware Support:

• Cross-attention and Encoder-Decoder models support on x86 CPU backend (#9089)
• Support AWQ for CPU backend (#7515)
• Add async output processor for xpu (#8897)
• Add on-device sampling support for Neuron (#8746)

Architectural Enhancements

• Progress in vLLM's refactoring to a core core:
  • Spec decode removing batch expansion (#8839, #9298).
  • We have made block manager V2 the default. This is an internal refactoring for cleaner and more tested code path (#8678).
  • Moving beam search from the core to the API level (#9105, #9087, #9117, #8928)
  • Move guided decoding params into sampling params (#8252)
• Torch Compile:
  • You can now set an env var VLLM_TORCH_COMPILE_LEVEL to control torch.compile various levels of compilation control and integration (#9058). Along with various improvements (#8982, #9258, #906, #8875), using VLLM_TORCH_COMPILE_LEVEL=3 can turn on Inductor's full graph compilation without vLLM's custom ops.

Others

• Performance enhancements to turn on multi-step scheeduling by default (#8804, #8645, #8378)
• Enhancements towards priority scheduling (#8965, #8956, #8850)

What's Changed

• [Misc] Update config loading for Qwen2-VL and remove Granite by @ywang96 in #8837
• [Build/CI] Upgrade to gcc 10 in the base build Docker image by @tlrmchlsmth in #8814
• [Docs] Add README to the build docker image by @mgoin in #8825
• [CI/Build] Fix missing ci dependencies by @fyuan1316 in #8834
• [misc][installation] build from source without compilation by @youkaichao in #8818
• [ci] Soft fail Entrypoints, Samplers, LoRA, Decoder-only VLM by @khluu in #8872
• [Bugfix] Include encoder prompts len to non-stream api usage response by @Pernekhan in #8861
• [Misc] Change dummy profiling and BOS fallback warns to log once by @mgoin in #8820
• [Bugfix] Fix print_warning_once's line info by @tlrmchlsmth in #8867
• fix validation: Only set tool_choice auto if at least one tool is provided by @chiragjn in #8568
• [Bugfix] Fixup advance_step.cu warning by @tlrmchlsmth in #8815
• [BugFix] Fix test breakages from transformers 4.45 upgrade by @njhill in #8829
• [Installation] Allow lower versions of FastAPI to maintain Ray 2.9 compatibility by @DarkLight1337 in #8764
• [Feature] Add support for Llama 3.1 and 3.2 tool use by @maxdebayser in #8343
• [Core] Rename PromptInputs and inputs with backward compatibility by @DarkLight1337 in #8876
• [misc] fix collect env by @youkaichao in #8894
• [MISC] Fix invalid escape sequence '' by @panpan0000 in #8830
• [Bugfix][VLM] Fix Fuyu batching inference with max_num_seqs>1 by @Isotr0py in #8892
• [TPU] Update pallas.py to support trillium by @bvrockwell in #8871
• [torch.compile] use empty tensor instead of None for profiling by @youkaichao in #8875
• [Kernel] AQ AZP 4/4: Integrate asymmetric quantization to linear method by @ProExpertProg in #7271
• [Bugfix] fix for deepseek w4a16 by @LucasWilkinson in #8906
• [Core] Multi-Step + Single Step Prefills via Chunked Prefill code path by @varun-sundar-rabindranath in #8378
• [misc][distributed] add VLLM_SKIP_P2P_CHECK flag by @youkaichao in #8911
• [Core] Priority-based scheduling in async engine by @schoennenbeck in #8850
• [misc] fix wheel name by @youkaichao in #8919
• [Bugfix][Intel] Fix XPU Dockerfile Build by @tylertitsworth in #7824
• [Misc] Remove vLLM patch of BaichuanTokenizer by @DarkLight1337 in #8921
• [Bugfix] Fix code for downloading models from modelscope by @tastelikefeet in #8443
• [Bugfix] Fix PP for Multi-Step by @varun-sundar-rabindranath in #8887
• [CI/Build] Update models tests & examples by @DarkLight1337 in #8874
• [Frontend] Make beam search emulator temperature modifiable by @nFunctor in #8928
• [Bugfix] Support testing prefill throughput with benchmark_serving.py --hf-output-len 1 by @heheda12345 in #8891
• [doc] organize installation doc and expose per-commit docker by @youkaichao in #8931
• [Core] Improve choice of Python multiprocessing method by @russellb in #8823
• [Bugfix] Block manager v2 with preemption and lookahead slots by @sroy745 in #8824
• [Bugfix] Fix Marlin MoE act order when is_k_full == False by @ElizaWszola in #8741
• [CI/Build] Add test decorator for minimum GPU memory by @DarkLight1337 in #8925
• [Build/CI] Set FETCHCONTENT_BASE_DIR to one location for better caching by @tlrmchlsmth in #8930
• [Model] Support Qwen2.5-Math-RM-72B by @zhuzilin in #8896
• [Model][LoRA]LoRA support added for MiniCPMV2.5 by @jeejeelee in #7199
• [BugFix] Fix seeded random sampling with encoder-decoder models by @njhill in #8870
• [Misc] Fix typo in BlockSpaceManagerV1 by @juncheoll in #8944
• [Frontend] Added support for HF's new continue_final_message parameter by @danieljannai21 in #8942
• [Kernel][Model] Varlen prefill + Prefill chunking support for mamba kernels and Jamba model by @mzusman in #8533
• [Model] support input embeddings for qwen2vl by @whyiug in #8856
• [Misc][CI/Build] Include cv2 via mistral_common[opencv] by @ywang96 in #8951
• [Model][LoRA]LoRA support added for MiniCPMV2.6 by @jeejeelee in #8943
• [Model] Expose InternVL2 max_dynamic_patch as a mm_processor_kwarg by @Isotr0py in #8946
• [Core] Make scheduling policy settable via EngineArgs by @schoennenbeck in #8956
• [Misc] Adjust max_position_embeddings for LoRA compatibility by @jeejeelee in #8957
• [ci] Add CODEOWNERS for test directories by @khluu in #8795
• [CI][SpecDecode] Fix spec decode tests, use flash attention backend for spec decode CI tests. by @LiuXiaoxuanPKU in #8975
• [Frontend][Core] Move guided decoding params into sampling params by @joerunde in #8252
• [CI/Build] Fix machete generated kernel files ordering by @khluu in #8976
• [torch.compile] fix tensor alias by @youkaichao in #8982
• [Misc] add process_weights_after_loading for DummyLoader by @divakar-amd in #8969
• [Bugfix] Fix Fuyu tensor parallel inference by @Isotr0py in #8986
• [Bugfix] Fix Token IDs Reference for MiniCPM-V When Images are Provided With No Placeholders by @alex-jw-brooks in #8991
• [Core] [Frontend] Priority scheduling for embeddings and in the OpenAI-API by @schoennenbeck in #8965
• [Doc] Update list of supported models by @DarkLight1337 in #8987
• Update benchmark_serving.py to read and write json-datasets, results in UTF8, for better compatibility with Windows by @vlsav in https://github.com...

