About

Mission

Talario's mission is to make PHP a first-class language for deep learning. We provide a hardware-agnostic, zero-allocation neural-network runtime as a native PHP extension: a PyTorch-style autograd tape recorded in a C arena, lowered to vendor-neutral SPIR-V and executed on Vulkan Compute across NVIDIA, AMD, Apple, Intel and ARM, with a portable CPU/SIMD fallback. We build in the open, credit the work we stand on, and work with the wider PHP community so that the frameworks and applications PHP developers already use can train and serve models without leaving PHP.

Talario is a native PHP 8.x Zend extension. It records a PyTorch-style autograd tape in a C arena, lowers it to vendor-neutral SPIR-V, and executes it on Vulkan Compute across NVIDIA, AMD, Apple, Intel and ARM, with a portable CPU/SIMD fallback. Its BLAS-shaped kernels (GEMM/AXPY) delegate to the Tessero numerical core, while the element-wise, activation and autograd kernels are Talario's own SPIR-V. A Laravel bridge is on the roadmap.

The story

Every PHP developer who has needed machine learning knows the moment. You have built the app, with its models, queues and dashboards, and then a requirement lands that needs a trained network: a classifier, a policy, an embedding. PHP, the language running a huge share of the web, quietly hands you a bill: go stand up a second stack.

So you shell out to a Python service, ship features over a socket into PyTorch, run the model, and reverse the whole trip. Two runtimes, two dependency managers, two things to deploy and keep in sync. Worse, the GPU story is vendor-locked: CUDA first, everything else second. Talario exists to remove both taxes: train and serve models natively in PHP, on whatever silicon you have.

Where the name comes from

Talario evokes talaria, the winged sandals of Mercury, a nod to speed and motion. A define-by-run autograd engine is exactly that, the motion of data through a graph, recorded as it happens and run in reverse to learn, for a runtime whose whole job is to keep tensors moving across the fastest path to the metal.

Write once, run on any silicon

The portability model is the Java Virtual Machine, and the capability north-star is PyTorch. One vendor-neutral SPIR-V artifact runs on every major GPU through Vulkan, and the CPU fallback is designed to be gradient-identical, so you can develop on a laptop and deploy on a GPU server with no code change. What the library actually implements, op by op and shader by shader, is published in the capabilities census, generated from the source so it never drifts.

In detail: Talario exposes a define-by-run tape with reverse-mode autograd over a fixed, growing op set (matmul, conv2d, pooling, the common activations, bias, fused linear+ReLU, and the MSE and softmax-cross-entropy loss terminals), on the Vulkan and CPU backends. Reproducible benchmarks are not published yet; see the capabilities page for what the source implements.

How we work

• In the open. Development, issues and decisions happen in public repositories; see governance.
• Credit the work we stand on. Talario builds on the Tessero numerical core and the Vulkan ecosystem; see the Acknowledgments.
• With the PHP community. We aim to make the frameworks and applications PHP developers already use capable of training and serving models, a bolt-on to the wider PHP ecosystem, not a replacement for it.

Licensing

Talario is released under the Apache License 2.0, © 2026 the Talario contributors, as recorded in the library repository. See the license and third-party notices.

Independence

Talario is an independent project and is not affiliated with or endorsed by PyTorch, the Linux Foundation or Meta. Talario builds on and credits the Tessero numerical core (tessero.org).