⚙️ 算子与内核
FlashInfer、CUDA graph、kernel fusion。
Finish RL Training Faster: Bitwise-Consistent, High-Performance, Agent-Driven
We make vLLM rollout and the training forward give bitwise identical logprobs on DeepSeek-V4-Flash and Nemotron-3.5-Lightning-30B-A3B (NVFP4), keep 91% of rollout throughput on DeepSeek-V4-Flash, and open-source the agent skill.
vLLM Support for NVIDIA Vera Rubin NVL72: 7.8x Throughput over GB200 NVL72
vLLM now runs on NVIDIA Vera Rubin NVL72 with day-0 model support, Rubin-tuned FlashInfer kernels, and locality-aware MoE, delivering 7.8x the per-GPU throughput of GB200 NVL72 on AgentX.
DeepSeek-V4.1-Flash on vLLM: 5x Agentic Throughput Since Day 0
Within three weeks of release, vLLM made DeepSeek-V4.1-Flash 1.9x faster at low concurrency and lifted its throughput 5x on SemiAnalysis AgentX, with SWA bounded replay, CUDA graphs, DeepSeek's new kernels, and vLLM kernel fusions.
Watermarking in vLLM
How vLLM implements distribution-preserving Gumbel-max text watermarking with efficient GPU kernels, statistical detection, speculative decoding, and repeated-context safeguards.
Announcing vllm-metal: Concurrent Serving on Apple Silicon
vllm-metal brings vLLM's paged, continuously batched serving stack to Apple Silicon, with flatter TTFT under concurrent agent load, batched MTP, and automatic M5 prefill acceleration.
PD Serving of Qwen3.8-2.4T
本文是『PD 分离服务』路径末尾三篇生产实战的收官篇，开篇即引用 Qwen3.5（25K TPS）篇的结果并把目标从“冲吞吐”扩展为完整 Pareto 前沿。建议三篇中最后阅读，重点吸收其分步测量、先估容量再选拓扑的可复用调优方法论。
vLLM x Novita AI: Chord, Faster INT4 MoE for Kimi K2.x. Up to 1.3x on H200, 2.15x on Untuned B300
Novita AI has open-sourced Chord, a high-performance W4A16 MoE CUDA kernel for Kimi K2.x serving shapes, with a Humming-compatible indexed path and grouped SM90 operators.
How we trained the fastest DSpark for Kimi-K3 using GB300 NVL72
本文是『推测解码』学习路径中的 DSpark 训练篇:应在掌握推测解码基础与 DFlash/DSpark 算法原理之后阅读,聚焦训练数据(hidden states)如何跨节点规模化生产。它排在两篇自适应验证文章之前,为理解 confidence head 驱动的硬件感知验证调度提供背景与动机。
vime × RL-Kernel × AMD: Bitwise Train–Rollout Consistency on ROCm
vime and RL-Kernel align selected-token logprobs bit for bit across Megatron training and vLLM rollout on AMD Instinct MI300X, with zero mismatches across 200 GRPO steps.
Kimi K3 Performance Optimizations in vLLM: The Road to 2.8× Throughput
Kimi K3 serving optimizations across scheduling, KDA prefix caching, ReplaySSM state recovery, PD disaggregation and state offload, parallelism, MoE, and GPU kernels.
Tiered KV Cache Offloading in vLLM
A host-centric framework for scaling KV cache across host memory, filesystems, object stores, and remote peers — reducing recomputation and increasing serving capacity.
Following the Bottleneck: Optimizing MiniMax M3 on AMD Instinct MI355X
A performance model for LLM serving: inspect local shapes, remove repeated work, verify data movement and dispatch, then follow the queue.
GLM 5.3 Optimizations, Part 1: Hybrid HiSparse Offloading in vLLM
vLLM integrates HiSparse as a pressure-driven memory tier that composes with the Hybrid Memory Allocator and KV offloading, letting GLM 5.3 requests keep decoding when their KV no longer fits in GPU memory, so concurrency stays high.