v0.6.2

Highlights

Model Support

• Support Llama 3.2 models (#8811, #8822)

   vllm serve meta-llama/Llama-3.2-11B-Vision-Instruct --enforce-eager --max-num-seqs 16
  

• Beam search have been soft deprecated. We are moving towards a version of beam search that's more performant and also simplifying vLLM's core. (#8684, #8763, #8713)

  • ⚠️ You will see the following error now, this is breaking change!

    Using beam search as a sampling parameter is deprecated, and will be removed in the future release. Please use the vllm.LLM.use_beam_search method for dedicated beam search instead, or set the environment variable VLLM_ALLOW_DEPRECATED_BEAM_SEARCH=1 to suppress this error. For more details, see #8306

• Support for Solar Model (#8386), minicpm3 (#8297), LLaVA-Onevision model support (#8486)

• Enhancements: pp for qwen2-vl (#8696), multiple images for qwen-vl (#8247), mistral function calling (#8515), bitsandbytes support for Gemma2 (#8338), tensor parallelism with bitsandbytes quantization (#8434)

Hardware Support

• TPU: implement multi-step scheduling (#8489), use Ray for default distributed backend (#8389)
• CPU: Enable mrope and support Qwen2-VL on CPU backend (#8770)
• AMD: custom paged attention kernel for rocm (#8310), and fp8 kv cache support (#8577)

Production Engine

• Initial support for priority sheduling (#5958)
• Support Lora lineage and base model metadata management (#6315)
• Batch inference for llm.chat() API (#8648)

