🐛 Describe the bug It came as a surprise to me that setting PYTORCH_NO_CUDA_MEMORY_CACHING=0 is treated the same as PYTORCH_NO_CUDA_MEMORY_CACHING=1 in the native allocator code. The python code of torch handles this envvar by checking i...
c10::DeviceIndex device = 0; C10_CUDA_CHECK(c10::cuda::GetDevice(&device)); malloc(&r, device, nbytes, stream); } malloc(&r, device, nbytes, stream); return r; }@@ -3625,11 +3598,7 @@ class NativeCachingAllocator : public CUDAAllocator { ...