Question about THC\THC.h

I cloned a project from github which is NVlabs/wetectron.
And I follow the instruction to set the environment.
The ninja version is 1.10.2.3
conda version is 4.4.10
I use the following commend to download pytorch:
conda install pytorch torchvision torchaudio cudatoolkit=10.2 -c pytorch
so I got pytorch-1.11.0 and cuda-10.2
And then I clone the apex from nvidia, things still work pretty well now
and when I install the ninja with symbolic links
Failed occur:
fatal error: THC/THC.h: No such file or directory
and another error is

I search in google but I didn’t found anything useful.
I am not sure about it was caused by version or I can just simply download the file and paste it somewhere

The TH/THC namespace is dead and apex also removed it already, so you should update apex to the latest version.

2 Likes

Really appreciate for your help, but I think the problem is not caused by apex. I removed the conda environment and clone the apex using
git clone GitHub - NVIDIA/apex: A PyTorch Extension: Tools for easy mixed precision and distributed training in Pytorch
and install it follow the commend as follow:
pip install -v --disable-pip-version-check --no-cache-dir --global-option="–cpp_ext" --global-option="–cuda_ext" ./
The error is still the same, but I found that, it is some csrc/cuda/***.cu file couldn’t find the THC, for example :SigmoidFocalLoss_cuda.cu deform_pool_cuda.cu ROIPool_cuda.cu. Maybe I can find a newer version of these function somewhere? Or any advice I can solve if.

I might have misunderstood the issue as I thought you are running into the building error in apex.
Based on the function names, it seems you are trying to build maskrcnn_benchmark, which still uses the deprecated import here so you might either want to update it or skip its build.

Thank you for your help, I think I have solve the problem. It is caused by version of torch and cudatool. I reinstall pytorch for many times from 1.01 to 1.4 to 1.6 and the project can work now. Besides, I wonder why when I use conda commend to install pytorch and I get the cpu only version.

i am currently facing the same problem, can you tell me what version of cuda and pytorch you used so i could solve it ?

Hi there, I have the same issue with building the mmdetection. Could you tell me how can I update it?
Thank you so much

You could check which deprecated TH(C) methods are currently used in the code base and migrate them to the new ATen API. I would guess that the majority of actual function names should be equal but in a new namespace now (or without the TH tag in their name).

3 Likes

Cuda verison is 10.2, pytorch version is 1.6.0. remember to use the pip install commend from official website to install pytorch.

1 Like

Im trying to compile a convolutional layer repository named “DCNv2” and have the same problem. I have installed CUDA 11.3 (docker image) and pytorch 1.12 and i have tried many things but any of them can solve it. I cant use another pytorch version because my gpu driver isnt compatible with cuda 11.6 (the other option for downloading pytorch) and my gpu isnt compatible with higher driver versions. I have tried to update ATen version but i havent found a higher version. I would appreciate any help. Thanks in advance.

PyTorch is compatible for all CUDA versions >= 10.2 at the moment, so you can pick whichever version would work for you.
As already described, the TH/THC namespace is deprecated and functions from it were moved to the ATen namespace. To fix build issues, check which TH(C) methods are used in the repository you are trying to build and move them to the new ATen calls.

2 Likes

Hi there, I am looking for the replacement of THCudaCheck. Could you provide any clues for it?

1 Like

The macro AT_CUDA_CHECK should work.

1 Like

Hi @ptrblck, thank you for the quick response! I have also figured out a solution from this post. I do have another general question about how to find out which ATen calls are for which TH(C) calls? Are there more guidance or tutorial?

Often the function name is the same and the namespace was just moved so you could remove the TH(C) from the operation and search for this op in the source code. If that doesn’t give you any matches, search for the TH(C) method and try to find the commit which moved it. I don’t think there is any mapping or so of all moved functions.

Awesome! It is very helpful!!

What would be the replacement of THCCeilDiv in the ATen namespace, please?

I found it. First #include <ATen/ceil_div.h>, and then at::ceil_divide() :blush:

hi, I’m trying to compile dcnv2, and I’m getting the error that this lib is not found. Can anybody help me?