[KongchangAI]
Unverified50% confidenceTradeoffExact time

vLLM's preemption recovery mechanism offers two strategies—recomputation and swapping—where recomputation requires no additional storage but has high recovery latency, while swapping recovers faster but is limited by PCIe bandwidth.

1
Sources
50%
Confidence
Long-term
Relevance
8/7/2026
First Seen

Sources

vLLM架构深度解析：PagedAttention如何实现高吞吐LLM推理

Related Claims

Unverified主机卸载用带宽换显存空间，代价是数据传输延迟；激活重计算用算力换显存，二者可互补配合69% similarUnverifiedHNSW generally achieves higher recall but has large memory consumption, while IVF builds faster but requires more nprobe parameter tuning.64% similarUnverifiedHyperscale 将计算层与存储层解耦，通过 RDMA 技术实现计算节点与存储节点间低延迟通信63% similar

Cite This Claim

Stable URI
https://kongchang.com/claim/705563
API
curl https://kongchang.com/api/v1/knowledge/claims/705563
MCP
get_claim(id=705563)