Production Example

This page covers the production inference pattern with persisted model state and tensor-backed weights.

Production flow

Training -> save pipeline -> deploy inference server -> load pipeline -> predict

Example

use Pml\Pipeline;

$pipeline = Pipeline::load('saved_pipeline');
$dataset = Pml\Dataset::fromCSV('datasets/housing/serve.csv', labelColumn: 0);
$predictions = $pipeline->predict($dataset);

Internals

• Pipeline::load() reconstructs transformer metadata and tensor weights.
• SafeTensorsIO::load() supplies native tensor buffers without PHP serialization.
• predict() applies the same preprocessing path used in training.

Performance notes

• Production inference should keep the input dataset in Tensor mode.
• Reuse loaded models across requests to avoid repeated disk I/O.
• Ensure all tensors are contiguous before inference for best throughput.