Prebuilt sgl-kernel for AMD Radeon RDNA3 (gfx1100)

SGLang's HIP path imports sgl_kernel unconditionally, and the sglang-kernel wheel on PyPI is CUDA-only -- it looks for libnvrtc.so.13. Building for RDNA3 from source means fetching ~125 MB of SGLang sources and running a compiler, which is a lot to ask, and impossible on links that cannot hold a large transfer open.

This wheel is that kernel, already compiled for gfx1100, with the patches from https://github.com/AMD-AIM/sglang-radeon applied:

• the architecture allowlist widened past gfx942/gfx950/gfx1250
• RDNA3 given the 48 KB LDS budget it actually has, not the 128 KB one
• the GPTQ kernel compiled and its ops registered for ROCm, which upstream never does -- this is what lets a 27B model run 4-bit on a single 48 GB card

It installs the same sgl_kernel module as upstream, so it is a drop-in replacement. Uninstall sglang-kernel first if it is present.

Requires: ROCm 7.x, PyTorch with HIP, Python 3.12, an RDNA3 GPU. Verified on Radeon PRO W7900.

Use with the plugin, which supplies the runtime shims::

pip install --no-deps sglang-radeon-rdna3 sglang-kernel-rdna3
sglang-rdna3 doctor