Performance

• Introduce MQLLMEngine for API Server, boost throughput 30% in single step and 7% in multistep (#8157, #8761, #8584)
• Multi-step scheduling enhancements
  • Prompt logprobs support in Multi-step (#8199)
  • Add output streaming support to multi-step + async (#8335)
  • Add flashinfer backend (#7928)
• Add cuda graph support during decoding for encoder-decoder models (#7631)

Others

• Support sample from HF datasets and image input for benchmark_serving (#8495)
• Progress in torch.compile integration (#8488, #8480, #8384, #8526, #8445)

What's Changed

• [MISC] Dump model runner inputs when crashing by @comaniac in #8305
• [misc] remove engine_use_ray by @youkaichao in #8126
• [TPU] Use Ray for default distributed backend by @WoosukKwon in #8389
• Fix the AMD weight loading tests by @mgoin in #8390
• [Bugfix]: Fix the logic for deciding if tool parsing is used by @tomeras91 in #8366
• [Gemma2] add bitsandbytes support for Gemma2 by @blueyo0 in #8338
• [Misc] Raise error when using encoder/decoder model with cpu backend by @kevin314 in #8355
• [Misc] Use RoPE cache for MRoPE by @WoosukKwon in #8396
• [torch.compile] hide slicing under custom op for inductor by @youkaichao in #8384
• [Hotfix][VLM] Fixing max position embeddings for Pixtral by @ywang96 in #8399
• [Bugfix] Fix InternVL2 inference with various num_patches by @Isotr0py in #8375
• [Model] Support multiple images for qwen-vl by @alex-jw-brooks in #8247
• [BugFix] lazy init _copy_stream to avoid torch init wrong gpu instance by @lnykww in #8403
• [BugFix] Fix Duplicate Assignment of Class Variable in Hermes2ProToolParser by @vegaluisjose in #8423
• [Bugfix] Offline mode fix by @joerunde in #8376
• [multi-step] add flashinfer backend by @SolitaryThinker in #7928
• [Core] Add engine option to return only deltas or final output by @njhill in #7381
• [Bugfix] multi-step + flashinfer: ensure cuda graph compatible by @alexm-neuralmagic in #8427
• [Hotfix][Core][VLM] Disable chunked prefill by default and prefix caching for multimodal models by @ywang96 in #8425
• [CI/Build] Disable multi-node test for InternVL2 by @ywang96 in #8428
• [Hotfix][Pixtral] Fix multiple images bugs by @patrickvonplaten in #8415
• [Bugfix] Fix weight loading issue by rename variable. by @wenxcs in #8293
• [Misc] Update Pixtral example by @ywang96 in #8431
• [BugFix] fix group_topk by @dsikka in #8430
• [Core] Factor out input preprocessing to a separate class by @DarkLight1337 in #7329
• [Bugfix] Mapping physical device indices for e2e test utils by @ShangmingCai in #8290
• [Bugfix] Bump fastapi and pydantic version by @DarkLight1337 in #8435
• [CI/Build] Update pixtral tests to use JSON by @DarkLight1337 in #8436
• [Bugfix] Fix async log stats by @alexm-neuralmagic in #8417
• [bugfix] torch profiler bug for single gpu with GPUExecutor by @SolitaryThinker in #8354
• bump version to v0.6.1.post1 by @simon-mo in #8440
• [CI/Build] Enable InternVL2 PP test only on single node by @Isotr0py in #8437
• [doc] recommend pip instead of conda by @youkaichao in #8446
• [Misc] Skip loading extra bias for Qwen2-VL GPTQ-Int8 by @jeejeelee in #8442
• [misc][ci] fix quant test by @youkaichao in #8449
• [Installation] Gate FastAPI version for Python 3.8 by @DarkLight1337 in #8456
• [plugin][torch.compile] allow to add custom compile backend by @youkaichao in #8445
• [CI/Build] Reorganize models tests by @DarkLight1337 in #7820
• [Doc] Add oneDNN installation to CPU backend documentation by @Isotr0py in #8467
• [HotFix] Fix final output truncation with stop string + streaming by @njhill in #8468
• bump version to v0.6.1.post2 by @simon-mo in #8473
• [Hardware][intel GPU] bump up ipex version to 2.3 by @jikunshang in #8365
• [Kernel][Hardware][Amd]Custom paged attention kernel for rocm by @charlifu in #8310
• [Model] support minicpm3 by @SUDA-HLT-ywfang in #8297
• [torch.compile] fix functionalization by @youkaichao in #8480
• [torch.compile] add a flag to disable custom op by @youkaichao in #8488
• [TPU] Implement multi-step scheduling by @WoosukKwon in #8489
• [Bugfix][Model] Fix Python 3.8 compatibility in Pixtral model by updating type annotations by @chrisociepa in #8490
• [Bugfix][Kernel] Add IQ1_M quantization implementation to GGUF kernel by @Isotr0py in #8357
• [Kernel] Enable 8-bit weights in Fused Marlin MoE by @ElizaWszola in #8032
• [Frontend] Expose revision arg in OpenAI server by @lewtun in #8501
• [BugFix] Fix clean shutdown issues by @njhill in #8492
• [Bugfix][Kernel] Fix build for sm_60 in GGUF kernel by @sasha0552 in #8506
• [Kernel] AQ AZP 3/4: Asymmetric quantization kernels by @ProExpertProg in #7270
• [doc] update doc on testing and debugging by @youkaichao in #8514
• [Bugfix] Bind api server port before starting engine by @kevin314 in #8491
• [perf bench] set timeout to debug hanging by @simon-mo in #8516
• [misc] small qol fixes for release process by @simon-mo in #8517
• [Bugfix] Fix 3.12 builds on main by @joerunde in #8510
• [refactor] remove triton based sampler by @simon-mo in #8524
• [Frontend] Improve Nullable kv Arg Parsing by @alex-jw-brooks in #8525
• [Misc][Bugfix] Disable guided decoding for mistral tokenizer by @ywang96 in #8521
• [torch.compile] register allreduce operations as custom ops by @youkaichao in #8526
• [Misc] Limit to ray[adag] 2.35 to avoid backward incompatible change by @ruisearch42 in #8509
• [Benchmark] Support sample from HF datasets and image input for benchmark_serving by @Isotr0py in #8495
• [Encoder decoder] Add cuda graph support during decoding for encoder-decoder models by @sroy745 in #7631
• [Feature][kernel] tensor parallelism with bitsandbytes quantizati...