MiniMax H3 on vLLM-Omni: From System-Wide Optimization to Real-Time Serving with FastVideo’s FastH3
How vLLM-Omni optimizes and scales the complete MiniMax H3 stack, then integrates FastVideo’s four-step FastH3 for generation faster than playback.
Large-Scale Sharded Weight Transfer with Ray Direct Transport (RDT) in vLLM
We implement a native sharded weight transfer engine in vLLM utilizing Ray Direct Transport (RDT), achieving weight transfer for the Kimi K2 model in BF16 on 48 8xH100 nodes in 7.53s
IsoExec: Unified Execution to Eliminate Trainer-Inference Mismatch in SkyRL
IsoExec unifies numerical execution across SkyRL's vLLM and Megatron runtimes, reducing the average rollout-versus-training logprob difference below 1e-6 on Qwen3.5-35B-A3B with 25% overhead.
Distributed Layerwise Offload: Scaling Toward 200B+ DiT Models Efficiently in vLLM-Omni
Distributed Layerwise Offload shards and streams DiT weights across devices, serving a measured 124 GB Cosmos3 model on 64 GB HBM and estimating a path toward 200B+ models.
Adaptive Verification in vLLM: DSpark confidence-scheduled verification
本文是『推测解码』路径中自适应验证主题的第一篇，讲 DSpark 如何用置信度头按存活概率调度验证预算；下一篇是无 confidence head 的在线估计器方案，建议对比阅读以理解两类自适应验证的取舍。
Day 0 Support for Qwen3.8-2.4T-A95B on vLLM
Day-0 vLLM support for Qwen3.8-2.4T-A95B: a 2.4-trillion-parameter hybrid MoE model served out of the box, with FP8/BF16 checkpoints plus NVFP4 and MXFP4 quantized weights, and co-developed kernels on NVIDIA and AMD hardware.
Announcing Day-0 Support for NVIDIA Nemotron 3.5 Lightning on vLLM
How vLLM serves NVIDIA Nemotron 3.5 Lightning with OpenAI-compatible APIs, speculative decoding, and BF16/NVFP4 checkpoints across NVIDIA GPUs and edge systems.
vLLM Reaches 25K Total TPS/GPU on Qwen3.5
本文是『PD 分离服务』路径末尾三篇生产规模实战中建议最先读的一篇：它是 Qwen3.8 篇的直接前置（后者开篇引用本文结果），也是 GB200/GB300 NVL72 + 混合 GDN 架构 PD serving 的基础认知入口，可再顺读 GLM-5.2 篇对照 SLA 导向的取舍。
Optimizing vLLM on Arm CPUs
An overview of Arm CPU enablement and inference performance optimizations in vLLM.
Kimi K3 Is Here: Efficient Day-0 Support on vLLM
vLLM delivers day-0 Kimi K3 serving with hybrid KDA prefix caching, DSpark speculative decoding, production-scale disaggregation, and optimized kernels across NVIDIA and AMD GPUs.
Announcing vLLM AFD Plugin: Disaggregating Attention and FFN for Flexible MoE Serving
『PD 分离服务』路径的横向扩展篇:在掌握按阶段拆分的 PD 分离之后阅读,理解按组件(Attention vs FFN)拆分的另一维度;其异步 connector 本就部署在 PD 分离之上,与 hybrid SSM 篇(同一 PD 框架下的模型适配)互为补充。
From Day 0 to Production SLAs: Serving GLM-5.2 on 24 NVIDIA B300 GPUs with vLLM
本文是『PD 分离服务』学习路径末尾三篇生产规模实战之一，以 SLA 合规而非峰值吞吐为第一目标，与两篇 Qwen 吞吐导向文章形成方法论对照。建议在读完 Qwen3.5（25K TPS）篇建立 NVL72 + 混合架构 PD 基础后阅读，重点体会约束驱动的配置取舍与边界问题定位。
A Preview of Production-Scale Kimi K3 Support on vLLM
A preview of production-scale Kimi K3 support in vLLM, including KDA-aware prefix caching, fused kernels, optimized MXFP4 MoE, multimodal integration, and initial NVIDIA and AMD paths.
