forked from Karylab-cklius/vllm
Use CUDA 12.9.86 compiler binaries (nvcc, ptxas, cicc) for better GPU code generation while keeping CUDA 12.8 headers/libs for runtime compatibility with PyTorch cu128 wheels. Controlled by ARG NVCC_TOOLCHAIN_VERSION (default 12.9.86). Set to empty string to disable. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> (cherry picked from commit fa2f0455aa6d4ad69c510eeb2516f026b26f795d)