Releases

Install from PyPI or crates.io:

pip install libsdca==0.3.0
cargo add libsdca@0.3.0

The attached files are the published PyPI wheels and sdist and the crates.io crate.

Unified interfaces: Python package (fit, loss_and_grad, coordinate_update),
scikit-learn estimator (SDCAClassifier), PyTorch modules and functions, JAX loss
functions, Rust crate, C ABI (include/sdca/c_api.h), C++11 headers, and MATLAB/Octave
gateways sharing one set of objective, option and status names (API.md).
Numerical stability: Losses and gradients preserve tiny probability tails down to the
subnormal limit (avoiding softmax collapse), and distillation gradients are exact zero at
equal logits (numerics,
known limits).
Fused CUDA and CPU kernels: On an A100 (8192 × 32000 and 8192 × 128256),
cross-entropy takes 0.93–0.98× torch.compile time (0.33–0.41× eager), top-5
cross-entropy 0.23–0.26×, and distillation 0.62–0.81× (performance).
Certified convergence: status == "solved" requires the duality gap of the returned
weights and the feasible projection of the returned duals, evaluated in double with a
first-order rounding estimate, to meet
tol (API.md).

PyTorch (libsdca, libsdca.torch): CrossEntropyLoss (drop-in for
torch.nn.CrossEntropyLoss, except integer class labels only and no label smoothing), TopKCrossEntropyLoss, MulticlassHingeLoss,
TopKHingeLoss ($\alpha$ and $\beta$), MultilabelCrossEntropyLoss,
MultilabelHingeLoss, ProximalCrossEntropyLoss, DistillationLoss,
ProximalDistillationLoss, SDCALoss, and LinearCrossEntropyLoss (chunked linear head
- cross-entropy without materializing (N, V) logits) across float16, bfloat16,
  float32 and float64 on CPU and CUDA, with torch.compile(fullgraph=True) and
  DTensor support (TORCH.md).
JAX (libsdca.jax): Classification and distillation losses for jit, grad and
vmap, with zero-copy XLA FFI custom calls on CPU (JAX.md).
Python & scikit-learn: fit, loss_and_grad, distillation_loss_and_grad,
proximal_distillation_loss_and_grad, coordinate_update, and SDCAClassifier
(python/README.md).
Rust, C, C++ and MATLAB: Rust crate libsdca (rust/README.md),
versioned C ABI libsdca::sdca_c (docs/C.md), CMake target
libsdca::libsdca (docs/CPP.md), and MEX gateways matsdca_fit,
matsdca_loss and matsdca_prox (docs/MATLAB.md).

CUDA kernels: Fused Triton kernels for all losses with C++ cuLaunchKernel autograd
dispatch, multi-row CTA tiling for short rows, streaming tiles up to $2^{25}$ classes, and
in-place backward gradient writes into unreferenced logits tensors.
CPU kernels: JIT-compiled ATen/OpenMP C++ extension with C++ autograd nodes and SIMD
float32 exponentials + float64 row statistics for narrow cross-entropy and top-k
cross-entropy ($k \le 64$).
Shrinking and screening (shrinking="auto"): Combines LIBLINEAR-style active-set
shrinking with gap-safe screening for hinge fits, adapting the evaluation schedule to
shrunk epoch work (0.14–0.59× LinearSVC runtime; API.md).
$O(n \log n)$ projections and Taylor root warm-starts: Top-k cone, simplex, knapsack
and top-k entropy projections run in $O(n \log n)$, and proximal / entropy root solves
converge in 1.0–1.3 row evaluations on average.

Single rounding: 16-bit gradients are scaled by the upstream cotangent and reduction
factor in working precision before a single rounding to the output dtype, and hinge / top-k
target gradients are formed from the unrounded competitor sum.
Double iterate for float32 fits (precision="float64" default): Keeps internal
solver weights and duals in double (8(d + n)m bytes) and rounds once at each
evaluation so float32 fits reach float64 gap tolerances down to float32 output
rounding (3.5e-7 on digits; pass precision="float32" for pure float32 speed and
memory; numerics.md).
Compensated reductions & hardened projections: Compensated sums in multilabel
cross-entropy and distillation, warm-start residual preservation at large score offsets,
and checked float64 / extended-precision multilabel_hinge projections
(API.md).

Portable py3-none wheels for Linux x86_64 (manylinux_2_28) and macOS 11+ (arm64)
with built-in BLAS; configurable BLAS (SDCA_BLAS_BACKEND=AUTO|EXTERNAL|BUILTIN);
CMake 3.12+ package config; and GitHub Actions CI on Linux and macOS.

Navigation Menu