Autograd and the tape

Talario is define-by-run. You build a forward computation in plain PHP and Talario records it onto a tape: a bump-allocated arena of POD nodes in C, where each tensor is an integer id rather than a PHP object. There is no per-op Zval and no operator-object churn on the hot path.

The shape of a step#

use Talario\Device;
use Talario\Tape;

$dev = new Device('gpu');     // 'cpu', 'gpu', or 'auto'
$t   = new Tape($dev);

$x = $t->tensor($features, [$n, $in]);   // host -> device tensor id
$w = $t->tensor($weights,  [$in, $out], true);  // requiresGrad = true
$y = $t->matmul($x, $w);
$y = $t->relu($y);
$loss = $t->mseLoss($y, $t->tensor($targets, [$n, $out]));

$t->backward($loss);          // reverse-mode pass
$grad = $t->grad($w);         // read back the gradient

backward() walks the recorded tape in reverse, accumulating gradients on every tensor that was created with requiresGrad = true (and on any bound Parameter). Forward is lazy and the backward pass is batched, so a training step resolves to a handful of backend submissions rather than one per op.

Reading values back#

value($id) and grad($id) copy a tensor or its gradient back to a PHP array. These are device-to-host transfers, so call them when you need the numbers (for logging or a metric), not inside the inner loop.

Device memory#

Device buffers are owned by the Zend object handlers of the Tape and Parameter, so they are released deterministically the moment PHP drops the last reference. The tape arena itself is freed in one shot when the tape is dropped.

Edit this page on GitHub