Keeping vLLM Production Quality: A Look Inside CI, Benchmarking, and the Release Process
How vLLM maintains production quality with extensive CI across diverse accelerators, nightly performance benchmark and accuracy evaluation, and a two-week release process.
TML Inkling on vLLM: Day-0 Support with Optimized Performance
vLLM brings day-0 support to TML Inkling, a 1T-parameter multimodal model, with MTP, long-context serving, parallelism, and up to 380 tokens per second per user on NVIDIA GB200 GPUs.
EAGLE3 Speculative Decoding on AMD Instinct GPUs: Training and Serving with vLLM and AMD Quark
『推测解码』学习路径中的动手实践深化篇:在掌握 speculative decoding 基本原理与 EAGLE 思路后阅读,可学到 EAGLE3 从训练、量化到服务落地的完整工程闭环;之后再读 2026-07-28 并行起草一文,了解超越 EAGLE-3 自回归限制的 P-EAGLE、DFlash、DSpark。
vime + ROCm: End-to-End RL Post-Training on AMD Instinct™ GPUs
Announcing ROCm support for vime, now running end-to-end on AMD Instinct MI355X GPUs with prebuilt container.
vLLM × HPC-Ops: High-Performance Attention and MoE Backends from Tencent Hunyuan
How HPC-Ops integrates Hopper-optimized attention and FP8 MoE backends into vLLM for Tencent Hunyuan Hy3, improving mixed-length decode, MoE latency, TTFT, and TPOT on NVIDIA H20.
Experience and Lessons Learned from Serving Multi-Stage Qwen3-Omni in vLLM-Omni
How vLLM-Omni serves and optimizes Qwen3-Omni with staged Thinker-Talker-Code2Wav execution, batching, CUDA Graphs, async chunk, async output, replicas, hot-path cleanup, and perf validation.
Engineering TTS Inference in vLLM-Omni
How vLLM-Omni supports and optimizes Qwen3-TTS, VoxCPM2, Higgs Audio V3, and Fish Speech S2 Pro with staged serving, batching, CUDA Graphs, and model-specific kernels.
Beyond One Model: Fusion in vLLM Semantic Router
How vLLM Semantic Router Fusion runs a panel of models, uses a judge to analyze agreement and gaps, and synthesizes one answer while preserving routing policy, traces, and OpenAI-compatible serving.
MiniMax M3 in vLLM: Day-0 Serving for 1M-Token Multimodal Reasoning
How vLLM serves MiniMax M3 with MiniMax Sparse Attention, multimodal and reasoning parsers, MXFP8 weights, and long-context deployment recipes.
DiffusionGemma: The First Diffusion LLM (dLLM) Natively Supported in vLLM
How vLLM supports DiffusionGemma, the first native diffusion language model in vLLM, using Model Runner V2 state hooks, iterative denoising, bidirectional attention, and reused speculative decoding paths.
Announcing Day-0 Support for NVIDIA Nemotron 3 Ultra on vLLM
How to serve NVIDIA Nemotron 3 Ultra with vLLM for long-running agentic reasoning, including BF16 and NVFP4 checkpoints, supported GPU configurations, OpenAI-compatible deployment, and NeMo RL integration.
vLLM on the DGX Spark: Architecture, Configuration, and Local Evaluation
How to run vLLM on NVIDIA DGX Spark and GB10 systems, including unified memory behavior, NVFP4 Nemotron-3-Super serving, Docker deployment, Prometheus metrics, and local evaluation results.
