Inference

Serve models from the same cluster you train on.

Konduktor schedules training and inference on the same GPUs. Inference scales with traffic, training fills the gaps.

job.yaml

num_nodes: 64

accelerators: H100:8

priority-class: high-priority

serving:

min_replicas: 3

8x GPUs
Training
8x GPUs
Training
8x GPUs
Inference
llama-70b-serve Live
Replicas: 3 → 5 (autoscaling)
Latency: 42ms | 1.2K req/sec
Health: All nodes passing
Resource Allocation

One cluster. Training and inference, always balanced.

Training and inference share one GPU pool. When traffic spikes, the scheduler shifts GPUs to inference; when it drops, they go back to training. Zero traffic means zero GPUs.

Training● high
448 GPUs allocated
Inference↑ spiking
64 GPUs allocated
Rebalancing
64 GPUs → Inference
2m 5s
Training (448)
Inference (64)
Relocating (64)
Auto-scaling

Scale to zero. Actually zero.

Not just scale-down-to-one. No requests means no GPUs used. Those resources go straight back to training jobs. When requests come in, they scale back up automatically.

GPU Allocation512 GPUs total
Training
468 GPUs
Inference
16 GPUs
Inference Scaling Down
12 req/s
Reallocated44 → Training
Utilization98.2%
Cluster512 GPUs

Deploy any model, zero code changes

Open source or your own custom fine-tuned model. Zero code changes required, same workflow as the rest of Konduktor.

Model Registry
3 deployed
🦙
Llama 3.1 70BOpen source · vLLM
● Live
Custom Fine-tuned 7BPrivate · your-org/ft-model
● Live
Mistral 8x7B MoEOpen source · TGI
Deploying

Every endpoint monitored, out of the box

Public-facing endpoints are created automatically for every deployed model. Latency, throughput, and GPU utilization are tracked in real time. No extra setup, no third-party dashboards.

Inference
Llama 3.1 70B
● Live
llama-70b.trainy.ai/v1/completions
1,284 req/s
8x H100
42ms p99
Endpoint Metricslast 24h
1,284 req/sThroughput
42msLatency
99.98%Uptime
Comparison

Your cluster. Not ours.

On other platforms you rent their infrastructure. Konduktor serves models from your own cluster, so your data never leaves your environment.

OthersTrainy
Infrastructure
Shared, vendor-managedYour cluster, your VPC
Data Privacy
Leaves your environmentNever leaves your network
Scaling
Capped by provider limits1000s of GPUs, on demand
Commitment
Year-long contractsPay only when training
Vendor Lock-in
Tied to their platformCloud-agnostic, portable

Ready to scale your AI training? Get enterprise-grade GPU infrastructure up and running in 20 minutes.