v0.6.1.post2

Highlights

• This release contains an important bugfix related to token streaming combined with stop string (#8468)

What's Changed

• [CI/Build] Enable InternVL2 PP test only on single node by @Isotr0py in #8437
• [doc] recommend pip instead of conda by @youkaichao in #8446
• [Misc] Skip loading extra bias for Qwen2-VL GPTQ-Int8 by @jeejeelee in #8442
• [misc][ci] fix quant test by @youkaichao in #8449
• [Installation] Gate FastAPI version for Python 3.8 by @DarkLight1337 in #8456
• [plugin][torch.compile] allow to add custom compile backend by @youkaichao in #8445
• [CI/Build] Reorganize models tests by @DarkLight1337 in #7820
• [Doc] Add oneDNN installation to CPU backend documentation by @Isotr0py in #8467
• [HotFix] Fix final output truncation with stop string + streaming by @njhill in #8468
• bump version to v0.6.1.post2 by @simon-mo in #8473

Full Changelog: v0.6.1.post1...v0.6.1.post2

v0.6.1.post1

Highlights

This release features important bug fixes and enhancements for

• Pixtral models. (#8415, #8425, #8399, #8431)
  • Chunked scheduling has been turned off for vision models. Please replace --max_num_batched_tokens 16384 with --max-model-len 16384
• Multistep scheduling. (#8417, #7928, #8427)
• Tool use. (#8423, #8366)

Also

• support multiple images for qwen-vl (#8247)
• removes engine_use_ray (#8126)
• add engine option to return only deltas or final output (#7381)
• add bitsandbytes support for Gemma2 (#8338)

