OOM for first-time build with CUDA Toolkit 12.4

[kevinddchen]

I have noticed that when compiling CUDA extensions for gsplat, e.g. running for first time, my machine freezes and becomes unresponsive at this point:

gsplat/gsplat/cuda/_backend.py

Line 130 in dd66cbd

_C = load_extension(

. I gather that this is an OOM issue because it is resolved by setting MAX_JOBS=4 when running.

I am able to consistently reproduce this issue on my machine. I am using CUDA Toolkit 12.4 and 32 GB of CPU RAM. The GPU is A10G and the TORCH_CUDA_ARCH_LIST env var is not set. This issue does not occur with CUDA Toolkit 11.8 (all other specs unchanged).

Not suggesting any change at the moment, but more curious of others have encountered this issue as well.