trunk/74b55cafb21cf7d14bf8adc466e256ef2b994533: Make the int4mm sm_80 requirement an actual runtime check (#195697)
- PyTorch: 2455 events in the last 90 days
- PyTorch: 2422th Release in the last 90 days
- Previous: earlier the same day · trunk/64373c6e0d225f9809e0f9cb98741a77cf67603d
What happened
_weight_int4pack_mm_cuda / _convert_weight_to_int4pack_cuda gate on !defined(__CUDA_ARCH__) || (__CUDA_ARCH__ >= 800) , with a TORCH_CHECK(false, ...) in the other arm. Both are host functions, and __CUDA_ARCH__ is only defined during nvcc's device compilation pass -- so !defined(__CUDA_ARCH__) is always true when this is preprocessed for the host, and the check is unreachable in every build. The "requires sm_80" co…
Summary assembled by rule from the sources below