vLLM 0.29.0

v0.29.0

Highlights

This release features 594 commits from 277 contributors (91 new)!

• Model Runner V2 is now the default for all models (#53183), completing the rollout that began with pooling models (#48290). MRV2 also gained CUDA graph memory profiling for KV cache auto-sizing (#53306), batch-sharded sampling that cuts per-step logits memory by 1/TP (#50465), prompt embeds (#42963), extract_hidden_states speculation (#49811), padded FULL cudagraph dispatch for uniform decode under spec decode (#53407), and DP-sync skipping before EAGLE/MTP draft prefill (#53694). MRV1 remains in use for a few ROCm models and features MRV2 does not yet support.
• New models: Hy4-preview, Tencent's 770B/49B-active MoE with Gated DeepSeek Sparse Attention and native MTP (#54160); Qwen3.8-Flash-Next with BF16/FP8/NVFP4 and MTP (#53896); GraniteSWA and GraniteMoeSWA (#52706); NemotronH_Omni_Reasoning_V3 with MTP (#52929, #53121); Kimi K3 NVFP4 checkpoints (#53132).
• Kimi-K3 and DeepSeek V4 performance: fused MXFP4 top-k finalization in the K3 latent tail (about 5% E2E latency, #53152), K3 Mamba metadata preparation in one Triton launch (6.6-7.6x kernel speedup, #52388), tuned Hopper low-latency GEMM (#54088) now also dispatched on SM100 (#53534) and used for eh_proj (12.9-25.2% kernel speedup, #53942), GEMM-RS extended to GEMM-AR (#53053), MLA gate merged into the QKV-A projection (#54015), K3 DCP with DSpark (#52188) and DCP partial prefix cache hits (#50493); DeepSeek V4 shared experts fused into MegaMoE (#53040), adaptive top-k width re-landed (#52823), a native SwiGLU clamp kernel for Humming MoE (#53685), and an opt-in FlashInfer moe_ep expert backend (#49636).
• Speculative decoding: per-request acceptance stats in OpenAI API responses via --per-request-spec-decode-metrics (#48915), adaptive verification extended to logprobs (#52242), SM100 sparse MLA for GLM-5.2 (#52783) and DeepSeek V4 on SM90 (#52795), Qwen3-Omni DSpark drafts (#52560), PLaMo3 EAGLE-3/DFlash (#54239), and DFlash2 loading from speculators format (#53797).
• RL weight sync: a new sharded_rdt P2P backend where each worker pulls only its TP/EP slice over NIXL or Ray Direct Transport (#43375), rank-local IPC weight updates (#52497), sparse checkpoint-coordinate updates through native weight loaders (#50723, #53751), and routed expert loading for gpt-oss (#52209).
• Mamba prefix caching: internal prefill checkpoints deliver a 9%-25% TTFT improvement (#52789); prefix_cache_retention_interval is now a CLI argument defaulting to 0 (#52216), with dense retention automatically restored for hybrid models using EAGLE/MTP (#55760, #55861).
• New defaults: FlashInfer all-reduce enabled by default for TP CUDA groups, opt out with VLLM_ALLREDUCE_USE_FLASHINFER=0 (#52998); prefix-cache NONE_HASH is deterministic by default so distributed KV cache users no longer need to pin PYTHONHASHSEED (#51875); new --max-num-queued-reqs / --max-num-queued-tokens admission-control flags (#49445).
• Breaking changes: ten deprecated model architectures removed (#53608); FlexOlmo, Olmo3 and Hunyuan V1/VL migrated to the Transformers modeling backend (#53615); PyAV video decoder backend removed (#54231); python -m vllm.entrypoints.openai.api_server deprecated in favor of vllm serve (#52131); VLLM_TEST_FORCE_FP8_MARLIN (#52182) and VLLM_ROCM_USE_AITER_FP4_ASM_GEMM (#53141) removed.

