Unverified50% confidenceFactExact time
Model Runner V2 在 vLLM v0.28.0 中获得了 E/P/D 分离（Encode/Prefill/Decode 解耦部署）和权重卸载两项能力
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
9/12/2026
First Seen
Valid until: 12/11/2026
Sources
Related Entities
Related Claims
UnverifiedvLLM v0.28.0rc2引入了名为DFlash2的投机解码新特性，该特性结合了局部卷积与候选选择器两大技术组件（PR #52816）68% similarUnverifiedModel Runner V2（MRV2）能力扩展至非生成式工作负载，支持仅编码器注意力、序列池化、BGE-M3 池化及 CPU 上运行多模态66% similarUnverifiedvLLM v0.27.0 引入简化版容错框架，专为 DP+EP 外部负载均衡器部署场景设计，并提供弹性 EP 扩展的异步准备能力66% similarUnverifiedvLLM v0.28.0 在推测解码领域引入了 DFlash2 以及 DSpark 置信度调度验证（confidence-scheduled verification）63% similarUnverifiedvLLM宣布在AMD GPU上支持推测解码(Speculative Decoding)技术60% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/912417API
curl https://kongchang.com/api/v1/knowledge/claims/912417MCP
get_claim(id=912417)