DeepGEMM built from deepseek-ai/DeepGEMM@8b1392b (the DeepGEMM commit vLLM 0.28.1rc1.dev359 vendors) against torch 2.11.0+cu130, nvcc 13.0.88, GCC 11.4, cp312, x86_64 (glibc >= 2.35). For SkyRL's Megatron/FSDP extras, where vLLM's vendored copy (linked against torch 2.13) fails to import. Build: DG_FORCE_BUILD=1 CUDA_HOME=/usr/local/cuda-13.0 python setup.py bdist_wheel. Kernels are JIT-compiled at runtime (needs nvcc/nvrtc on the host).