search

pytorch / pytorch

Tensors and Dynamic neural networks in Python with strong GPU acceleration
https://pytorch.org
Other
84.6k stars 22.79k forks source link

[AOTI] error: template a rgument 2 is invalid #122980

Open chenyang78 opened 8 months ago

chenyang78 commented 8 months ago

🐛 Describe the bug

How to repro:

comment out the line below:

https://github.com/pytorch/pytorch/blob/57a9a64e10b84fa8d932482ae9417aa3fc3fbf44/test/inductor/test_aot_inductor.py#L2293

and run the following command:

$ python test/inductor/test_aot_inductor.py -k test_fft_c2c_abi_compatible_cpu_with_stack_allocation_and_minimal_arrayref_interface
...
error: template a
rgument 2 is invalid                                                                                                                                               
  588 |       outputs = model->run_impl_minimal_arrayref_interface<AOTInductorModelInputs, AOTInductorModelOutputs>(                                               
      |                 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~^                                               
  589 |           inputs,                                                                                                                                          
      |           ~~~~~~~                                                                                                                                          
  590 |           (torch::aot_inductor::DeviceStreamType)nullptr,                                                                                                  
      |           ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~                                                                                                  
  591 |           nullptr);                                                                                                                                        
      |           ~~~~~~~~

Versions

PyTorch version: 2.3.0a0+git57a9a6 Is debug build: True CUDA used to build PyTorch: 12.0 ROCM used to build PyTorch: N/A

cc @ezyang @msaroufim @bdhirsh @anijain2305 @zou3519 @chauhang @desertfire

int3 commented 7 months ago

It looks like the problem is that DTYPE_TO_CPP maps torch.complex64 to complex64, which is not a valid C++ type. I guess we could use std::complex<double> instead. However, I am wondering if the whole DTYPE_TO_CPP mapping is really necessary. It seems to be used in ABI-compatible mode as an alternative to DTYPE_TO_ATEN, but it seems that these at::-namespaced types are just aliases to types defined in libc10. As previously discussed, we can safely use c10 headers as long as we're careful not to link model.so against libc10, so maybe using DTYPE_TO_ATEN is safe after all?

int3 commented 7 months ago

cc @desertfire

yushangdi commented 3 months ago

re-open because #132810 only fixed it partially. Changes in #132347 can fix this, but those changes broke other tests internally.