cb.fabric-parity-gate

Fabric parity gate validates CUDA backend implementations against the PyTorch Fabric reference by enforcing equality across forward outputs, materialized final state, input gradients, and full parameter gradients.

The best way to test Fabric backend parity is applying a gate across active routes, touched owners, transition buckets, chunking policies, resets, and materialization modes to validate outputs, states, and gradients.

To enforce gradient equality, the parity gate checks input gradients, provided initial carry gradients, and full parameter gradients against the PyTorch Fabric reference across multiple backend paths.

Yes, you can build and validate a minimal parity matrix covering active routes, reset behavior, and materialization modes to certify readiness of fabric backend paths before production rollout.

Fabric parity checking works across materialization modes, chunking policies, demotions, and surface keys to ensure the active CUDA backend path matches the reference implementation comprehensively.

Run the Fabric parity gate at startup to verify outputs, states, and gradients across the active backend path before enabling performance measurements or closing performance validation.