Skip to content

feature/gpupreparation - #35

Merged
JoshuaTetzner merged 3 commits into
mainfrom
feature/gpupreparation
Oct 14, 2026
Merged

JoshuaTetzner merged 3 commits into
mainfrom
feature/gpupreparation

Conversation

@JoshuaTetzner

Copy link
Copy Markdown
Owner

No description provided.

…atrices

- Fully pivoted ACA (CPU and batched CUDA kernel) gets a `sampled` mode that
  rescales the residual of representative column samples, used by NCA's
  batched bottom-up compressor
- localkernelmatrix(matrix, rows, nothing) / (matrix, nothing, columns)
  prepares only the sampled side for BEAST kernel matrices
- tofarquadstrat for BEAST.DoubleNumSauterQstrat
- ACABEASTCUDA throws a clear error when the loaded BEAST has no
  BEASTCUDAExt; GPUMatrixData is documented as experimental
- Typed block comprehensions in the symmetric assemblenears: Julia 1.10
  infers Vector{Any} for an empty block list, which assemble_blocks rejects
- Compat entries for the stdlibs and all test-only packages
- Re-enable Aqua's compat check
@JoshuaTetzner JoshuaTetzner changed the title Feature/gpupreparation feature/gpupreparation Oct 14, 2026
@JoshuaTetzner
JoshuaTetzner merged commit b5a46c1 into main Oct 14, 2026
5 checks passed
@JoshuaTetzner
JoshuaTetzner deleted the feature/gpupreparation branch October 14, 2026 13:57
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant