# vLLM v0.10.2 — Release & Upgrade Intelligence
- **Product:** vLLM (https://vllm.ai)
- **Version:** v0.10.2
- **Released:** Sep 13, 2025
- **Impact Score:** 7/10
- **Breaking Changes:** Source flagged
- **Canonical Analysis:** https://changelogs.us/vllm/updates/v0.10.2/

## Executive Summary (TL;DR)
vLLM v0.10.2. Source highlight: This release contains 740 commits from 266 contributors (97 new)!

## Breaking Changes & Deprecations
- Breaking Changes**: This release includes PyTorch 2.8.0 upgrade, V0 deprecations, and API changes - please review the changelog carefully.
- Breaking change**: FlashMLA disabled on Blackwell GPUs due to compatibility issues (#24521).
- Breaking change**: Removed original Marlin quantization format (#23204).
- Major updates**: PyTorch 2.8.0 upgrade (#20358) - breaking change requiring environment updates, FlashInfer v0.3.0 upgrade (#24086), and FlashInfer 0.2.14.post1 maintenance update (#23537).
- API breaking changes**: prompt_token_ids fallback removal from LLM.generate and LLM.embed (#18800), LoRA extra vocab size deprecation warning (#23635), LoRA bias parameter deprecation (#24339), and metrics naming change from TPOT to ITL (#24110).
- Breaking Changes 1. **PyTorch 2.8.0 upgrade** - Environment dependency change requiring updated CUDA versions 2. **FlashMLA Blackwell restriction** - FlashMLA disabled on Blackwell GPUs due to compatibility issues

## New Features
- **Pipeline parallelism expansion**: Added PP support for Hunyuan (#24212), Ovis2.5 (#23405), GPT-OSS (#23680), and Kimi-VL-A3B-Thinking-2506 (#23114).
- **LoRA ecosystem expansion**: Added LoRA support to Voxtral (#24517), Qwen-2.5-Omni (#24231), and DeepSeek models V2/V3/R1-0528 (#23971), with significantly faster LoRA startup performance (#23777).
- **V1 engine maturation**: Extended V1 support to compute capability < 8.0 (#23614, #24022), added cross-attention KV cache for encoder-decoder models (#23664), request-level logits processor integration (#23656), and KV events from connectors (#19737).
- **Hybrid and Mamba model improvements**: Enabled full CUDA graphs by default for hybrid models (#22594), disabled prefix caching for hybrid/Mamba models (#23716), added FP32 SSM kernel support (#23506), full CUDA graph support for Mamba1 (#23035), and V1 as default for Mamba models (#23650).
- [XPU][CI]add xpu env vars in CI scripts
- [Misc] Add @tdoublep as a maintainer of hybrid model and Triton-attention related code

## Improvements
- **New model families and enhancements**: Apertus (#23068), LFM2 (#22845), MiDashengLM (#23652), Motif-1-Tiny (#23414), Seed-Oss (#23241), Google EmbeddingGemma-300m (#24318), GTE sequence classification (#23524), Donut OCR model (#23229), KeyeVL-1.5-8B (#23838), R-4B vision model (#23246), Ernie4.5 VL (#22514), MiniCPM-V 4.5 (#23586), Ovis2.5 (#23084), Qwen3-Next with hybrid attention (#24526), InternVL3.5 with video support (#23658), Qwen2Audio embeddings (#23625), NemotronH Nano VLM (#23644), 
- **Classification and pooling enhancements**: Multi-label classification support (#23173), logit bias and sigmoid normalization (#24031), and FP32 precision heads for pooling models (#23810).
- **Performance optimizations**: Removed unnecessary CUDA sync from GLM-4.1V (#24332) and Qwen2VL (#24334) preprocessing, eliminated redundant all-reduce in Qwen3 MoE (#23169), optimized InternVL CPU threading (#24519), and GLM4.5-V video frame decoding (#24161).
- **Multimodal enhancements**: Multimodal cache tracking with mm_hash (#22711), UUID-based multimodal identifiers (#23394), improved V1 video embedding estimation (#24312), and simplified multimodal UUID handling (#24271).
- **NVIDIA Blackwell/SM100 generation**: FP8 MLA support with CUTLASS backend (#23289), DeepGEMM Linear with 1.5% E2E throughput improvement (#23351), Hopper DeepGEMM E8M0 for DeepSeekV3.1 (#23666), SM100 FlashInfer CUTLASS MoE FP8 backend (#22357), MXFP4 fused CUTLASS MoE (#23696), default MXFP4 MoE on Blackwell (#23008), and GPT-OSS DP/EP support with 52,003 tokens/s throughput (#23608).
- **Breaking change**: FlashMLA disabled on Blackwell GPUs due to compatibility issues (#24521).

## Bug Fixes
- **Performance core improvements**: `--safetensors-load-strategy` for NFS based file loading acceleration (#24469), critical CUDA graph capture throughput fix (#24128), scheduler optimization for single completions (#21917), multi-threaded model weight loading (#23928), and tensor core usage enforcement for FlashInfer decode (#23214).
- **Supporting updates**: XGrammar 0.1.23 (#22988), TPU core dump fix with tpu_info 0.4.0 (#23135), and compressed tensors version bump (#23202).
- [Bugfix] fix Qwen2.5-Omni processor output mapping
- [Bugfix][CI] Machete kernels: deterministic ordering for more cache hits
- [Misc] Fix backward compatibility from #23030
- [XPU] Fix compile size for xpu

## Upgrade Command
```bash
python -m pip install 'vllm==0.10.2'
```