chore: bump vllm from 0.8.5.post1 to 0.10.2

[dependabot]

Bumps vllm from 0.8.5.post1 to 0.10.2.

Release notes

Sourced from vllm's releases.

v0.10.2

Highlights

This release contains 740 commits from 266 contributors (97 new)!

Breaking Changes: This release includes PyTorch 2.8.0 upgrade, V0 deprecations, and API changes - please review the changelog carefully.

aarch64 support: This release features native support for aarch64 allowing usage of vLLM on GB200 platform. The docker image vllm/vllm-openai should already be multiplatform. To install the wheels, you can download the wheels from this release artifact or install via

uv pip install vllm==0.10.2 --extra-index-url https://wheels.vllm.ai/0.10.2/ --torch-backend=auto

Model Support

• 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), BLOOM V1 engine support (#23488), and Whisper encoder-decoder for V1 (#21088).
• Pipeline parallelism expansion: Added PP support for Hunyuan (#24212), Ovis2.5 (#23405), GPT-OSS (#23680), and Kimi-VL-A3B-Thinking-2506 (#23114).
• Data parallelism for vision models: Enabled DP for ViT across Qwen2.5VL (#22742), MiniCPM-V (#23948, #23327), Kimi-VL (#23817), and GLM-4.5V (#23168).
• 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).
• 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).

Engine Core

• 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).
• Backend expansion: Terratorch backend integration (#23513) enabling non-language model tasks like semantic segmentation and geospatial applications with --model-impl terratorch support.
• 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).
• 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).
• 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).
• Sampling and structured outputs: Support for all prompt logprobs (#23868), final logprobs (#22387), grammar bitmask optimization (#23361), and user-configurable KV cache memory size (#21489).
• Distributed: Support Decode Context Parallel (DCP) for MLA (#23734)

Hardware & Performance

• 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).
• Kernel and attention optimizations: FlashAttention MLA with CUDA graph support (#14258, #23958), V1 cross-attention support (#23297), FP8 support for FlashMLA (#22668), fused grouped TopK for MoE (#23274), Flash Linear Attention kernels (#24518), and W4A8 support on Hopper (#23198).
• Performance improvements: 13.7x speedup for token conversion (#20413), TTIT/TTFT improvements for disaggregated serving (#22760), symmetric memory all-reduce by default (#24111), FlashInfer warmup during startup (#23439), V1 model execution overlap (#23569), and various Triton configuration tuning (#23748, #23939).
• Platform expansion: Apple Silicon bfloat16 support for M2+ (#24129), IBM Z V1 engine support (#22725), Intel XPU torch.compile (#22609), XPU MoE data parallelism (#22887), XPU Triton attention (#24149), XPU FP8 quantization (#23148), and ROCm pipeline parallelism with Ray (#24275).
• Model-specific optimizations: Hardware-tuned MoE configurations for Qwen3-Next on B200/H200/H100 (#24698, #24688, #24699, #24695), GLM-4.5-Air-FP8 B200 configs (#23695), Kimi K2 optimization (#24597), and QWEN3 Coder/Thinking configs (#24266, #24330).

Quantization

• New quantization capabilities: Per-layer quantization routing (#23556), GGUF quantization with layer skipping (#23188), NFP4+FP8 MoE support (#22674), W4A8 channel scales (#23570), and AMD CDNA2/CDNA3 FP4 support (#22527).
• Advanced quantization infrastructure: Compressed tensors transforms for linear operations (#22486) enabling techniques like SpinQuantR1R2R4 and QuIP quantization methods.
• FlashInfer quantization integration: FP8 KV cache for TRTLLM prefill attention (#24197), FP8-qkv attention kernels (#23647), and FP8 per-tensor GEMMs (#22895).
• Platform-specific quantization: ROCm TorchAO quantization enablement (#24400) and TorchAO module swap configuration (#21982).
• Performance optimizations: MXFP4 MoE loading cache optimization (#24154) and compressed tensors version updates (#23202).
• Breaking change: Removed original Marlin quantization format (#23204).

API & Frontend

• OpenAI API enhancements: Gemma3n audio transcription/translation endpoints (#23735), transcription response usage statistics (#23576), and return_token_ids parameter (#22587).
• Response API improvements: Streaming support for non-harmony responses (#23741), non-streaming logprobs (#23319), MCP tool background mode (#23494), MCP streaming+background support (#23927), and tool output token reporting (#24285).
• Frontend optimizations: Error stack traces with --log-error-stack (#22960), collective RPC endpoint (#23075), beam search concurrency optimization (#23599), unnecessary detokenization skipping (#24236), and custom media UUIDs (#23449).
• Configuration enhancements: Formalized --mm-encoder-tp-mode flag (#23190), VLLM_DISABLE_PAD_FOR_CUDAGRAPH environment variable (#23595), EPLB configuration parameter (#20562), embedding endpoint chat request support (#23931), and LM Format Enforcer V1 integration (#22564).

... (truncated)

Commits

• 26b999c [CI Failure] Fix test_flashinfer_cutlass_mxfp4_mxfp8_fused_moe (#24750)
• da3fa78 [Compilation Bug] Fix Inductor Graph Output with Shape Issue (#24772)
• bbb7003 Enable conversion of multimodal models to pooling tasks (#24451)
• 89da8d9 [Qwen3Next] Fixes the cuda graph capture conditions under large batch sizes (...
• 01085b1 [Qwen3-Next] MoE configs for H100 TP=1,2 and TP2/EP (#24739)
• 66160a9 [BugFix] Fix Qwen3-Next PP (#24709)
• eaca762 [Qwen3-Next] MoE configs for H20 TP=1,2,4,8 (#24707)
• 880c741 [Bugfix] fixes the causal_conv1d_update kernel update non-speculative decodin...
• 40b6c91 [V1] feat:add engine v1 tracing (#20372)
• 2e6bc46 [Startup] Make DeepGEMM warmup scale with max-num-batched-tokens (#24693)
• Additional commits viewable in compare view

Dependabot will resolve any conflicts with this PR as long as you don't alter it yourself. You can also trigger a rebase manually by commenting @dependabot rebase.

---

Dependabot commands and options

You can trigger Dependabot actions by commenting on this PR:

• @dependabot rebase will rebase this PR
• @dependabot recreate will recreate this PR, overwriting any edits that have been made to it
• @dependabot merge will merge this PR after your CI passes on it
• @dependabot squash and merge will squash and merge this PR after your CI passes on it
• @dependabot cancel merge will cancel a previously requested merge and block automerging
• @dependabot reopen will reopen this PR if it is closed
• @dependabot close will close this PR and stop Dependabot recreating it. You can achieve the same result by closing it manually
• @dependabot show <dependency name> ignore conditions will show all of the ignore conditions of the specified dependency
• @dependabot ignore this major version will close this PR and stop Dependabot creating any more for this major version (unless you reopen the PR or upgrade to it yourself)
• @dependabot ignore this minor version will close this PR and stop Dependabot creating any more for this minor version (unless you reopen the PR or upgrade to it yourself)
• @dependabot ignore this dependency will close this PR and stop Dependabot creating any more for this dependency (unless you reopen the PR or upgrade to it yourself)

[dependabot]

Superseded by #52.