nemo_automodel.components.models.kimi_k3.kda_fused.chunk_kda_fwd_cuda
nemo_automodel.components.models.kimi_k3.kda_fused.chunk_kda_fwd_cuda
Source text of the fused chunked Kimi-K3 KDA forward (CUDA) and its torch binding.
kda_fused._forward_ext hands these strings to torch.utils.cpp_extension.load_inline as
cpp_sources = [HEADER, BINDING] and cuda_sources = [HEADER, KERNEL]; torch writes them to main.cpp and
cuda.cu in the build directory and builds the extension once, so the wheel ships Python files only. The header is
inlined into both translation units (there is no file to #include), which is why it carries no #pragma once.