EAGLE 3.1: Advancing Speculative Decoding Through Collaboration Between the EAGLE Team, vLLM, and TorchSpec
『推测解码』路径的第三篇（进阶收尾，也是当前最新进展）：默认读者已掌握前两篇的推测解码原理与 Eagle3 hidden-state 机制，聚焦 EAGLE 3.1 对 attention drift 的架构级修复与生产落地，建议最后阅读。
Elastic Expert Parallelism in vLLM
How Elastic Expert Parallelism lets vLLM scale Mixture-of-Experts serving up or down at runtime by changing data-parallel workers, redistributing experts, and coordinating live topology changes without server restarts.
vLLM Tops the Artificial Analysis Leaderboard
How vLLM achieved leading Artificial Analysis results for DeepSeek V3.2, MiniMax-M2.5, and Qwen 3.5 397B using open-source kernel fusion, speculative decoding, Blackwell optimizations, and model-specific serving work.
Serving Agentic Workloads at Scale with vLLM x Mooncake
How vLLM integrates Mooncake Store as a distributed KV cache for agentic workloads, reusing shared prefixes across turns and instances to improve throughput, TTFT, end-to-end latency, and multi-GPU scaling.
DeepSeek V4 in vLLM: Efficient Long-context Attention
A first-principles walkthrough of DeepSeek V4's long-context attention, and how we implemented it in vLLM.
The State of FP8 KV-Cache and Attention Quantization in vLLM
What vLLM FP8 KV-cache validation found across Hopper and Blackwell, covering attention quantization, Flash Attention 3 fixes, memory savings, decode speedups, and layers to skip.
Disaggregated Serving for Hybrid SSM Models in vLLM
『PD 分离服务』路径的进阶深化篇:建议先读 vLLM 基础 PD 分离文章(介绍 NIXL connector 的 v0-disagg-prefill)掌握标准流程,再读本文理解 hybrid SSM 模型带来的底层改造;可与 AFD Plugin(按组件拆分)、TileRT(decode 可插拔)两篇生态文对照阅读。
vLLM Korea Meetup 2026 Wrap-Up
What the vLLM Korea Meetup 2026 covered: community growth, vLLM V1 updates, production stack adoption, accelerator integration, vllm-playground, and real-world LLM serving.
Model Runner V2: A Modular and Faster Core for vLLM
How Model Runner V2 reworks vLLM's execution core with modular model logic, GPU-native input preparation, stable persistent batching, async-first scheduling, and no API changes.
P-EAGLE: Faster LLM inference with Parallel Speculative Decoding in vLLM
推测解码学习路径中的深度原理压轴篇:建议先读 Laguna XS.2 + DFlash 案例建立直观认识,再读 Speculators v0.5.0 掌握训练生态,最后读本文深入并行起草的架构、训练算法与 vLLM 内核级实现。
vLLM Semantic Router v0.2 Athena: ClawOS, Model Refresh, and the System Brain
What vLLM Semantic Router v0.2 Athena adds: refreshed multilingual and multimodal routing models, ONNX and ROCm acceleration, safety and memory signals, long-context handling, and ClawOS orchestration.
vLLM Triton Attention Backend Deep Dive
A technical walkthrough of the vLLM Triton attention backend, covering performance-portable paged attention kernels, backend selection, autotuning, CUDA graph behavior, benchmarks, and NVIDIA, AMD, and Intel support.
Beyond Porting: How vLLM Orchestrates High-Performance Inference on AMD ROCm
How vLLM orchestrates high-performance inference on AMD ROCm with multiple attention backends, workload-aware prefill, extend, and decode routing, AITER primitives, MLA support, and MI300X-class benchmarks.
Efficiently serve dozens of fine-tuned models with vLLM on Amazon SageMaker AI and Amazon Bedrock
How vLLM serves many fine-tuned MoE and dense models with Multi-LoRA, including fused MoE LoRA kernels, Triton compiler fixes, Split-K and CTA swizzling optimizations, and SageMaker AI and Bedrock tuning.
DeepSeek-V3.2 on GB300: Performance Breakthrough
What DeepSeek-V3.2 and DeepSeek-R1 benchmark results show on NVIDIA GB300 with vLLM, covering NVFP4 quantization, TP and EP deployment, throughput, and reproducible setup details.
