Undefined symbol _ZN6caffe26detail37_typeMetaDataInstance_preallocated_32E

Hi,
I'm getting an undefined symbol _ZN6caffe26detail37_typeMetaDataInstance_preallocated_32E when running amp.initialize(...).

$ ipython3

In [1]: import torch, torchvision 
     ...: from apex import amp 
     ...: model = torch.nn.Linear(1024, 512).cuda() 
     ...: optimiser = torch.optim.SGD(model.parameters(), lr=1e-3) 
     ...: model, optimiser = amp.initialize(model, optimiser, opt_level='O2')
Selected optimization level O2:  FP16 training with FP32 batchnorm and FP32 master weights.

Defaults for this optimization level are:
enabled                : True
opt_level              : O2
cast_model_type        : torch.float16
patch_torch_functions  : False
keep_batchnorm_fp32    : True
master_weights         : True
loss_scale             : dynamic
Processing user overrides (additional kwargs that are not None)...
After processing overrides, optimization options are:
enabled                : True
opt_level              : O2
cast_model_type        : torch.float16
patch_torch_functions  : False
keep_batchnorm_fp32    : True
master_weights         : True
loss_scale             : dynamic
Warning:  multi_tensor_applier fused unscale kernel is unavailable, possibly because apex was installed without --cuda_ext --cpp_ext. Using Python fallback.  Original ImportError was: ImportError('/home/brett/.local/lib/python3.7/site-packages/amp_C.cpython-37m-x86_64-linux-gnu.so: undefined symbol: _ZN6caffe26detail37_typeMetaDataInstance_preallocated_32E')

Apex is built on an up-to-date Ubuntu 19.10 installation with CUDA 10.1, gcc 9.2.1 and nvcc 10.1.243 using gcc 8.3.0 as the ccbin (see https://github.com/NVIDIA/apex/issues/278#issuecomment-581789063).

Due to pip's habit of removing temporary build directories, I haven't been able to figure out where this symbol originates or what library should contain the symbol (libcaffe.so seemed likely, but I couldn't find the symbol there).

Start with the amp.initialize(...) path and the amp_C extension mentioned in the warning. Inspect how Apex is built with --cuda_ext --cpp_ext and identify which linked library should provide _ZN6caffe26detail37_typeMetaDataInstance_preallocated_32E. Done means the extension imports without the undefined-symbol error and the fused unscale kernel is available.