viable/strict/1787807055: Move expandable_segments() out of the header (#194872)
- PyTorch: 125 events in the last 90 days
- PyTorch: 124th Release in the last 90 days
- Previous: earlier the same day · trunk/a5f76a0e8ca67c5ef3c095061187dd7c318e7197: [cuDNN] Add guards for cuDNN SDPA decode (#194927)
What happened
Summary CUDAAllocatorConfig::expandable_segments() branches on PYTORCH_C10_DRIVER_API_SUPPORTED , which c10/cuda/CMakeLists.txt defines PRIVATE to c10_cuda . Being inline, the function recompiles in every translation unit that includes the header without that macro, so consumers see an unconditional false while the library itself sees the real answer. CUDACachingAllocatorReserveDeviceTest hits exactly this — it enab…
Summary assembled by rule from the sources below