What's Changed

• [MISC] Dump model runner inputs when crashing by @comaniac in #8305
• [misc] remove engine_use_ray by @youkaichao in #8126
• [TPU] Use Ray for default distributed backend by @WoosukKwon in #8389
• Fix the AMD weight loading tests by @mgoin in #8390
• [Bugfix]: Fix the logic for deciding if tool parsing is used by @tomeras91 in #8366
• [Gemma2] add bitsandbytes support for Gemma2 by @blueyo0 in #8338
• [Misc] Raise error when using encoder/decoder model with cpu backend by @kevin314 in #8355
• [Misc] Use RoPE cache for MRoPE by @WoosukKwon in #8396
• [torch.compile] hide slicing under custom op for inductor by @youkaichao in #8384
• [Hotfix][VLM] Fixing max position embeddings for Pixtral by @ywang96 in #8399
• [Bugfix] Fix InternVL2 inference with various num_patches by @Isotr0py in #8375
• [Model] Support multiple images for qwen-vl by @alex-jw-brooks in #8247
• [BugFix] lazy init _copy_stream to avoid torch init wrong gpu instance by @lnykww in #8403
• [BugFix] Fix Duplicate Assignment of Class Variable in Hermes2ProToolParser by @vegaluisjose in #8423
• [Bugfix] Offline mode fix by @joerunde in #8376
• [multi-step] add flashinfer backend by @SolitaryThinker in #7928
• [Core] Add engine option to return only deltas or final output by @njhill in #7381
• [Bugfix] multi-step + flashinfer: ensure cuda graph compatible by @alexm-neuralmagic in #8427
• [Hotfix][Core][VLM] Disable chunked prefill by default and prefix caching for multimodal models by @ywang96 in #8425
• [CI/Build] Disable multi-node test for InternVL2 by @ywang96 in #8428
• [Hotfix][Pixtral] Fix multiple images bugs by @patrickvonplaten in #8415
• [Bugfix] Fix weight loading issue by rename variable. by @wenxcs in #8293
• [Misc] Update Pixtral example by @ywang96 in #8431
• [BugFix] fix group_topk by @dsikka in #8430
• [Core] Factor out input preprocessing to a separate class by @DarkLight1337 in #7329
• [Bugfix] Mapping physical device indices for e2e test utils by @ShangmingCai in #8290
• [Bugfix] Bump fastapi and pydantic version by @DarkLight1337 in #8435
• [CI/Build] Update pixtral tests to use JSON by @DarkLight1337 in #8436
• [Bugfix] Fix async log stats by @alexm-neuralmagic in #8417
• [bugfix] torch profiler bug for single gpu with GPUExecutor by @SolitaryThinker in #8354
• bump version to v0.6.1.post1 by @simon-mo in #8440

New Contributors

• @blueyo0 made their first contribution in #8338
• @lnykww made their first contribution in #8403
• @vegaluisjose made their first contribution in #8423

Full Changelog: v0.6.1...v0.6.1.post1

v0.6.1

Highlights

Model Support

• Added support for Pixtral (mistralai/Pixtral-12B-2409). (#8377, #8168)
• Added support for Llava-Next-Video (#7559), Qwen-VL (#8029), Qwen2-VL (#7905)
• Multi-input support for LLaVA (#8238), InternVL2 models (#8201)

Performance Enhancements

• Memory optimization for awq_gemm and awq_dequantize, 2x throughput (#8248)

Production Engine

• Support load and unload LoRA in api server (#6566)
• Add progress reporting to batch runner (#8060)
• Add support for NVIDIA ModelOpt static scaling checkpoints. (#6112)

Others

• Update the docker image to use Python 3.12 for small performance bump. (#8133)
• Added CODE_OF_CONDUCT.md (#8161)

