Ollama

LLM inference server on RTX 5070 Ti GPU.

Overview

PropertyValue
CDK8s fileplatform/cdk8s/cots/ai/ollama.go
Namespaceollama
Helm chartollama v1.41.0 (otwld.github.io/ollama-helm)
HTTPRouteollama.madhan.appollama:11434
UINo (REST API only)
Nodek8s-worker4 (GPU)

Purpose

Runs open-source LLMs locally using the NVIDIA RTX 5070 Ti. Exposes an OpenAI-compatible REST API at http://ollama.madhan.app.

Supported models include: llama3.2, mistral, deepseek-r1, and any model in the Ollama library.

GPU Configuration

runtimeClassName: nvidia
resources:
  requests:
    memory: "2Gi"
  limits:
    nvidia.com/gpu: "1"
env:
  - name: NVIDIA_VISIBLE_DEVICES
    value: all

API Usage

# List available models
curl http://ollama.madhan.app/api/tags

# Run inference (streaming)
curl http://ollama.madhan.app/api/generate \
  -d '{"model": "llama3.2", "prompt": "Hello!"}'

# Pull a new model
curl http://ollama.madhan.app/api/pull \
  -d '{"name": "mistral"}'

Node Scheduling

The pod is scheduled to k8s-worker4 via a nodeSelector or nodeName constraint targeting the GPU node.

Coexistence with ComfyUI

Ollama (2 Gi RAM) and ComfyUI (1 Gi RAM request) run simultaneously on the same node via GPU time-slicing. Total RAM on node is ~5.4 Gi allocatable — the combined 3 Gi fits within limits.