:warning: Model Runner V1 Deprecation

Now that Model Runner V2 is used by default, we are considering Model Runner V1 deprecated and are targeting v0.32 for its removal. We do not intend to accept any more MRV1-specific improvements or optimizations.

Some features are not yet supported in MRV2 but we are planning for these gaps to be closed within the next 2-3 weeks. These include sequence parallelism, dual-batch overlap, elastic expert parallellism, custom logits processors and certain speculative decoding methods. For now, vLLM will still fall back to use MRV1 if any of these features are configured.

Release Artifacts

Python Wheels

| Platform | Install | |---|---| | PyPI (CUDA 13.0) | pip install vllm | | PyPI (CUDA 13.0, uv) | uv pip install vllm --torch-backend=auto | | ROCm | pip install vllm --extra-index-url https://wheels.vllm.ai/rocm/0.29.0/rocm723 | | XPU | uv pip install vllm --extra-index-url https://wheels.vllm.ai/0.29.0/xpu --extra-index-url https://download.pytorch.org/whl/xpu --index-strategy unsafe-best-match |

Docker Images

| Platform | Docker Image | |---|---| | CUDA 13.0 (Default) | docker pull vllm/vllm-openai:v0.29.0 (v0.29.0-cu130 also works) | | CUDA 12.9 | docker pull vllm/vllm-openai:v0.29.0-cu129 | | CUDA 13.0 + Ubuntu 24.04 | docker pull vllm/vllm-openai:v0.29.0-ubuntu2404 | | CUDA 12.9 + Ubuntu 24.04 | docker pull vllm/vllm-openai:v0.29.0-cu129-ubuntu2404 | | ROCm | docker pull vllm/vllm-openai-rocm:v0.29.0 | | CPU | docker pull vllm/vllm-openai-cpu:v0.29.0 | | XPU | docker pull vllm/vllm-openai-xpu:v0.29.0 |

Other Artifacts

Pre-built release artifacts are available in the Assets section at the bottom of this page, including:

• Source distribution tarball
• CUDA 12.9 Python wheels for x86_64 and arm64
• CUDA 13.0 Python wheels for x86_64 and arm64
• CPU Python wheels for x86_64, arm64, and macOS
• XPU Python wheel for x86_64

Model Support

• New models: Hy4-preview (#54160), Qwen3.8-Flash-Next (#53896), GraniteSWA and GraniteMoeSWA via the existing Granite implementation (#52706), NemotronH_Omni_Reasoning_V3 (#52929) with MTP for Nemotron VL models (#53121), bidirectional attention for DeepSeek-backbone embedding models (#52948), FP8 ModernBERT (#53101), and Kimi K3 NVFP4 checkpoints (#53132).
• Transformers modeling backend: FlexOlmo, Olmo3 and Hunyuan V1/VL migrated off native implementations (#53615), multimodal path hardened (#51827), and RMSNormFuser.fuse performance fixed (#52766).
• Weight loading: weight tying now inspects the checkpoint so a real lm_head is loaded even when the config claims tied embeddings (#51665, #53170).
• LoRA: DeepSeek V4 (#53361), Qwen3-Omni multimodal LoRA (#52786, #53557), tower/connector LoRA for LLaVA-NeXT (#49788) and LFM2-VL (#51498), plus fixes for Muse-Glimmer (#53513), Qwen3.5 embedding modules (#48850), partial LoRA on Qwen3.5/3.6 GatedDeltaNet (#47640), int32 overflow in punica kernels at long context (#53034), false target matches on unsupported module types (#52313), and base-layer/routed-expert prefix ordering (#52552).
• Multimodal performance: ViT full CUDA graph for Idefics3 and SmolVLM (#47625), packed encoder attention for Pixtral (#52185), fused Kimi vision Q/K RoPE kernel (#50400), Dots3 NOTE runtime (#53517) and Omni encoder optimizations (#53460), MM tensors no longer broadcast to workers for prefix-cache-covered items (#52041), redundant placeholder scans skipped (#52925), common token sequences cached (#53560), async media resolved concurrently across modalities (#54537), and H2D copies pinned to avoid stream stalls (#53412, #54292, #54299).
• Multimodal inputs: video embeds accepted by the Python frontend (#54242), modality-scoped mm_processor_kwargs honored (#53808), Qwen3-VL profiling honors cap_pixels_per_frame (#54380), video frame sampling respects MP4 edit-list trims (#48608), oversized items skip the MM processor cache instead of crashing (#53016), encoder cache entries survive until their last use (#54284), and repeated multimodal requests with the SHM cache no longer terminate the engine (#54994).
• Correctness: PaliGemma stale image scaling removed (#52692), Mistral3 placeholder grid with processor size overrides (#52874), MiniCPM-o on Transformers v5 (#54501), JinaVL cache order (#53553), mixed CLIP/SigLIP pooling batches (#53165), XD-RoPE models on prefix-cache hits (#53456), Qwen3-Omni audio encoder with non-divisible TP (#50858), GLM-5.2 no longer uses dense MHA (#52512), Moondream3 MoE all-reduce (#54152), deepseek-vl2 config defaults (#51302), HY-V3 compressed-tensors ignore matching (#48682), MiniMax-M3 FP8 query allocation under CUDA graph replay (#51203), and GraniteMoeHybrid quantized expert loading (#54052).

