numr 0.7.0

Tensors, linear algebra, FFT, and autograd behind one API on CPU, CUDA, and WebGPU.
Every kernel is written in-house — no cuBLAS, cuSOLVER, or MKL.

Added

• Tensors — Tensor<R> generic over the backend. Broadcasting, zero-copy views (reshape, transpose, slice, permute), fallible constructors that never panic on allocation failure.
• Backends — CPU (#[target_feature] AVX-512 / AVX2+FMA / NEON, optional Rayon), CUDA (graph capture, arena allocator, kernel cache), WebGPU (WGSL per dtype). CPU and CUDA cover every dtype; WebGPU is F32 / I32 / U32 / Bool.
• DTypes — F64, F32, F16, BF16, FP8E4M3, FP8E5M2, I64, I32, I16, I8, U64, U32, U16, U8, Bool. Narrow dtypes accumulate in a wider type. WebGPU is F32 / I32 / U32 / Bool.
• Element-wise and reductions — unary, binary, scalar, compare, logical, conditional. Reductions over any axis set, cumulative ops, sorting with a NaN-aware total order, gather/scatter indexing.
• Matmul — batched and broadcast, tiled CPU kernel with a transposed-B path, fused GEMM epilogue, FP8, min-plus/max-plus semirings, einsum, 2:4 structured sparsity.
• Linear algebra — LU, QR, SVD, Cholesky, Schur, QZ, polar, eigen. Solvers, lstsq, inverse, slogdet, cond, matrix_rank. Matrix functions (expm, logm, sqrtm, signm, funm). Tucker, HOSVD, CP, tensor-train.
• Iterative solvers — CG, BiCGSTAB, CGS, GMRES, LGMRES, MINRES, QMR, Jacobi, SOR, sparse eigensolvers, svds, AMG.
• FFT — 1D / 2D / ND, forward and inverse. Bluestein covers arbitrary sizes on all three backends.
• Convolution — conv1d, conv2d, depthwise_conv2d, conv_transpose1d with native kernels and autograd on each.
• Autograd — reverse mode via Var<R> and GradFn, forward mode via DualTensor, gradient checkpointing. backward takes a needed-gradient mask so ops skip gradients the driver discards.
• Random — uniform, normal, seeded normal, advanced and multivariate distributions, Sobol sequences, Philox on GPU. Seeds reproduce per backend, not across backends.
• Statistics and special functions — quantiles, histograms, correlation, distance metrics, polynomials. Gamma, Bessel, Airy, Fresnel, elliptic, hypergeometric, orthogonal families.
• Sparse (sparse) — CSR, CSC, COO with conversions, element-wise ops, SpMM, SpMV, and sparse LU / QR with COLAMD ordering.
• Distributed (distributed, nccl) — a Communicator trait over NCCL, nexar, hierarchical, and no-op backends, plus process groups.

Semantics

Behaviour a caller must know. These are contracts, not defaults.

• Output dtype is a function of input dtypes, never input values. pow_scalar on an integer tensor returns F64 unless the exponent is a whole non-negative number. matmul_output_dtype is public so a caller can size a bias against it.
• Integer accumulators saturate at the dtype's bound: sum, prod, mean, cumsum, cumprod, matmul, scatter_reduce.
• Integer element-wise ops wrap: add, sub, mul, and the fused forms.
• Integer division by zero returns 0, and INT_MIN / -1 returns INT_MIN. Neither panics.
• I8 matmul widens to I32. matmul_bias on I8 takes an I32 bias and returns I32. The GEMM epilogue ops (matmul_bias_activation, matmul_bias_residual, and the backward form) reject I8 rather than read a wider bias buffer as I8.

Notes

• tests/backend_parity/ checks CUDA and WebGPU against CPU for every operation, at dtype-appropriate tolerances.
• WebGPU stays 32-bit by design. WGSL has no native F64.
• ROCm is planned.

Full diff: v0.6.1...v0.7.0