Implementation:Kserve Kserve Batcher Sample Input

Knowledge Sources	Kserve_Kserve
Domains	Kubernetes, Model Serving, Batch Inference
Last Updated	2026-02-13 00:00 GMT

Overview

This file contains sample input data for testing the KServe inference batcher with image classification, providing pixel data in the KServe v1 prediction format.

Description

The file is a JSON document structured as a KServe v1 prediction request with an instances array containing multi-dimensional arrays of floating-point pixel values representing images in CIFAR-10 format (3 channels x 32 x 32 pixels). At 3,276 lines, the file provides a realistically sized payload to demonstrate meaningful batch processing behavior, where the batcher accumulates multiple inference requests and submits them as a single batched call to the model server.

Usage

Use this input file to test the KServe batcher feature. Send it as the request body to an InferenceService endpoint that has batcher annotations enabled. The large payload size helps verify that the batcher correctly handles substantial input data and demonstrates performance improvements from batching.

Code Reference

Source Location

• Repository: Kserve_Kserve
• File: docs/samples/batcher/basic/input.json

Signature

{
   "instances": [
      [
         [
            [
               0.23921573162078857,
               0.24705886840820312,
               0.29411768913269043,
               ...
            ],
            ...  // 32 values per row
         ],
         ...  // 3 channels (RGB)
      ],
      ...  // 32 rows per channel
   ]
}

Import

# Send to an InferenceService with batcher enabled
curl -v -H "Content-Type: application/json" \
  http://${SERVICE_HOSTNAME}/v1/models/${MODEL_NAME}:predict \
  -d @docs/samples/batcher/basic/input.json

I/O Contract

Input Structure

Field	Type	Description
instances	array	Top-level array of input instances (KServe v1 format)
instances[i]	3D array	Single image as [channels][height][width] (3 x 32 x 32)
pixel values	float	Normalized floating-point values (approximately -1.0 to 1.0)

Data Format

Property	Value
Format	KServe v1 Prediction Request
Image Format	CIFAR-10 (3x32x32)
Channels	3 (RGB)
Height	32 pixels
Width	32 pixels
Pixel Type	float64 (normalized)
Total Lines	3,276

Expected Output

Field	Type	Description
predictions	array	Array of classification predictions for each input instance

Usage Examples

# Deploy an InferenceService with batcher annotations
# Then send the sample input
MODEL_NAME=cifar10
SERVICE_HOSTNAME=$(kubectl get inferenceservice ${MODEL_NAME} \
  -o jsonpath='{.status.url}' | cut -d "/" -f 3)

curl -v -H "Content-Type: application/json" \
  -H "Host: ${SERVICE_HOSTNAME}" \
  http://${INGRESS_HOST}:${INGRESS_PORT}/v1/models/${MODEL_NAME}:predict \
  -d @docs/samples/batcher/basic/input.json

Related Pages

• Kserve_Kserve_Triton_MMS_Perf_Test - Performance testing with similar CIFAR-10 input data format