How can I resolve this build issue with libtorch

I’ve run USE_CUDA=ON BUILD_SHARED_LIBS=OFF python build, and also tried make local, but when it gets to building Caffe2, I get:

[1/625] Building CUDA object caffe2/CMakeFiles/torch_cuda.dir/__/aten/src/ATen/native/transformers/cuda/flash_attn/kernels/
FAILED: caffe2/CMakeFiles/torch_cuda.dir/__/aten/src/ATen/native/transformers/cuda/flash_attn/kernels/ 
/usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler -DAT_PER_OPERATOR_HEADERS -DCPUINFO_SUPPORTED_PLATFORM=1 ...

The command prints out “nvcc fatal : Could not open output file caffe2/CMakeFiles/torch_cuda.dir/__/aten/src/ATen/native/transformers/cuda/flash_attn/kernels/”, and in fact caffe2/CMakeFiles doesn’t exist at all.

Came across the same error, found a fix here: Build consumes 32GB RAM triggers OOM crash when compiling CUDA object · Issue #111526 · pytorch/pytorch · GitHub