Engine Core

• Model Runner V2: default for all models (#53183) and pooling models (#48290), CUDA graph memory reservation (#53306), batch-sharded sampling (#50465), prompt embeds (#42963), extract_hidden_states (#49811), padded FULL cudagraph dispatch (#53407), DP-sync skipping for drafts (#53694), decoupled draft/target gumbel noise streams (#54282), encoder-only path split out (#53176), and memory released correctly on shutdown and sleep (#53508, #54246, #54162, #53955, #53682).
• Speculative decoding: per-request acceptance stats (#48915), adaptive verification with logprobs (#52242), on SM100 sparse MLA (#52783) and DSv4 + SM90 (#52795), varlen trtllm-gen decode (#52157), FlashInfer MLA for DSpark drafting under DCP (#54277), fused GDN MTP for all Qwen head ratios (#52539), widest uniform decode batch captured by default (#50488) with memory-safe graph sizes (#54418), and fixes for short_conv/LFM2 targets (#50272), speculators-format config overrides (#42376), DFlash draft RoPE layout (#54373), draft models with a different GQA ratio (#53002), draft models with a larger hidden size under TP (#52193), DSpark backend inheritance scoped to DeepSeek V4 (#52809), generic DSparkDraftModel configs for Qwen3 (#52197), and Gemma4 MTP under CUDA graphs (#53884).
• Prefix caching & scheduling: Mamba internal prefill checkpoints (#52789), prefix_cache_retention_interval argument (#52216, #55760, #55861), deterministic NONE_HASH (#51875), queue admission control (#49445), KV null block reserved when validating max_model_len (#47272), spec decode no longer padded up to max_model_len (#53962), and negative external block allocation prevented (#52707).
• RL workflows: sharded_rdt P2P weight sync (#43375), rank-local IPC updates (#52497), sparse checkpoint updates (#50723, #53751), gpt-oss routed expert loading (#52209), stable DeepSeek V4 mHC broadcast buffers across weight sync (#52626), and packed weight transfer stream reuse capping reserved-memory waste (#52951).
• Determinism: trace_decode_token_ids for deterministic decode replay (#46701), per-arch tuned batch-invariant matmul configs (about 3x decode kernels on RTX 4090D/H20, #53247), Blackwell autotuning with 33.6% E2E latency reduction (#53649), deterministic MoE combine under DP+EP (#45683), and fuse_allreduce_rms disabled under batch invariance (#51292).
• Kernels: fused embedding kernel (#53677), FA4 re-enabled for head_size=256 on Blackwell (#52980), vectorized sparse MLA mask loads (#52217), masked MHA prefill for GLM-5 head dimensions (#53785), GLM-5.2 sparse MLA Q concatenation fused with head padding (#53878), fused QK-norm + partial MRoPE + gate for Qwen3.6 (#52676), FlashInfer CuTeDSL BF16 low-latency GEMM as opt-in --linear-backend flashinfer_cutedsl (#50572), replicated embedding and norm fusion for DSV3 flat models (#48484), standardized fused shared-expert selection (#51695), GPT-OSS MoE topk metadata reuse (#45457), tuned cooperative topk (#53382), and tuned FP8 fused_moe for Qwen3.5 on L40S (+7%, #53819).
• Robustness: KV cache layout standardized under a KVCacheLayout enum (#51718), JIT warmup provider registry (#50174), --cpu-offload-params now reaches vision/audio towers (#53120), attention backend probe failures no longer crash init (#51703), FlashInfer XQA falls back on unsupported head_dim (#53111), FlashInfer prefill LSE normalized before merging to fix prefix-cache logits divergence (#52796), seed preserved when a batch mixes seeded and unseeded requests (#51866), startup thread allocation accounts for local DP workers (#52385), int32 overflow fixes in fused SiLU block quant (#53409) and LoRA kernels (#53034), a shared-memory race in fused groupwise RMSNorm quantization (#54111), BLHNC addressing for FlashInfer sparse MLA (#54465), Mamba state copy race (#50729), and a start_profile no-op after auto-stop (#51839).