Driving vLLM WideEP and Large-Scale Serving Toward Maturity on Blackwell (Part I)
How vLLM improves WideEP and large-scale DeepSeek-style MoE serving on NVIDIA GB200 with NVFP4 and FP8 kernels, fusion, prefill/decode disaggregation, weight offloading, and reduced chunking overhead.
GPT-OSS Performance Optimizations on NVIDIA Blackwell: Pushing the Pareto Frontier
How vLLM and NVIDIA optimized GPT-OSS on Blackwell with FlashInfer, torch.compile fusion, FP8 KV cache, async scheduling, stream interval tuning, and deployment recipes that improve throughput and interactivity.
Inside vLLM’s New KV Offloading Connector: Smarter Memory Transfer for Maximizing Inference Throughput
How vLLM's asynchronous KV offloading connector stores KV cache in CPU memory to reduce recomputation, improve throughput under memory pressure, and support pluggable offload backends.
vLLM Semantic Router v0.1 Iris: The First Major Release
What vLLM Semantic Router v0.1 Iris introduces: signal-decision plugin architecture, model selection, safety filtering, semantic caching, hallucination detection, LoRA-based routing models, and production-ready MoM routing.
vLLM-Omni Diffusion Cache Acceleration
How vLLM-Omni speeds up diffusion model inference with Cache-DiT and TeaCache, reusing intermediate computations across timesteps to deliver 1.5x to 2x image generation speedups with minimal quality loss.
vLLM Large Scale Serving: DeepSeek @ 2.2k tok/s/H200 with Wide-EP
How vLLM reaches 2.2k tokens per second per H200 for DeepSeek-style MoE serving with Wide-EP, async scheduling, dual-batch overlap, disaggregated serving, CUDA graphs, DeepGEMM, and expert load balancing.
Run Highly Efficient and Accurate AI Agents with NVIDIA Nemotron 3 Nano on vLLM
How to serve NVIDIA Nemotron 3 Nano with vLLM for efficient agentic AI, including BF16, FP8, and NVFP4 checkpoints, 1M-token context, hybrid MoE architecture, Thinking Budget, supported GPUs, and OpenAI-compatible deployment.
Tracing Hanging and Complicated GPU Kernels Down To The Source Code
How vLLM developers debug hanging and complex CUDA kernels by triggering GPU core dumps, identifying stuck kernels, and mapping failures back to source code lines for faster kernel debugging.
Fast and Affordable LLMs serving on Intel Arc Pro B-Series GPUs with vLLM
How vLLM serves LLMs on Intel Arc Pro B-Series GPUs with MoE optimizations, persistent kernels, multi-GPU scaling, LoRA, speculative decoding, structured outputs, and mixed-precision recipes.
No More Train-Inference Mismatch: Bitwise Consistent On-Policy Reinforcement Learning with vLLM and TorchTitan
How vLLM and TorchTitan demonstrate bitwise consistent on-policy RL by matching training and inference numerics, using batch-invariant kernels to reduce train-inference mismatch and stabilize reinforcement learning.
From Monolithic to Modular: Scaling Semantic Routing with Extensible LoRA
How vLLM Semantic Router refactors its Rust classification layer with modular model support, Qwen3-Embedding, EmbeddingGemma, LoRA-based multi-task classification, and concurrent routing execution.
Zero-Reload Model Switching with vLLM Sleep Mode
How vLLM Sleep Mode enables fast model switching by hibernating weights to CPU RAM or discarding them while preserving process state, CUDA graphs, allocators, and kernel warmup to avoid full reloads.
vLLM TPU: A New Unified Backend Supporting PyTorch and JAX on TPU
How the redesigned vLLM TPU backend uses tpu-inference, JAX-to-XLA lowering, Torchax, ragged paged attention, and unified PyTorch and JAX support to improve TPU performance and model coverage.
