LoginSign Up
← Back to all sparks
G

glmbayes

ANALYTICS
Velocity6.3

R package glmbayes by knygren — release notes from GitHub.

github.com/knygren/glmbayes ↗

A GPU-accelerated Bayesian GLM package buys its way into the standard R Bayesian toolchain

bayesian-statisticsgeneralized-linear-modelsopenclr-packagegpu-acceleration
◆Current state
glmbayes fits Bayesian generalized linear models with optional OpenCL acceleration. The last four months moved it from a package with its own vocabulary to one that answers the insight and bayestestR generics the rest of the R Bayesian ecosystem is built on, while pushing the OpenCL kernels out into a separate nmathopencl dependency that carries CRAN Windows binaries. It returned to CRAN in August after an archival over a configure policy issue.
◆Where it's heading
The arc is about removing reasons not to use it. GPU support was previously blocked on Windows because the OpenCL kernels were vendored; splitting them into a CRAN package with binaries fixed that. The ecosystem work does the same thing for tooling — a glmb fit now responds to get_parameters, get_priors, simulate_prior and check_prior, so it drops into workflows built around easystats rather than requiring its own. The CRAN archival and the configure fixes that followed show how much of the effort goes into distribution rather than modelling.
◆Prediction
get_priors() returning the full prior specification rather than a marginal table is the kind of detail that invites further bayestestR integration, and the diagnostic surface is the least built-out part of what has shipped so far.

◆Recent moves

  1. 10d ago

    Back on CRAN after a configure policy fix

    A CRAN return after archival over a configure policy issue. OpenCL now compiles only when a non-PoCL platform exposes a GPU, so CRAN's Debian checks stop writing PoCL cache files. The notes state there are no API or modelling changes.

    View source ↗
  2. 22d ago

    OpenCL split out to nmathopencl; insight and bayestestR integration

    ⚡ SPARK

    OpenCL kernels move out to a hard dependency on nmathopencl, which ships CRAN Windows binaries and unblocks GPU builds there. S3 methods arrive for the insight accessors and for bayestestR's simulate_prior(), check_prior() and describe_prior(), with both packages promoted from Suggests to Imports. get_priors() returns the full prior specification including the complete Sigma rather than a marginal table.

    View source ↗
  3. 1mo ago

    Multi-response models and conjugate GLM priors

    Multi-response lmb() and mlmb arrive, along with conjugate GLM priors covering dBeta and dGamma identity-link paths and a vignette restructure with Bayes Rules! and LearnBayes companion appendices. A fix stops rindepNormalGamma_reg() accepting prior-dominated dispersion envelopes.

    View source ↗
  4. 3mo ago

    OpenCL kernels restructured and a binomial GPU bug fixed

    An OpenCL-focused release: kernels and fragments are consolidated, the GPU program is assembled from explicit pieces rather than ad hoc concatenation, and a binomial kernel bug is fixed so aggregated binomial data matches CPU semantics. OpenCL tests gain skip_on_cran() so CRAN machines do not run heavy GPU work.

    View source ↗
  5. 3mo ago

    Version bump for CRAN resubmission

    A version bump in response to CRAN resubmission feedback, with no described change.

    View source ↗
  6. 1y ago

    CRAN-ready beta with the core S3 interface

    The CRAN-ready beta, establishing the S3 interface — glmb(), summary(), predict(), confint() — and vignettes comparing results against glm().

    View source ↗