Tensor operation kernels and backend dispatch for the Coeus stack.
Operation families
- Elementwise — [
unary] and [binary] kernels dispatched via [ElementwiseOps::elementwise_unary] / [ElementwiseOps::elementwise_binary]. - Linear algebra — [
matmul()], [sparse] SpMM/SpMV. - Reductions — [
reduction]:sum,mean,max/min,argmax/argmin,topk, cumulative sum/product scans, plus vector arithmeticdot(flat inner product) andcross(per-channel 3-vector cross alongdim), plus matrix normsfrobenius_norm/frobenius_norm_batched(compose onnormfortorch.linalg.matrix_norm(A, ord='fro')). - Convolution — 1-D/2-D/3-D forward+backward routed through
BackendOps::conv1d/conv2d/conv3d. - Pooling — max and average pooling (2-D/3-D) with backward gradients.
- Attention — [
attention::scaled_dot_product_attention] with causal/padding mask support. - Fused expressions — [
fuse] lazy expression DAG evaluated in a single pass on CPU. - Optimizer steps — provider-dispatched
sgd_step,adam_step,adamw_step,rmsprop_step, andadagrad_stepon [OptimizerOps].