SemiAnalysis InferenceMAX: vLLM and NVIDIA Accelerate Blackwell Inference
How vLLM and NVIDIA optimize Blackwell inference for SemiAnalysis InferenceMAX, improving gpt-oss 120B and Llama 3.3 70B throughput with FP4 kernels, scheduling work, and Pareto-frontier benchmarking.
DeepSeek-V3.2-Exp in vLLM: Fine-Grained Sparse Attention in Action
How vLLM supports DeepSeek-V3.2-Exp with DeepSeek Sparse Attention, lightning indexer caches, separate prefill and decode paths, FlashMLA sparse attention, DeepGEMM kernels, and Blackwell deployment.
vLLM Now Supports Qwen3-Next: Hybrid Architecture with Extreme Efficiency
How vLLM supports Qwen3-Next with hybrid attention, Gated DeltaNet, full attention, high-sparsity MoE, multi-token prediction, hybrid KV cache management, Triton kernels, and CUDA graphs.
Inside vLLM: Anatomy of a High-Throughput LLM Inference System
How vLLM's inference engine works, covering PagedAttention, continuous batching, prefix caching, speculative decoding, multi-GPU serving, scheduling, and benchmarking for high-throughput LLM workloads.
Introduction to torch.compile and How It Works with vLLM
How torch.compile works inside vLLM, including TorchDynamo graph capture, TorchInductor code generation, custom compiler passes, graph breaks, model compilation strategy, and performance optimization.
CUDA Core Dump: An Effective Tool to Debug Memory Access Issues and Beyond
How to debug vLLM CUDA illegal memory access errors with CUDA core dumps, environment variables, cuda-gdb, and GPU state inspection when Python stack traces or CUDA_LAUNCH_BLOCKING are insufficient.
vLLM Now Supports gpt-oss
How vLLM supports gpt-oss 20B and 120B on NVIDIA Blackwell, Hopper, and AMD GPUs, with MXFP4 MoE kernels, efficient attention, hybrid KV cache allocation, and built-in tool support.
MiniMax-M1 Hybrid Architecture Meets vLLM: Long Context, Fast Inference
How vLLM serves MiniMax-M1's hybrid MoE architecture for long-context inference, covering model deployment, memory management, batched serving, backend optimizations, and Docker-based setup.
PTPC-FP8: Boosting vLLM Performance on AMD ROCm
How PTPC-FP8 quantization improves vLLM performance on AMD ROCm by combining per-token activation scaling and per-channel weight scaling for near-BF16 accuracy with FP8 speed.
vLLM V1: A Major Upgrade to vLLM's Core Architecture
What changed in vLLM V1: a re-architected engine with a simpler scheduler, near-zero-overhead prefix caching, cleaner tensor parallelism, multiprocessing API server, and default optimizations for higher-throughput serving.
Installing and Developing vLLM with Ease
How to install and develop vLLM using stable releases, nightly wheels, uv, source builds, Python and C++/CUDA workflows, and version tracking for production deployments.
Serving LLMs on AMD MI300X: Best Practices
Best practices for serving LLMs with vLLM on AMD MI300X, covering ROCm setup, Llama 3.1 70B and 405B benchmarks, chunked prefill, multi-step scheduling, prefix caching, graph capture, and AMD tuning.
vLLM’s Open Governance and Performance Roadmap
vLLM's open governance and performance roadmap, covering LF AI and Data incubation, public benchmarks, optimized kernels, async scheduling, API frontend overhead, torch.compile, disaggregated prefill, and community research.
Notes on vLLM v.s. DeepSpeed-FastGen
A performance comparison of vLLM and DeepSpeed-FastGen, explaining when Dynamic SplitFuse helps, where vLLM is faster, and how memory allocation, output length, and workload shape affect throughput.
vLLM: Easy, Fast, and Cheap LLM Serving with PagedAttention
What the original vLLM launch announced: PagedAttention for KV cache management, up to 24x throughput over Hugging Face Transformers, and lower-cost high-throughput LLM serving.