Extends workflows to build an release cuda12/linux based images - #52
Conversation
|
The CUDA 12 addition needs one matrix correction before merge: the cuda12 arm64 entry still sets cuda_archs to 90;121-real. Compute capability 12.1 / GB10 is a CUDA 13 target, so a CUDA 12.9 toolchain should not be asked to compile 121-real. Please either omit the CUDA 12 arm64 variant or restrict it to architectures supported by that toolkit. @mudler this also has no PR CI because pull requests currently run only the CPU matrix. |
localai-org-maint-bot
left a comment
localai-org-maint-bot
left a comment
There was a problem hiding this comment.
@mudler The requested matrix correction is resolved at 9c767f9: the CUDA 12.9 container arms now explicitly target 70-real/90-real, while the release path uses the pinned ggml toolkit-gated defaults. The cuda12 variants are consistently included in build, merge, release bundle, rpath, and shared-library packaging branches. Good to merge from my review; this repository still has no PR workflow execution for these GPU matrix arms.
3 participants
I can't fully test, because the workflows belong to your repo.