PHPAiModel-Transformer

PHPAiModel-Transformer is a lightweight, pure PHP implementation of a character-level Transformer model runtime. It includes a simple chat UI for interacting with models, supporting inference on JSON-encoded weights. Designed for PHP 7.4+, it features pre-LayerNorm, GELU activation, Multi-Head Attention (MHA), and a tied output head for efficient generation.

This project enables running Transformer-based language models directly in PHP without external dependencies, ideal for educational purposes, prototyping, or lightweight deployments.

Overview

PHPAiModel-Transformer provides a backend runtime (aicore.php) for Transformer inference and a frontend chat interface (index.php). The runtime loads model weights from JSON files, performs forward passes for logit computation, and generates text using temperature and top-k sampling. The chat UI allows selecting models, setting parameters, and interacting via prompts.

Key components:

• Character-level Tokenization: Simple mapping of characters to IDs.
• Causal Self-Attention: Ensures autoregressive generation.
• Feed-Forward Layers: With GELU activation and configurable dimensions.
• Generation: Supports context management up to max_seq length.

Features

• Pure PHP implementation (no extensions required beyond standard ones like mbstring and json).
• Supports pre-LayerNorm architecture with Multi-Head Attention.
• Tied embeddings for input and output (efficient memory use).
• Configurable sampling: temperature and top-k.
• Simple web-based chat UI with model selection and parameter controls.
• Debug metadata in responses (e.g., model dimensions).
• MIT licensed, open-source.

Requirements

• PHP 7.4 or higher (tested for compatibility).
• Web server (e.g., Apache, Nginx) for serving the UI.
• Recommended: Increase memory_limit to 2048M or more for larger models.
• Recommended: Set max_execution_time to 600 seconds for long generations.

License

This project is licensed under the MIT License. See the LICENSE file for details.