What's Changed

• [Doc] [Misc] Create CODE_OF_CONDUCT.md by @mmcelaney in #8161
• [bugfix] Upgrade minimum OpenAI version by @SolitaryThinker in #8169
• [Misc] Clean up RoPE forward_native by @WoosukKwon in #8076
• [ci] Mark LoRA test as soft-fail by @khluu in #8160
• [Core/Bugfix] Add query dtype as per FlashInfer API requirements. by @elfiegg in #8173
• [Doc] Add multi-image input example and update supported models by @DarkLight1337 in #8181
• Inclusion of InternVLChatModel In PP_SUPPORTED_MODELS(Pipeline Parallelism) by @Manikandan-Thangaraj-ZS0321 in #7860
• [MODEL] Qwen Multimodal Support (Qwen-VL / Qwen-VL-Chat) by @alex-jw-brooks in #8029
• Move verify_marlin_supported to GPTQMarlinLinearMethod by @mgoin in #8165
• [Documentation][Spec Decode] Add documentation about lossless guarantees in Speculative Decoding in vLLM by @sroy745 in #7962
• [Core] Support load and unload LoRA in api server by @Jeffwan in #6566
• [BugFix] Fix Granite model configuration by @njhill in #8216
• [Frontend] Add --logprobs argument to benchmark_serving.py by @afeldman-nm in #8191
• [Misc] Use ray[adag] dependency instead of cuda by @ruisearch42 in #7938
• [CI/Build] Increasing timeout for multiproc worker tests by @alexeykondrat in #8203
• [Kernel] [Triton] Memory optimization for awq_gemm and awq_dequantize, 2x throughput by @rasmith in #8248
• [Misc] Remove SqueezeLLM by @dsikka in #8220
• [Model] Allow loading from original Mistral format by @patrickvonplaten in #8168
• [misc] [doc] [frontend] LLM torch profiler support by @SolitaryThinker in #7943
• [Bugfix] Fix Hermes tool call chat template bug by @K-Mistele in #8256
• [Model] Multi-input support for LLaVA and fix embedding inputs for multi-image models by @DarkLight1337 in #8238
• Enable Random Prefix Caching in Serving Profiling Tool (benchmark_serving.py) by @wschin in #8241
• [tpu][misc] fix typo by @youkaichao in #8260
• [Bugfix] Fix broken OpenAI tensorizer test by @DarkLight1337 in #8258
• [Model][VLM] Support multi-images inputs for InternVL2 models by @Isotr0py in #8201
• [Model][VLM] Decouple weight loading logic for Paligemma by @Isotr0py in #8269
• ppc64le: Dockerfile fixed, and a script for buildkite by @sumitd2 in #8026
• [CI/Build] Use python 3.12 in cuda image by @joerunde in #8133
• [Bugfix] Fix async postprocessor in case of preemption by @alexm-neuralmagic in #8267
• [Bugfix] Streamed tool calls now more strictly follow OpenAI's format; ensures Vercel AI SDK compatibility by @K-Mistele in #8272
• [Frontend] Add progress reporting to run_batch.py by @alugowski in #8060
• [Bugfix] Correct adapter usage for cohere and jamba by @vladislavkruglikov in #8292
• [Misc] GPTQ Activation Ordering by @kylesayrs in #8135
• [Misc] Fused MoE Marlin support for GPTQ by @dsikka in #8217
• Add NVIDIA Meetup slides, announce AMD meetup, and add contact info by @simon-mo in #8319
• [Bugfix] Fix missing post_layernorm in CLIP by @DarkLight1337 in #8155
• [CI/Build] enable ccache/scccache for HIP builds by @dtrifiro in #8327
• [Frontend] Clean up type annotations for mistral tokenizer by @DarkLight1337 in #8314
• [CI/Build] Enabling kernels tests for AMD, ignoring some of then that fail by @alexeykondrat in #8130
• Fix ppc64le buildkite job by @sumitd2 in #8309
• [Spec Decode] Move ops.advance_step to flash attn advance_step by @kevin314 in #8224
• [Misc] remove peft as dependency for prompt models by @prashantgupta24 in #8162
• [MISC] Keep chunked prefill enabled by default with long context when prefix caching is enabled by @comaniac in #8342
• [Bugfix] Ensure multistep lookahead allocation is compatible with cuda graph max capture by @alexm-neuralmagic in #8340
• [Core/Bugfix] pass VLLM_ATTENTION_BACKEND to ray workers by @SolitaryThinker in #8172
• [CI/Build][Kernel] Update CUTLASS to 3.5.1 tag by @tlrmchlsmth in #8043
• [Misc] Skip loading extra bias for Qwen2-MOE GPTQ models by @jeejeelee in #8329
• [Bugfix] Fix InternVL2 vision embeddings process with pipeline parallel by @Isotr0py in #8299
• [Hardware][NV] Add support for ModelOpt static scaling checkpoints. by @pavanimajety in #6112
• [model] Support for Llava-Next-Video model by @TKONIY in #7559
• [Frontend] Create ErrorResponse instead of raising exceptions in run_batch by @pooyadavoodi in #8347
• [Model][VLM] Add Qwen2-VL model support by @fyabc in #7905
• [Hardware][Intel] Support compressed-tensor W8A8 for CPU backend by @bigPYJ1151 in #7257
• [CI/Build] Excluding test_moe.py from AMD Kernels tests for investigation by @alexeykondrat in #8373
• [Bugfix] Add missing attributes in mistral tokenizer by @DarkLight1337 in #8364
• [Kernel][Misc] Add meta functions for ops to prevent graph breaks by @bnellnm in #6917
• [Misc] Move device options to a single place by @akx in #8322
• [Speculative Decoding] Test refactor by @LiuXiaoxuanPKU in #8317
• Pixtral by @patrickvonplaten in #8377
• Bump version to v0.6.1 by @simon-mo in #8379

