Unverified50% confidenceFactExact time
In Ray Serve, the num_gpus in @serve.deployment(ray_actor_options={"num_gpus": 0.25}) is a logical resource quota declared to the scheduler, and at most 4 actor replicas can reside simultaneously on one GPU.
1
Sources
50%
Confidence
Long-term
Relevance
10/14/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedKubernetes 默认将 GPU 视为不可分割的扩展资源，一个 Pod 请求 1 块 GPU 即独占其全部算力和显存69% similarUnverifiedThe rule of thumb for GPU VRAM requirements is: VRAM is approximately the number of parameters multiplied by 0.5 (Q4 quantization) to multiplied by 1.0 (Q8 quantization), in GB.69% similarUnverifiedGPU computing supply can be divided into four major categories: hyperscale cloud providers, dedicated GPU clouds, computing marketplaces, and co-located owned hardware.69% similarUnverifiedGPU clouds (such as Lambda, RunPod, and CoreWeave) offer raw compute rental, where users deploy their own model weights and are billed per GPU hour.68% similarUnverifiedRun:ai introduces queues, quotas, priorities, and preemption mechanisms on top of Kubernetes, turning GPUs from exclusive devices into pooled resources that can be shared and scheduled by multiple teams.67% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1028885API
curl https://kongchang.com/api/v1/knowledge/claims/1028885MCP
get_claim(id=1028885)