vLLM releases v0.22.0 with 28.9% FP8 latency cuts and KV offloading

• vllm_project's release post says vLLM 0.22.0 landed with 459 commits from 230 contributors, including 63 first-time contributors.
• The headline perf change is a batch-invariant Cutlass FP8 path that, according to vllm_project's hardware thread, cuts end-to-end latency by 28.9%.
• vllm_project's engine-core breakdown also turns DeepSeek V4 support into a more serious serving path, with NVFP4 fused MoE, fuller CUDA graph support, ROCm work, and multi-tier KV cache offloading.
• The upgrade is not frictionless: vllm_project's upgrade notes removes deprecated tokenizer and chat-template locations, drops old MLA prefill arguments, and clears out dead CUDA code.

The release thread is unusually dense for a point update. You get engine changes like an in-tree Rust frontend and Model Runner V2 work, hardware-specific kernels for Blackwell, ROCm, CPU, and RISC-V, plus new model architecture support for MiniCPM-V 4.6, InternS2 Preview, and OpenVLA.

Most of the release weight sits in the serving stack, not model count. vllm_project's engine-core breakdown says DeepSeek V4 now lives in a dedicated package with NVFP4 fused MoE, full and piecewise CUDA graphs, MTP speculative decoding on ROCm, and more fused kernels.

The same post adds four Model Runner V2 changes:

• oracle selection for Qwen3 dense by default
• sleep-mode weight reload
• update_config
• shared KV-cache layers

There is also an escape hatch. When KV connectors are present, the engine-core thread says the runtime falls back to Model Runner V1.

The cleanest number in the release is the FP8 latency claim. vllm_project's hardware thread attributes a 28.9% end-to-end improvement to the batch-invariant Cutlass FP8 path, and adds compile-mode support on SM80 plus an NVFP4 Cutlass linear path.

The hardware matrix also widened:

• Blackwell gets FlashInfer MoE and FP4 GEMM on SM120 and SM121, plus per-tensor FP8 CUTLASS on SM12.1, per the hardware thread.
• ROCm picks up more DeepSeek V4 functionality, flash sparse MLA Triton kernels, and gluon paged MQA logits, again in the same post.
• CPU and RISC-V get RVV-optimized attention kernels, fused GDN for AMX, and experimental Triton with MRv2, according to vllm_project.

Cache management also moved. The engine-core thread says KV cache offloading now supports multiple tiers, including a Python filesystem secondary tier, DeepSeek V4 support, and Mooncake disk offloading.

The release adds support for MiniCPM-V 4.6, InternS2 Preview, and OpenVLA, according to vllm_project's model-support post. The same thread also calls out speculative decoding updates, including custom callable proposers and Gemma3 and Gemma4 multi-GPU fixes with batched vision encoders.

The migration gotchas are short but real. vllm_project's upgrade notes says vLLM removed deprecated get_tokenizer and resolve_hf_chat_template locations, eliminated deprecated MLA prefill arguments, and deleted dead CUDA kernels and code.

One day later, vllm_project's Laguna XS.2 post showed the team already using the release as a serving showcase: a DFlash speculator drafting eight tokens per forward pass for Laguna XS.2, with claimed 2x to 3x faster decoding and quantized FP8, NVFP4, and INT4 checkpoints.