vLLM 0.20.0
#15 in LLM-Inferenz & ServingVllm · v0.20.0 · 4× · tolest 30. Juni 2026
1
Momentum
vLLM 0.20.0 is an Open Source release of the quelloffenen Inference and Serving Engine vLLM (Apache-2.0 license), designed for high throughput and efficient memory management (PagedAttention) in LLM Inference. The release encompasses 752 commits from 320 contributors and introduces initial DeepSeek-V4 support, TurboQuant 2-bit KV-Cache quantization for quadrupled KV-Cache capacity, FA4 as the standard for MLA-Prefill, and a DeepSeek-specific MegaMoE path on Blackwell GPUs. Additionally, the standard toolchain switches to CUDA 13.0, PyTorch 2.11, and HuggingFace Transformers v5.
Momentum-Verloop
19.05.17.08.
Features
| Deployment (Self-host/Cloud) | Self-Hosting via pip/Docker (vllm/vllm-openai Image), Kubernetes, Multi-GPU/Multi-Node; keine eigene Cloud-Hosting-Option |
| Durchsatz/Latenz | TurboQuant 2-Bit-KV-Cache: 4x KV-Kapazität; fused RMSNorm: ca. 2,1% End-to-End-Latenzverbesserung |
| Lizenz | Apache 2.0 (permissiv, kommerzielle Nutzung erlaubt) |
| Plattform | Linux; NVIDIA CUDA 13.0 (Standard), AMD ROCm, Intel XPU, TPU, CPU (inkl. ARM/RISC-V/PowerPC); PyTorch 2.11, Python bis 3.14 |
| Preis | Kostenlos, Open Source (keine Lizenzgebühren) |
| Protokoll-Kompatibilität | OpenAI-kompatibler HTTP-Server (Completions, Chat, Embeddings, Rerank u.a. Endpunkte) |
| Release-Datum | 28. April 2026 (v0.20.0) |
| Unterstützte Modelle/Provider | DeepSeek V4 (initial), Hunyuan v3 (Preview), plus breite HF-Transformers-v5-Modellbasis (Llama, Mistral, Qwen u.a.) |