burn_gdn

burn_gdn contains the Burn-native Gated DeltaNet 2 memory block used by JEPA autocode experiments.

The crate keeps the model/reference recurrence separate from the chunk-WY kernel boundary:

• GatedDeltaNet2Memory owns projections, gates, learned decay, output projection, and recurrent state shape.
• GatedDeltaNet2Executor::Reference uses the exact graph recurrence on any Burn backend.
• GatedDeltaNet2Executor::ChunkWy uses the CubeCL chunk-WY forward/custom backward path on supported f32 CUDA/WGPU backends and otherwise falls back to the reference path.
• gdn2_kernel_path::<B>() and gdn2_kernel_status::<B>() report what the selected executor will use for a backend.

Runtime controls are crate-local and keep older aliases for compatibility:

• BURN_GDN2_CHUNK_WY_CUSTOM_BACKWARD=0 disables custom backward and keeps only the forward kernel before reference/autodiff fallback.
• BURN_GDN2_CUDA_TENSOR_CORE_BACKWARD=0 disables the default CUDA tensor-core blocked analytic backward and uses the fused CubeCL runtime backward policy.
• BURN_GDN2_CUDA_SERIAL_BACKWARD=1 and BURN_GDN2_CUDA_COLUMN_PARTS_BACKWARD=1 select diagnostic CUDA backward variants.
• BURN_GDN2_CUDA_BLOCK_PARTS_BACKWARD=1 enables the block-parts CUDA diagnostic fallback when tensor-core backward is disabled.

The same names prefixed with BURN_JEPA_GDN2_ and the older BURN_DRAGON_GDN2_ aliases are also accepted.