summarylogtreecommitdiffstats
path: root/PKGBUILD
AgeCommit message (Collapse)Author
2026-04-27Cap nvcc parallelism via MAKEFLAGSWill Handley
2026-04-27Bump to sglang 0.5.10.post1Will Handley
2026-04-13Update to 0.4.1 (sglang 0.5.10)Will Handley
2026-03-19Patch THCAtomics for torch 2.10, add build flagsWill Handley
- Remove THC/THCAtomics.cuh include (removed in torch 2.10) - Set CMAKE_POLICY_VERSION_MINIMUM for newer cmake - Default SGL_KERNEL_COMPILE_THREADS=1 to prevent OOM - Skip build dependency check for torch version mismatch Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
2026-03-18Use python-pytorch instead of python-pytorch-cudaWill Handley
Avoids conflict with python-pytorch-opt-cuda which also provides python-pytorch. CUDA toolkit dependency is separate. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
2026-03-18Add torch and sympy to makedependsWill Handley
Both needed at build time by scikit-build-core. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
2026-03-18Fix source URL to use sglang release tagWill Handley
sgl-kernel 0.3.21 ships inside sglang v0.5.9 tarball. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
2026-03-18Build from source against system torchWill Handley
Pre-built wheel was compiled against torch 2.9.1, causing ABI mismatch with Arch's torch 2.10.0. Building from source links against the system torch and targets the local GPU architecture. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
2026-03-18Initial PKGBUILD for sgl-kernel 0.3.21Will Handley
Pre-built CUDA kernels for SGLang. Uses abi3 wheel (Python 3.10+). Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>