cedana / use cases · elastic inference

Run inference elastically.
On your infrastructure.

Every GPU-hour either produces SLA-compliant tokens or it's waste. Cedana's elastic provisioning reclaims the wasted hours and turns them into lower cost per token.
60%
less idle capacity
vs static overprovisioning
10×
faster cold starts
600 s → 60 s
0
code changes
drop-in helm install
Mechanism · Status: LIVE

One primitive: GPU workload portability.

Cedana checkpoints running inference workers, capturing the entire state — memory, KV cache, CUDA context, container — and resumes them on a different GPU.

native · 50% overprovisioned● steady
provisionedidle (waste)demand
workers8/8 provisioned · 6/8 serving
cold start600 s
idle / waste52%
slaok
cedana · 20% buffer● tracking
provisionedidle (waste)demand
workers7/8 provisioned · 6/8 serving
cold start60 s · 10× faster
idle / waste21%
slaok
How it decides

Inside the loop: the policy engine.

What's happening inside the chart above. Cedana's policy engine elastically resizes warm GPU buffers to match live demand — scaling capacity up before a spike, releasing it after. You set the policy. Cedana provisions to track actual load.

policy engine · track-demand● ACTIVE
inputs
demand
64 %
p99 lat
146 ms
queue
15 req
$/1k tok
0.00055
actions logpolicy → live
10:32:12warm +24 → 6 gpu
10:32:15·holdp99 ok
10:32:18drain 1kv preserved
10:32:21warm +1spike forecast
spikewarm +N
steadyhold
dropdrain · kv preserved
What you get

Unlock operational leverage.

01

Resilience for mission-critical workloads

When a worker is lost to failure or preemption, Cedana resumes it automatically — keeping valuable KV cache and in-flight sessions intact.

02

Less overprovisioning

Match provisioning closer to demand with 2–10× faster cold starts.

03

Idle GPUs get to work

Training and inference share the same GPUs. Long jobs immediately yield to inference spikes, with no lost work.

04

Right workload, right GPU

Dynamically migrate running inference workers across nodes without losing state. Put latency-critical traffic on faster GPUs, batch work on cheaper ones, and rebalance whenever cost or demand shifts.

Install · Status: MINUTES

Seamless integration.

Integrates with a Helm chart in minutes. No code or config changes. Supports NVIDIA Dynamo.

the result● shipped
01reliability
Mission-critical SLAs
● met under load
02economics
Lower cost per token
↓ less waste, more tokens
03ownership
On your infrastructure
your stack, your control
~ / cedana / deploy ready

Command your
compute.

Your compute, liquid — checkpoint, migrate, and resume live GPU jobs across the fleet.

deployK8s helm chart · SLURM plug-in
first migration<30 min
code changes0
workloadstraining · inference · HPC
01
AWS
02
Google Cloud
03
NVIDIA
04
K8s
05
SLURM
06
Nvidia Dynamo