Hardware & Performance

• NVIDIA: DeepSeek V3.2 / GLM-5.2 DSA routed to the optimized CUDA path on all GPUs (#52861), PCP for DSv3.2 sparse MLA (#52046), cuBLAS out_dtype router GEMM on all archs including GB10 (#54048), FlashInfer all-reduce tuning on SM103 (#53318, #53606), FA4 hdim256 on SM100 (#52980), SM120 sparse MLA fixes (#51395, #53574), DeepSeek V3.2 fused kernel grids hardened for 65k+ token launches (#52381), FlashMLA sparse decode workspace fix (#53755), MNNVL Lamport corruption fix (#53000), and opt-in Rubin Docker builds for CUDA 13.4/13.5 (#53443).
• AMD ROCm: dual-stream decode with hipgraphs (#52033), W4A4 preshuffled asm GEMM by default (+15% throughput on Llama-3.3-70B MXFP4, #53141), ROCr/CLR update fixing graph replay segfaults with up to about 20% TPOT improvement (#53712), fused KDA decode on MI325X (#52293), FULL cudagraphs for AITER MLA spec decode (#51171), FP8 asm MLA prefill for non-divisor head counts (#51040), DCP causal multi-token verification (#51705) and prefix cache hits for Kimi-K3 (#53598), AITER PA gluon decode for MiniMax-M3 (#52849), DeepSeek-V4 fusions for mHC/RMSNorm (#52737), C4A top-k (#52882), C4 compressor GEMMs (#53838) and SWA q/kv norm + FP8 quant (#53540), fused shared experts for block-FP8 (#53097), CPU offload on ROCm 7.13+ (#43018), TheRock 7.14 preview docker (#49925), int4/int8 quantization fixes (#52112, #48998, #51632, #53110), CUDA graphs captured on the current stream (#53818), AITER metadata preserved across graph replay (#53821), and improved ROCm detection under WSL (#38434).
• Intel XPU: INC int4 W4A8 linear backend (#50501), AutoRound MXFP8 MoE (#51248), EC connector KV offloading (#49532), and fixes for HunyuanOCR XD-RoPE (#52174), sparse-MLA metadata sync (#52066), Mamba state pointer overflow (#48109), oneCCL warm-up at world size 1 (#52389), and MRoPE (#53201).