New Contributors

• @mmcelaney made their first contribution in #8161
• @elfiegg made their first contribution in #8173
• @Manikandan-Thangaraj-ZS0321 made their first contribution in #7860
• @sumitd2 made their first contribution in #8026
• @alugowski made their first contribution in #8060
• @vladislavkruglikov made their first contribution in #8292
• @kevin314 made their first contribution in #8224
• @TKONIY made their first contribution in #7559
• @akx made their first contribution in #8322

Full Changelog: v0.6.0...v0.6.1

v0.6.0

Highlights

Performance Update

• We are excited to announce a faster vLLM delivering 2x more throughput compared to v0.5.3. The default parameters should achieve great speed up, but we recommend also try out turning on multi step scheduling. You can do so by setting --num-scheduler-steps 8 in the engine arguments. Please note that it still have some limitations and being actively hardened, see #7528 for known issues.
  • Multi-step scheduler now supports LLMEngine and log_probs (#7789, #7652)
  • Asynchronous output processor overlaps the output data structures construction with GPU works, delivering 12% throughput increase. (#7049, #7911, #7921, #8050)
  • Using FlashInfer backend for FP8 KV Cache (#7798, #7985), rejection sampling in Speculative Decoding (#7244)

Model Support

• Support bitsandbytes 8-bit and FP4 quantized models (#7445)
• New LLMs: Exaone (#7819), Granite (#7436), Phi-3.5-MoE (#7729)
• A new tokenizer mode for mistral models to use the native mistral-commons package (#7739)
• Multi-modality:
  • multi-image input support for LLaVA-Next (#7230), Phi-3-vision models (#7783)
  • Ultravox support for multiple audio chunks (#7963)
  • TP support for ViTs (#7186)

Hardware Support

• NVIDIA GPU: extend cuda graph size for H200 (#7894)
• AMD: Triton implementations awq_dequantize and awq_gemm to support AWQ (#7386)
• Intel GPU: pipeline parallel support (#7810)
• Neuron: context lengths and token generation buckets (#7885, #8062)
• TPU: single and multi-host TPUs on GKE (#7613), Async output processing (#8011)

Production Features

• OpenAI-Compatible Tools API + Streaming for Hermes & Mistral models! (#5649)
• Add json_schema support from OpenAI protocol (#7654)
• Enable chunked prefill and prefix caching together (#7753, #8120)
• Multimodal support in offline chat (#8098), and multiple multi-modal items in the OpenAI frontend (#8049)

Misc

• Support benchmarking async engine in benchmark_throughput.py (#7964)
• Progress in integration with torch.compile: avoid Dynamo guard evaluation overhead (#7898), skip compile for profiling (#7796)

