Native Support or Extension Layer for Running AI/ML Models in PHP

[gavriel-adi]

As AI and machine learning continue transforming how applications interact with users, many PHP developers are starting to integrate AI-powered features — from image analysis and text generation to recommendation engines and embeddings. However, PHP currently lacks a robust, standardized way to run or interact with large AI models directly within the language.

Proposal: Introduce a native extension or FFI interface to enable safe and efficient execution of AI/ML models — either via integration with ONNX Runtime, TensorFlow Lite, or even through PHP bindings for stable diffusion and LLMs (e.g., ggml, llama.cpp).

This could include:

php_ai_run_model('model.onnx', $inputData)

Memory-efficient tensor structures

Streamed inference results

Optional GPU acceleration (via Vulkan, CUDA, or Metal)

Benefits:

Allows developers to build AI-enhanced PHP apps without relying on external APIs or services

Reduces latency, cost, and privacy concerns associated with cloud AI

Empowers CMSs, e-commerce platforms, chat systems, and media platforms with embedded intelligence

Aligns with the direction of true_async and high-performance PHP runtimes

Today, many developers rely on exec() to call Python or external binaries, which is slow, unsafe, and limits scaling. A native solution would significantly expand what PHP is capable of.

[pronskiy]

This is a really cool idea!

[pronskiy]

@gavriel-adi what do you think about https://github.com/ankane/onnxruntime-php ?

[gavriel-adi]

Thanks for the link! I haven’t tested the ankane/onnxruntime-php library yet, but it definitely looks promising — especially because it has GPU support, which is something the library I did try (codewithkyrian/transformers-php) doesn’t offer.

I found that transformers-php had limited performance, weak output quality, and poor compatibility with more advanced models. Also, it lacked flexibility and proper session management.

I’m hoping onnxruntime-php will be a more capable and scalable option. Looking forward to testing it.