78 posts
vLLM supports Kimi K3.1 Pro on day 0: sparse attention on a KDA hybrid, a packed KV cache with FP8 or NVFP4 latents and MXFP4 index keys, token-sharded prefill, decode context parallelism, and fused low-latency decode.
With P/D disaggregation and Wide-EP on a GB300 NVL72 rack, vLLM serves DeepSeek-V4.1-Flash on SemiAnalysis AgentX at 954K tokens/s per GPU at a P90 of 196 tokens/s per user, 7.9x our aggregated single-node result.
vLLM serves GLM-5.3 on 16 H200s with PCP and DCP prefill and Hybrid HiSparse decode, reaching 1M-token contexts and 1.7–3.2× the per-user speed of data-parallel P/D on the AgentX agentic benchmark.
Day-0 vLLM support for MiniMax-M3.1-Flash on NVIDIA Blackwell, NVIDIA Rubin, and AMD MI350-series GPUs, with NVFP4 MSA kernels, NVFP4 MegaMoE, DSpark speculative decoding, AgentX results, and day-0 RL.
vLLM supports both Flint checkpoints (40B total, 2.5B active) from day 0, decoding a single request at about 500 tokens/s on one GPU.
Introducing Humming, a quantized inference kernel library from Ant Group and the vLLM community, with broad GPU and quantization support, dense and MoE performance results, and vLLM integration.
How agent-assisted CuTe DSL optimizations accelerate Kimi K3 decode on NVIDIA GB300, and how scoped changes, correctness tests, and reproducible benchmarks help ship them responsibly.
Inside the Kimi K3 optimization effort: understanding the bottlenecks and the decisions that made inference faster
vLLM supports Reflection's Beam-1.0 (501B total, 23B active) from day 0, with agent-driven optimization that cut decode latency 2.71× and DFlash2 speculative decoding with adaptive verification.
An online-fitted acceptance estimator that lets every draft-model speculator use Adaptive Verification, by predicting the target's acceptance probability from the drafter's own logits and refitting continuously during serving.
A push-based NIXL connector for vLLM disaggregated serving: the prefill instance writes KV directly into decode memory over RDMA, taking the router relay, decode-side allocation, and transfer setup off the TTFT critical path.
Bi-directional KV transfer lets vLLM prefill nodes reuse previously computed KV from decode nodes on later conversation turns, cutting redundant prefill recompute and reducing TTFT by up to 3.3x on long multi-turn prompts.
vLLM-Omni runs Sand AI's MAGI-2 Preview audio-video generation model on eight Moore Threads MTT S5000 GPUs through torchada's CUDA compatibility layer, then optimizes MoE, mHC, attention, parallelism, and compilation.
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.
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.
How vLLM implements distribution-preserving Gumbel-max text watermarking with efficient GPU kernels, statistical detection, speculative decoding, and repeated-context safeguards.
Shipping Python bytecode saved about 18 seconds per startup in H100 server benchmarks, with images and models already local.
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.
How vLLM reaches 5K throughput and 180 interactivity on Qwen3.8-2.4T with GB300 NVL72 PD serving and how to reproduce results yourself.
How to leverage NVIDIA Hardware Video Decoders to Achieve Multi-GPU Scaling in Video Captioning and Description tasks.
How Speculators and Mooncake enabled multi-node DSpark training for Kimi K3.
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.
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 serving optimizations across scheduling, KDA prefix caching, ReplaySSM state recovery, PD disaggregation and state offload, parallelism, MoE, and GPU kernels.
A performance model for LLM serving: inspect local shapes, remove repeated work, verify data movement and dispatch, then follow the queue.
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.
How vLLM optimizes KV cache management, parallelism, scheduling, and P/D disaggregation for agentic workloads, validated on SemiAnalysis AgentX with up to 130K tokens per GPU-second and a 14.6x-106x serving-cost advantage over Opus 5.
How vLLM-Omni optimizes and scales the complete MiniMax H3 stack, then integrates FastVideo’s four-step FastH3 for generation faster than playback.
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 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.
A release focused on higher-throughput diffusion rollout, reusable omni adapters, and broader recipe coverage.
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.
Sizing the DSpark draft-verification budget from per-request confidence instead of verifying every drafted token, so one configuration holds the throughput/latency frontier from batch size 1 to 256.
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.
Decode Context Parallelism (DCP) in vLLM shards KV cache across GPUs by sequence dimension, enabling 3× higher throughput on long-context agentic workloads compared to standard tensor parallelism.
How vLLM reaches 25K total TPS/GPU on Qwen3.5-397B-A17B-NVFP4 with GB200 NVL72 disaggregated serving, Blackwell GDN kernels, HMA cache transfer, async scheduling fixes, and srt-slurm recipes.
An overview of Arm CPU enablement and inference performance optimizations in 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.
How we took GLM-5.2-NVFP4 from 40 ms to 17 ms mean TPOT on 24 B300 GPUs with vLLM: P/D disaggregation, MTP speculative decoding, Model Runner V2, and the SLA-first trade-offs behind the final configuration.
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.
How vLLM maintains production quality with extensive CI across diverse accelerators, nightly performance benchmark and accuracy evaluation, and a two-week release process.
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.
vLLM prefill paired with TileRT decode through vLLM V1's connector interface: a specialized, latency-optimized decode engine that coexists with native vLLM decode behind one shared serving layer, with zero changes to vLLM.
How AMD Quark trains, quantizes, and serves EAGLE3 speculative-decoding drafts with vLLM on AMD Instinct GPUs, delivering up to 2.00x throughput gains for Kimi-K2.5 and 1.79x for MiniMax-M2.5.
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.
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.
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.
How Session-Aware Agentic Routing in vLLM Semantic Router preserves long-horizon agent continuity with session memory, safe model-switch boundaries, prefix-cache-aware switch pricing, and replayable traces.
How vLLM Semantic Router hardens multimodal routing by turning visual evidence into trustworthy signals, debugging a Rust/Candle vision-encoder parity issue, and validating image signal correctness for production policy.
How EAGLE 3.1 improves speculative decoding robustness in vLLM with FC normalization, post-norm hidden-state feedback, TorchSpec training support, and config-driven compatibility with EAGLE 3 checkpoints.
How PegaFlow integrates with vLLM as an external KV cache service, using a Rust daemon, CUDA IPC, RDMA, SSD caching, and the external KV connector to improve startup, sharing, throughput, and cache lifecycle.
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.
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.
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.
How P-EAGLE brings parallel speculative decoding to vLLM by generating multiple draft tokens in one forward pass, with pre-trained drafter heads, config support, and B200 speedups over EAGLE-3.
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.
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.
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.
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.
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.
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.
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.
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.
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.
How shared memory IPC caching in vLLM reduces redundant data transfers for multimodal and multi-process inference, improving prefill throughput and TTFT by sharing large inputs across coordinator and worker processes.
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.
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.
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.
How torch.compile works inside vLLM, including TorchDynamo graph capture, TorchInductor code generation, custom compiler passes, graph breaks, model compilation strategy, and performance optimization.
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.
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.
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.
How structured decoding works in vLLM, covering JSON outputs, grammar-guided generation, outlines, XGrammar, TPOT improvements, constrained decoding, and agentic workflow use cases.
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.
How speculative decoding works in vLLM, covering EAGLE, Medusa, n-gram proposals, draft and target runners, scheduler and memory-manager changes, and continuous batching for lower token latency.
What changed in vLLM v0.6.0 to improve throughput and latency, including API server isolation, reduced CPU overhead, multi-step scheduling, async execution, and benchmarks against earlier vLLM versions.
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.
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.
