cusna (native) — GPU-accelerated SAOM/RSiena, ERGM, and friends for R

cusna is a self-contained native engine — a C ABI over CUDA kernels and C++ host logic (libcusna) — callable from R without a Python runtime. It is the native counterpart of the reticulate-based cusna R wrapper, modelled on the R torch package:

the CRAN build is CPU-only from source (pure C++, needs only a C++17 compiler);
the GPU path is compiled from source when a CUDA 12.x toolkit (nvcc) is detected at configure time.

cusna_has_cuda() reports which build you have.

What the package provides

Family	Functions	Validated against
SAOM (RSiena)	`saom_data()`, `cusna_effect()`, `mom_estimate()`, `cusna_fit` methods; behavior co-evolution, composition change, `mom_estimate_multinet()`, `cusna_fran()`	data/masks/targets bit-identical to the reference; estimates within simulation SE; RSiena targets to machine zero
ERGM	`ergm_simulate()` (TNT sampler), `ergm_stats()`, `ergm_mple()`, `ergm_mcmle()`	sampler ≡ `ergm::simulate`; MLE matches `ergm::ergm()`
Temporal ERGM	`tergm_mple()` (+ block bootstrap), `tergm_simulate()`, `stergm_cmle()`	matches `btergm` to machine precision; `tergm` CMLE within SE
ALAAM	`alaam_mple()`, `alaam_mcmle()`, `alaam_simulate()`	MPLE ≡ `glm`; MLE recovers observed moments
Low-level	`cusna_network_stats()`, `cusna_behavior_stats()`, `cusna_gof_distribution()`	RSiena Appendix B conventions, machine zero

The underlying C ABI is bit-for-bit validated in native/test (see native/VALIDATION.md).

Reviewer quickstart (CPU-only, no GPU needed)

# from a checkout of the monorepo (configure vendors ../../native sources):
install.packages("cpp11")            # build-time only
# then, with a C++17 toolchain (Rtools on Windows):
#   R CMD INSTALL Rpkg-native
library(cusna)
cusna_has_cuda()                     # FALSE on the CPU-only build

# a two-wave panel and a Method-of-Moments SAOM fit, all native:
set.seed(7)
w1 <- matrix(as.integer(runif(400) < 0.12), 20, 20); diag(w1) <- 0L
w2 <- w1; flip <- sample(400, 40); w2[flip] <- 1L - w2[flip]; diag(w2) <- 0L
fit <- mom_estimate(saom_data(list(w1, w2)),
                    effects = list(cusna_effect("density"), cusna_effect("recip")))
summary(fit)

# an ERGM maximum-likelihood fit on the same data:
ergm_mcmle(w1, list(ergm_term("edges"), ergm_term("mutual")), directed = TRUE)

See vignette("cusna") for the full tour (covariates, co-evolution, multi-network models, TERGM/STERGM/ALAAM) and vignette("siena07-backend") for driving RSiena’s siena07() on the native simulator.

Building

See BUILD.md. In short: the CPU-only build needs a C++17 compiler (Rtools on Windows); the GPU build additionally needs a CUDA 12.x toolkit. The native sources are vendored from ../../native/ by configure.

License

MIT (see LICENSE). We compare outputs against RSiena but do not link its code.