What's Changed

• [Core] Add multi-step support to LLMEngine by @alexm-neuralmagic in #7789
• [Bugfix] Fix run_batch logger by @pooyadavoodi in #7640
• [Frontend] Publish Prometheus metrics in run_batch API by @pooyadavoodi in #7641
• [Frontend] add json_schema support from OpenAI protocol by @rockwotj in #7654
• [misc][core] lazy import outlines by @youkaichao in #7831
• [ci][test] exclude model download time in server start time by @youkaichao in #7834
• [ci][test] fix RemoteOpenAIServer by @youkaichao in #7838
• [Bugfix] Fix Phi-3v crash when input images are of certain sizes by @zifeitong in #7840
• [Model][VLM] Support multi-images inputs for Phi-3-vision models by @Isotr0py in #7783
• [Misc] Remove snapshot_download usage in InternVL2 test by @Isotr0py in #7835
• [misc][cuda] improve pynvml warning by @youkaichao in #7852
• [Spec Decoding] Streamline batch expansion tensor manipulation by @njhill in #7851
• [Bugfix]: Use float32 for base64 embedding by @HollowMan6 in #7855
• [CI/Build] Avoid downloading all HF files in RemoteOpenAIServer by @DarkLight1337 in #7836
• [Performance][BlockManagerV2] Mark prefix cache block as computed after schedule by @comaniac in #7822
• [Misc] Update qqq to use vLLMParameters by @dsikka in #7805
• [Misc] Update gptq_marlin_24 to use vLLMParameters by @dsikka in #7762
• [misc] fix custom allreduce p2p cache file generation by @youkaichao in #7853
• [Bugfix] neuron: enable tensor parallelism by @omrishiv in #7562
• [Misc] Update compressed tensors lifecycle to remove prefix from create_weights by @dsikka in #7825
• [Core] Asynchronous Output Processor by @megha95 in #7049
• [Tests] Disable retries and use context manager for openai client by @njhill in #7565
• [core][torch.compile] not compile for profiling by @youkaichao in #7796
• Revert #7509 by @comaniac in #7887
• [Model] Add Mistral Tokenization to improve robustness and chat encoding by @patrickvonplaten in #7739
• [CI/Build][VLM] Cleanup multiple images inputs model test by @Isotr0py in #7897
• [Hardware][Intel GPU] Add intel GPU pipeline parallel support. by @jikunshang in #7810
• [CI/Build][ROCm] Enabling tensorizer tests for ROCm by @alexeykondrat in #7237
• [Bugfix] Fix phi3v incorrect image_idx when using async engine by @Isotr0py in #7916
• [cuda][misc] error on empty CUDA_VISIBLE_DEVICES by @youkaichao in #7924
• [Kernel] Expand MoE weight loading + Add Fused Marlin MoE Kernel by @dsikka in #7766
• [benchmark] Update TGI version by @philschmid in #7917
• [Model] Add multi-image input support for LLaVA-Next offline inference by @zifeitong in #7230
• [mypy] Enable mypy type checking for vllm/core by @jberkhahn in #7229
• [Core][VLM] Stack multimodal tensors to represent multiple images within each prompt by @petersalas in #7902
• [hardware][rocm] allow rocm to override default env var by @youkaichao in #7926
• [Bugfix] Allow ScalarType to be compiled with pytorch 2.3 and add checks for registering FakeScalarType and dynamo support. by @bnellnm in #7886
• [mypy][CI/Build] Fix mypy errors by @DarkLight1337 in #7929
• [Core] Async_output_proc: Add virtual engine support (towards pipeline parallel) by @alexm-neuralmagic in #7911
• [Performance] Enable chunked prefill and prefix caching together by @comaniac in #7753
• [ci][test] fix pp test failure by @youkaichao in #7945
• [Doc] fix the autoAWQ example by @stas00 in #7937
• [Bugfix][VLM] Fix incompatibility between #7902 and #7230 by @DarkLight1337 in #7948
• [Core][Kernels] Use FlashInfer backend for FP8 KV Cache when available. by @pavanimajety in #7798
• [Kernel] [Triton] [AMD] Adding Triton implementations awq_dequantize and awq_gemm to support AWQ by @rasmith in #7386
• [TPU] Upgrade PyTorch XLA nightly by @WoosukKwon in #7967
• [Doc] fix 404 link by @stas00 in #7966
• [Kernel/Model] Migrate mamba_ssm and causal_conv1d kernels to vLLM by @mzusman in #7651
• [Bugfix] Make torch registration of punica ops optional by @bnellnm in #7970
• [torch.compile] avoid Dynamo guard evaluation overhead by @youkaichao in #7898
• Remove faulty Meta-Llama-3-8B-Instruct-FP8.yaml lm-eval test by @mgoin in #7961
• [Frontend] Minor optimizations to zmq decoupled front-end by @njhill in #7957
• [torch.compile] remove reset by @youkaichao in #7975
• [VLM][Core] Fix exceptions on ragged NestedTensors by @petersalas in #7974
• Revert "[Core][Kernels] Use FlashInfer backend for FP8 KV Cache when available." by @youkaichao in #7982
• [Bugfix] Unify rank computation across regular decoding and speculative decoding by @jmkuebler in #7899
• [Core] Combine async postprocessor and multi-step by @alexm-neuralmagic in #7921
• [Core][Kernels] Enable FP8 KV Cache with Flashinfer backend. + BugFix for kv_cache_dtype=auto by @pavanimajety in #7985
• extend cuda graph size for H200 by @kushanam in #7894
• [Bugfix] Fix incorrect vocal embedding shards for GGUF model in tensor parallelism by @Isotr0py in #7954
• [misc] update tpu int8 to use new vLLM Parameters by @dsikka in #7973
• [Neuron] Adding support for context-lenght, token-gen buckets. by @hbikki in #7885
• support bitsandbytes 8-bit and FP4 quantized models by @chenqianfzh in #7445
• Add more percentiles and latencies by @...