mirror of
https://github.com/kvcache-ai/sglang.git
synced 2026-07-02 21:37:11 +00:00
Upstream sglang removed the gptq_marlin* kernels to the JIT system. - Added to JIT in sglang upstream PR #18543 - Removed from sgl-kernel in sglang upstream PR #19241 At that moment the kt_ep_wrapper should have been updated but was not. In many cases the trouble is not observed since sgl-kernel is often installed from (stale) prebuilt wheels. When building from source however, e.g. for CUDA123/sm120, the produced binary does not have gptq_marlin_repack and kt_ep_wrapper is not imported.