mirror of
https://github.com/hanzoai/ml.git
synced 2026-11-19 23:50:51 +00:00
cuda_include_dir skipped CUDA_HOME, the variable every other build script in the stack reads first, so a host that selects a toolkit that way compiled its headers with whatever /usr/local/cuda points at. flash-attn-v3 took NVCC or a bare nvcc from PATH while reading CUDA_HOME for the lib dir, so the compiler and the libraries it linked could come from different toolkits. Both now resolve NVCC, then CUDA_HOME/CUDA_PATH/CUDA_ROOT, then PATH, and the compiler that wins decides the libraries. Co-authored-by: Hanzo Dev <[email protected]>