queue: sgl-kernel -> parked, ai-ml-sdk-vgf-library -> ci-running

sgl-kernel: unconditional CUDA (project(... LANGUAGES CXX CUDA), enable_language(CUDA),
find_package(CUDAToolkit REQUIRED), find_package(Torch REQUIRED)), 55 .cu sources, no
sdist published. Not portable to riscv64.

ai-ml-sdk-vgf-library: plain CMake/pybind11 C++ library, no Vulkan driver needed at build
time. PR #3001 opened.

Author: luhenry

.queue.yml

@@ -9940,16 +9940,16 @@ packages:
   version: 0.3.21
   home: https://github.com/sgl-project/sglang/tree/main/sgl-kernel
   repo: https://github.com/sgl-project/sglang/tree/main/sgl-kernel
-  status: queued
+  status: parked
   pr: null
-  notes: '2 Linux wheels upstream (abi: cp310); no riscv64 on PyPI or pypi.riseproject.dev.'
+  notes: 'Not feasible: unconditional CUDA, confirmed from the real CMakeLists.txt (current path python/sglang/kernels/aot/CMakeLists.txt, the old sgl-kernel/CMakeLists.txt at the 0.3.21 era). `project(sgl-kernel LANGUAGES CXX CUDA)`, `enable_language(CUDA)` and `find_package(CUDAToolkit REQUIRED)` run unconditionally -- no CPU/ROCm option, no `if()` gate around any of it. It also `find_package(Torch REQUIRED)` and `FetchContent`s NVIDIA cutlass, flashinfer-ai/flashinfer and sgl-project/sgl-attn (a flash-attention fork), all CUDA-only. 55 `.cu` files under csrc/. PyPI classifies it `Environment :: GPU :: NVIDIA CUDA`; no sdist is published (`pip download --no-binary :all:` finds nothing), so there is nothing to build from even ignoring the CUDA requirement. The aarch64 wheel is not a CPU fallback: the CMakeLists only uses `CMAKE_SYSTEM_PROCESSOR MATCHES aarch64` to flip default GPU gencode flags for SM90 (Grace Hopper), i.e. it targets NVIDIA Grace-Hopper (arm64 CPU + Hopper GPU) systems, not ARM CPUs generally. No worktree/branch/PR created.'
 - pkg: ai-ml-sdk-vgf-library
   version: 0.10.0
-  home: https://pypi.org/project/ai-ml-sdk-vgf-library/
-  repo: https://pypi.org/project/ai-ml-sdk-vgf-library/
-  status: queued
-  pr: null
-  notes: '2 Linux wheels upstream (abi: cp312); no riscv64 on PyPI or pypi.riseproject.dev.'
+  home: https://github.com/arm/ai-ml-sdk-vgf-library
+  repo: https://github.com/arm/ai-ml-sdk-vgf-library
+  status: ci-running
+  pr: https://github.com/riseproject-dev/python-wheels/pull/3001
+  notes: 'Open source (Apache-2.0), Arm ML SDK for Vulkan VGF (Vulkan Graph Format) encoder/decoder library. Build is plain CMake + pybind11 C++17, confirmed from the real v0.10.0 source: no Vulkan headers/loader/driver needed at build or test time at all (VGF is a flatbuffers-based file format, independent of the Vulkan runtime). Needs 4 small build-time deps with no FetchContent fallback in this repo (argparse, nlohmann/json, flatbuffers, pybind11, pinned by arm/ai-ml-sdk-manifest''s vgf-lib group) and a one-line patch to pip_package/setup.py''s BDistWheel.get_tag(), which hard-asserts machine == "x86_64" on non-aarch64 Linux. No sdist on PyPI and no upstream CI workflow to mirror, so the workflow drives scripts/build.py directly inside manylinux_2_39_riscv64. PR #3001 opened as draft; not locally validated (sandboxed dev environment refuses to compile/run third-party native code), so this is the first real CI run.'
 - pkg: fastdigest
   version: 0.12.0
   home: https://github.com/moritzmucha/fastdigest