gke-inference

GKE AI/ML Inference

This reference covers deploying AI/ML inference workloads on GKE using Google's Inference Quickstart (GIQ) and best practices for LLM serving.

MCP Tools: apply_k8s_manifest, get_k8s_resource, get_k8s_logs, get_k8s_rollout_status, describe_k8s_resource, list_k8s_events. CLI-only: gcloud container ai profiles *

When to Use

• Deploy an AI model (Llama, Gemma, Mistral, etc.) to GKE
• Generate optimized Kubernetes manifests for inference
• Select GPU/TPU accelerators for model serving
• Configure autoscaling for LLM inference

Prerequisites

• A golden path GKE Autopilot cluster (GPU workloads are supported via

ComputeClasses and NAP)

• gcloud CLI authenticated
• Sufficient GPU/TPU quota in the target region

Workflow

1. Discovery: Find Models and Hardware

# List all supported models
gcloud container ai profiles models list --quiet

# Find valid accelerator/server combinations for a model
gcloud container ai profiles list --model=<MODEL_NAME> --quiet

# Example: what can run Gemma 2 9B?
gcloud container ai profiles list --model=gemma-2-9b-it --quiet

2. Generate Manifest

gcloud container ai profiles manifests create \
  --model=<MODEL_NAME> \
  --model-server=<SERVER> \
  --accelerator-type=<ACCELERATOR> \
  --target-ntpot-milliseconds=<NTPOT> --quiet > inference.yaml

Parameters:

• --model: Model ID (e.g., gemma-2-9b-it, llama-3-8b)
• --model-server: Inference server (vllm, tgi, triton, tensorrt-llm)
• --accelerator-type: GPU/TPU type (nvidia-l4, nvidia-tesla-a100,

nvidia-h100-80gb)

• --target-ntpot-milliseconds: Target Normalized Time Per Output Token

(optional, for latency optimization)

Example:

gcloud container ai profiles manifests create \
  --model=gemma-2-9b-it \
  --model-server=vllm \
  --accelerator-type=nvidia-l4 \
  --target-ntpot-milliseconds=50 --quiet > inference.yaml

3. Review and Deploy

# Review for placeholders (HF tokens, PVCs)
cat inference.yaml

# Deploy
kubectl apply -f inference.yaml

# Monitor
kubectl get pods -w
kubectl logs -f <POD_NAME>

Some models require Hugging Face tokens. Create a Kubernetes Secret and reference it in the manifest.

GPU ComputeClass for Inference

For Autopilot clusters, create a ComputeClass to target GPU nodes:

apiVersion: cloud.google.com/v1
kind: ComputeClass
metadata:
  name: l4-inference
spec:
  priorities:
  - machineFamily: g2
    gpu:
      type: nvidia-l4
      count: 1
    minCores: 4
    minMemoryGb: 16

Accelerator Selection Guide

: : lightweight legacy : : : : : models : : : | NVIDIA L4 (G2) | Small-medium model | 24 GB | Low | : : inference, video, : : : : : graphics : : : | NVIDIA RTX PRO 6000 | Multimodal AI, | 96 GB | Medium | : (G4) : high-fidelity 3D, : : : : : fine-tuning : : : | Cloud TPU v5e | Cost-effective | Varies | Medium | : : transformer inference : : : | Cloud TPU v5p | High-performance | Varies | High | : : training : : : | Cloud TPU v6e | High-efficiency next-gen | 32 GB/chip | Medium-High | : (Trillium) : training & serving : : : | Cloud TPU v7x | Ultra-scale inference & | 192 GB/chip | High | : (Ironwood) : agentic workflows : : : | NVIDIA A100 | Large model inference, | 40/80 GB | High | : : enterprise ML : : : | NVIDIA H100 / H200 | Frontier model training, | 80/141 GB | Highest | : : high throughput : : : | NVIDIA B200 (A4) | Blackwell-scale | 192 GB | Highest | : : training, FP4 precision : : : | NVIDIA GB200 (A4X) | Rack-scale AI (Grace | Massive | Highest | : : Blackwell Superchip) : : :

Autoscaling LLM Inference

GPU-based autoscaling

Use custom metrics for GPU utilization:

apiVersion: autoscaling/v2
kind: HorizontalPodAutoscaler
metadata:
  name: llm-hpa
spec:
  scaleTargetRef:
    apiVersion: apps/v1
    kind: Deployment
    name: llm-server
  minReplicas: 1
  maxReplicas: 10
  metrics:
  - type: Pods
    pods:
      metric:
        name: gpu_duty_cycle
      target:
        type: AverageValue
        averageValue: "80"

Best practices for inference autoscaling

1. Use DCGM metrics: Golden path enables DCGM monitoring for GPU

utilization metrics

1. Set appropriate minReplicas: At least 1 for always-on serving; 0 for

batch/on-demand

1. Tune scale-down delay: LLM model loading is slow; use longer

stabilization windows

1. Consider queue depth: Scale on pending requests rather than pure GPU

utilization for latency-sensitive workloads

Optimization Tips

• Quantization: Use quantized models (GPTQ, AWQ) to reduce GPU memory and

increase throughput

• Batching: Configure model server batch size for throughput vs latency

trade-off

• Tensor parallelism: Split large models across multiple GPUs within a

node

• KV cache optimization: Tune --gpu-memory-utilization in vLLM for KV

cache allocation

Troubleshooting

: model/accelerator : : profiles list : : combination : : --model=<MODEL>` : | GPU quota exceeded | Regional quota limit | Request quota increase or | : : : try a different region : | OOM on GPU | Model too large for | Use larger GPU, enable | : : accelerator : quantization, or use tensor : : : : parallelism : | Slow cold start | Large model loading from | Use local SSD for model | : : registry : caching; pre-pull images :