The codec decoder for ExecuTorch's Vulkan backend: forward(codes int64[8, T]) -> audio float32[T * 960], T 4-1024, with get_model_schema like the other decoders; fp16 weights and compute, fp32 accumulation. It needs a Vulkan runtime built from software-mansion-labs/executorch PR #32 (the et_vk.conv1d_fused op and buffer-backed conv1d kernels): react-native-executorch's native libs 0.11.1 cannot load it. On a Galaxy S24 FE it decodes 10 s of audio in 143 ms (XNNPACK fp16: 633 ms), 11-12 dB closer to the fp32 decoder.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Add the XNNPACK decoder: fp16, its last block and output stage in fp32, with get_model_schema