Deterministic machine learning. Without leaving PHP.
A portable, deterministic ML runtime for PHP. Your model lives in the web worker: no serving sidecar, no IPC hop per prediction.
Pre-release. No public build yet.
Bit-identical CPU == GPU by SHA-256 at full scale on 26 GPU models: 22 NVIDIA and 4 AMD Instinct.
f32, no Tensor Cores on the bit-identical path. Default-mode subnormal and NaN bit patterns can differ. Each NVIDIA driver is attested.
What has been attested
CPU == GPU
Forward and gradients, compared by SHA-256 of the output bytes.
Training replay
Bit-identical to the CPU on 19 NVIDIA models, on x86 and Arm hosts. Not run on AMD Instinct.
Limits
AMD Radeon and Intel GPUs: exact ops at small sizes only. Intel Gaudi is not supported.
Software Vulkan is never counted as evidence. Full scope and limits
Get a written report
GPU test run by us
We run the Talario test suite on your GPU model and send you a test record. From USD 487: a flat price by GPU class
Model reproducibility review
Planned: we run one version of your ONNX model and report whether its results reproduce. USD 5,000 per model version. Re-review USD 2,500.
Run your model in PHP
Export to ONNX, import it, and call it from your PHP code.
Read the ONNX guideuse Talario\Device;
use function Talario\Onnx\onnx_import;
$dev = new Device('auto');
$model = onnx_import(
$dev, 'model.onnx', [1, 784]
);
$logits = $model($features);